The Production Tool Stack for AI Automation

Key takeaways (as of 12 October 2026)

• Serving: vLLM (Apache-2.0, 93,564 stars, v0.31.0 on 5 Oct 2026) is the default open GPU engine and SGLang the main alternative. Hugging Face's TGI was archived read-only on 21 Mar 2026; do not use it for new work.
• Routing: LiteLLM's MIT proxy is free and already has budgets, rate limits and fallbacks. Hosted routers charge on top of model spend: OpenRouter 5.5% (Standard) or 8% (Business). No vendor's own API processes requests in India; AWS Bedrock's in. profiles keep calls to the previous Claude generation, Kimi K3 and GPT-5.6 Terra/Luna in India, while the newest models have no documented India processing.
• Retrieval: pgvector, Docling (MIT) and open-weight embeddings (Qwen3-Embedding, BGE-M3) carry no licence fees. Hosted parsing costs about $1.25 (LlamaParse basic), $15 (Unstructured) or $10–$40 (Reducto) per 1,000 pages.
• Memory: Mem0 leads on stars (66,978) and disclosed funding ($24M). Audit logs, SSO, SOC 2 and HIPAA BAAs are enterprise-tier features at almost every vendor read, and no vendor page read details retention or PII handling. Claude org owners cannot read members' memories; Microsoft 365 Copilot memories are discoverable in Purview but are not covered by retention policies.
• Ownership churn: Weights & Biases (CoreWeave, $1.0B, May 2025), Invariant Labs (Snyk), Protect AI (Palo Alto Networks, $634.5M, July 2025), Prompt Security (SentinelOne, about $180M, September 2025), Lakera (Check Point), Langfuse (ClickHouse), Helicone (Mintlify; now in maintenance mode), promptfoo (OpenAI) and Traceloop (ServiceNow) have all been acquired or agreed to be acquired since May 2025.
• Licence traps: n8n's Sustainable Use License limits use to "your own internal business purposes or for non-commercial or personal use". Dify may not be run as a multi-tenant service without Dify's written authorisation. Phoenix's ELv2 bans providing it to third parties as a hosted or managed service. Flowise was archived on 13 Aug 2026 and reached end of life on 31 Aug 2026. LM Studio's app terms allow only personal or internal business use and bar SaaS, ASP or service-bureau use. All five were re-read on 12 Oct 2026.
• Voice: Retell's own page puts all-in voice-agent cost at $0.07–$0.31 a minute. Vapi's own page lists a $0.05 a minute platform fee plus component costs, and secondary comparisons put the all-in cost at 2.6–6.2 times that fee. Three vendors publish INR prices (vendor-reported): Sarvam (STT ₹30 an hour, TTS ₹3 per 1,000 characters), Plivo India (calls ₹0.38 and ₹0.25 a minute, transcription ₹0.81 a minute) and Bolna (voice agents ₹5.52 a minute on its pilot plan).
• WhatsApp India: businesses whose Meta Sold-To country is India must move all their WhatsApp Business Accounts (WABAs) to INR billing by 31 Dec 2026, or Meta stops delivering those accounts' messages on 1 Jan 2027. Two vendor pages describe a charge from 1 Oct 2026 for service messages beyond 1,000 free per business phone number a month, at the utility rate (₹0.1150 excluding GST). Meta's own page did not confirm it (vendor-reported).

This document covers the layers around the model, not the agent framework

A model API on its own does not make a production automation. A working deployment also needs something to serve or route model calls, a place to store and search company documents, a parser that turns PDFs and scans into clean text, a memory that persists across sessions, traces and tests that show what the system did, guardrails that screen what goes in and out, and the workflow, voice, messaging and builder tools that connect it all to a business process. This document maps those supporting layers as of 12 October 2026 (licences, prices, deals and India-specific rates were re-checked against primary sources that day), for Indian SMEs, US and global companies, and the IT and operations leaders who run their automation. It leaves out agent SDKs, protocols such as MCP and A2A, coding agents and browser agents, which are covered in Agent frameworks and protocols. GPU choices are in Hardware, unit economics and pricing models in Costs, regulation including India's DPDP Act in Governance, and terms in the Glossary.

Method. GitHub stars, licences, archived flags and latest releases come from the GitHub REST API, queried on 11 Oct 2026. "Latest release" means the release GitHub flags as latest. Sometimes that is a sub-package or SDK tag rather than the product's headline version, and the tables say so where it happens. Prices are list prices read from vendor pricing pages on the same day. They are in USD unless stated otherwise, and INR billing, GST treatment and India-region availability were not confirmed for any USD-priced tool vendor (AWS Bedrock's India routing for model calls is noted under gateways). Labels used throughout: "vendor claim" marks a vendor's own statement about performance or traction; "secondary" marks third-party blogs, trackers or aggregators; "single source" marks a figure that only one source supports. Where research passes disagree, the later reading of a primary page is preferred, and the conflict is shown.

Serving and routing: one engine to self-host, one gateway in front of everything

vLLM and SGLang lead GPU serving after TGI's archival

vLLM is the engine to evaluate first. It has 93,564 stars and shipped v0.31.0 on 5 Oct 2026 (vLLM release). SGLang, the main challenger, has 36,990 stars and shipped v0.5.21 on 2 Oct 2026. Both are Apache-2.0, so self-hosting has no per-seat fee; the costs are GPU hours and staff time. An earlier research pass cited v0.30.0 (22 Sep 2026) from a Wikipedia snapshot (Wikipedia: vLLM). The GitHub API shows that snapshot is stale, and the API value is used here.

Hugging Face's Text Generation Inference shows what happens when an engine is retired. Its README says TGI "is now in maintenance mode" and will accept only minor bug fixes, documentation and lightweight maintenance. The repository was archived and made read-only on 21 Mar 2026 (banner re-read on 11 Oct 2026), and its last release was v3.3.7 on 19 Dec 2025 (TGI repository; TGI release). TGI's own documentation points users to "vllm, SGLang, as well as local engines with inter-compatibility such as llama.cpp or MLX" (TGI docs). A third-party migration guide dates the switch to maintenance mode to 11 Dec 2025, and sources disagree on whether TGI still gets security fixes (secondary, unverified) (Spheron migration guide). Existing TGI deployments can keep running, but each needs a migration plan.

No independent 2026 benchmark separates the leading engines. The only throughput comparison found is a single secondary benchmark on one A100 80GB configuration. It gave vLLM 2–24x TGI's throughput, while TGI had a lower time-to-first-token at low concurrency, and it has not been reproduced (Markaicode benchmark). Choosing between vLLM, SGLang and TensorRT-LLM therefore comes down to benchmarking each on the target model and hardware.

Above a single node, NVIDIA Dynamo and llm-d take over. NVIDIA announced Dynamo 1.0 as an open-source inference platform, with integrations for llm-d, vLLM, SGLang, LangChain and LMCache and three standalone modules: KVBM (KV-cache memory manager), NIXL (GPU-to-GPU transfer) and Grove (scaling). NVIDIA claims up to 7x higher inference performance on Blackwell (vendor claim) (NVIDIA press release). The March 2026 date for 1.0 comes from secondary coverage, and the current release is v1.5.1, so 1.0 is now an early milestone. llm-d "extends the capabilities of vLLM with Kubernetes-native architecture for large-scale inference deployments" and uses NIXL to move the KV cache between prefill and decode. Its community launched at Red Hat Summit 2025 (NVIDIA Developer Blog). Both add Kubernetes and NVIDIA dependencies, so they are worth adopting only when multi-node or disaggregated serving is a real requirement.

Local runtimes are developer tools rather than production servers. Ollama has 182,754 stars, the most in the category, and Stack Overflow's 2025 survey found that 51% of developers building agents use it (Stack Overflow 2025 AI survey). That figure measures developer reach, not support for multi-tenant production. GPU sizing and hardware cost are covered in Hardware. INR pricing for GPU cloud in Indian data centres was not checked here; India quotes E2E Networks' current H100 rates.

Gateways split between a free MIT proxy and hosted routers that charge 5–8 percent

A gateway gives one API key across model providers and is where budgets, rate limits, fallbacks and logs are enforced. LiteLLM is the open-source default: an MIT core with 60,964 stars, v1.105.0 released on 11 Oct 2026, and an enterprise/ directory under a separate licence (LiteLLM licence; enterprise directory). The free self-hosted tier already includes virtual keys, users and teams, spend tracking, budgets, rate limits, fallbacks and logging. Enterprise is billed annually, "sized to your usage" and "never per token". It adds SSO with SCIM, OIDC/JWT, audit logs, secret-manager integration, RBAC, a multi-region control plane, a 24/7 SLA and air-gapped deployment (LiteLLM pricing). The pricing page still shows 58.8k stars, which the API figure supersedes. Self-hosting is not free in practice: it needs a server or Kubernetes pod, Redis for caching and engineering time. That cost list comes from a competitor's marketing, so it is biased (TKMX AI).

Hosted routers charge as a share of model spend, or else on logs. On $1,000 of monthly model spend, OpenRouter's fee is $55 on Standard or $80 on Business. On $1,000 of credits, Cloudflare's Unified Billing adds $50. (These are computed from list percentages.) The governance features worth paying for are tied to tiers. OpenRouter's in-region routing for the EU and US needs Business. Portkey's VPC deployment, SOC 2 Type 2 and HIPAA BAAs need Enterprise. Self-hosted LiteLLM keeps the whole data path inside your own cloud. Data residency therefore narrows the shortlist more than price does. Portkey's pricing page gives overages as "+$9 per additional 100k requests" while metering the tier in recorded logs, so confirm the unit on a live account. Azure API Management and OpenRouter's data-retention terms were not researched. AWS Bedrock's India-geography cross-region profiles (in. prefix) route only to Mumbai and Hyderabad for Claude Opus 5, Sonnet 5, Haiku 4.5, Kimi K3 and GPT-5.6 Terra/Luna; the newest models (Claude 5.5, GPT-6) are global-only from India. AWS says not to rely on the prefix alone (call GetInferenceProfile) (Bedrock regional availability).

The retrieval layer can start almost free and scale into monthly minimums

Start with pgvector; Pinecone's $50 minimum pays off only on steady traffic

Open-source vector stores are mature and mostly permissively licensed, but managed pricing comes in four shapes that resist like-for-like comparison: monthly minimums plus usage (Pinecone pricing), per-GiB writes and storage (Chroma pricing), cluster resources with a free-forever tier (Qdrant pricing), and minimums with bring-your-own-cloud (BYOC) options (Turbopuffer pricing). For most business projects with a few million vectors, pgvector in a Postgres you already run is the lowest-risk first step. A dedicated store earns its place once filtering, scale or latency demands it, and a monthly minimum is a fixed cost that low-traffic pilots never use up.

Two licence details need attention. Elasticsearch is licensed under AGPL-3.0, SSPL-1.0 or Elastic License 2.0 depending on the directory (Elasticsearch licence). Weaviate is licensed under BSD-3-Clause or the Weaviate License, also by directory (Weaviate licence). Check the directory before either is embedded in a product you ship or one a partner builds for you. Release cadence is also worth checking: Chroma's latest GitHub release dates from 5 May 2026, five months before this snapshot. The vector store is also an attack surface: OWASP's 2025 LLM Top 10 lists "Vector and Embedding Weaknesses" as LLM08 (OWASP LLM Top 10; see Governance). No 2026 independent benchmark of recall or latency was found, and this research did not confirm a managed India region for any vendor. That decides whether a managed store can keep your data in India.

Open-weight embeddings remove per-token fees; hosted ones start at $0.02 per million tokens

Open-weight embedding models are now competitive and permissively licensed. Qwen3-Embedding (0.6B, 4B and 8B) is Apache-2.0, and its model card states the 8B model ranked first on the MTEB multilingual leaderboard "as of June 5, 2025, score 70.58". That is a dated vendor claim and does not describe the 2026 ranking (Qwen3-Embedding-8B card). BGE-M3 is MIT and supports dense, multi-vector and sparse retrieval in 100+ languages, per its card (BGE-M3 card). The current MTEB leaderboard could not be read because it renders in JavaScript (MTEB leaderboard), so no 2026 ranking supports any model choice here.

Among hosted options, Voyage publishes the clearest per-token list prices. Embedding a 1-billion-token corpus once would cost about $20 on voyage-4-lite, $60 on voyage-4 or $120 on voyage-4-large, before the first 200M free tokens (computed from Voyage's per-1M-token list prices) (Voyage pricing). Cohere lists dedicated Embed 5 instances from $2,000 a month; per-token API rates also appear on its page, but their column mapping was unclear (Cohere pricing). Jina publishes rate limits rather than per-token prices in the captured text (Jina embeddings). OpenAI's embedding prices were not in the fetched text of its pricing page (OpenAI pricing). The researchers' working view is that adding a reranker usually buys more quality per dollar than switching embedding models. That is an inference, not a tested result. Retrieval quality for Hindi, other Indic scripts and code-mixed text was not tested for any model, so vendors' multilingual claims should not be read as covering Indian languages. Indian languages covers the Indic benchmarks and embedding options.

Document parsing ranges from free to $40 per 1,000 pages

Retrieval can only find what the parser extracted, and the open-source parsers are strong. Docling (MIT, 68,677 stars, v2.137.0 on 9 Oct 2026) converts PDFs and Office files into layout-aware Markdown or JSON, and Unstructured (Apache-2.0) handles ETL-style partitioning and chunking. Both run in-house. Hosted parsers charge per page, and at volume the differences are large. At list prices, 100,000 pages a month cost about $1,350 on Unstructured (90,000 billable pages after the 10,000 free, at $0.015) (Unstructured pricing). The same volume costs $1,000–$4,000 on Reducto, depending on endpoint tier (Reducto pricing), and about $112.50 on LlamaParse's basic mode (90,000 credits after the 10K monthly free credits, at $1.25 per 1,000). Agentic LLM or VLM parsing costs more credits per page (LlamaParse pricing). All three figures are computed from list prices.

Mistral prices OCR per 1,000 pages, but the dollar figure was missing from the fetched text. Batch processing halves the price, and Mistral claims up to 2,000 pages a minute on one node (vendor claim) (Mistral pricing; Mistral OCR). Azure Document Intelligence's page renders in JavaScript (Azure pricing), but the Azure Retail Prices API lists Read at $0.60 or $1.50 per 1,000 pages (two unlabelled rows, unresolved), prebuilt models at $10 and custom models at $30 per 1,000 pages. Google Document AI Enterprise OCR costs $1.50 per 1,000 pages (1,000–5M pages; $0.60 above 5M) (Document AI pricing), and AWS Textract Detect Document Text costs $0.0015 a page for the first 1M (Textract pricing). The sensible default is to start with Docling or Unstructured and add a hosted parser only where accuracy on tables or handwriting falls measurably short. Indic-script OCR (Hindi, Tamil, Bengali) was not tested for any tool, so run a pilot on your own scans before relying on any accuracy claim. Indian-language OCR and document AI, with rupee page prices, are compared in Indian languages.

GraphRAG earns its indexing bill only on multi-hop questions

Graph-based retrieval builds an entity-and-relation graph from the corpus. That helps with questions that span several documents, but every chunk costs LLM calls to index, and the cost recurs whenever the corpus changes. Microsoft GraphRAG (MIT, v3.3.0 on 8 Oct 2026) and LightRAG (MIT, 40,068 stars, v1.5.7 on 2 Sep 2026) are the two most-starred open frameworks (GraphRAG; LightRAG). LightRAG describes itself as an EMNLP 2025 paper, and its authors report faster indexing, which was not verified here. Neo4j is the mature graph-database option for GraphRAG pipelines. Its community repository is GPL-3.0, enterprise use requires a commercial licence, and its GitHub "latest release" is a 2017 alpha tag that cannot be used for version tracking (Neo4j). The pattern that fits is a stable, entity-heavy corpus such as contracts, policies or research. For FAQ-style support, GraphRAG usually costs more than plain RAG (inference, not benchmarked). No independent 2026 benchmark comparing GraphRAG with plain RAG on accuracy and cost was found.

The wider "RAG versus long context" debate is moving toward hybrids. A preprint submitted on 6 Oct 2026, UNREAL, uses one model-internal mechanism to select evidence both from a corpus and from inside a long prompt. It reports HotpotQA recall rising from 49.1% to 73.2%, 2WikiMultiHopQA from 31.7% to 60.1%, and NoLiMa accuracy from 1.0% to 24.83% at 128K tokens. It also reports lower FLOPs and time-to-first-token than full-context inference from about 32K tokens upward. These are the authors' own, non-peer-reviewed results (arXiv:2610.08463). The field is active: an arXiv query for "long context" AND RAG returned 249 papers on 11 Oct 2026 (arXiv query). Vendors' own memory features are built on retrieval: Claude's chat-memory search "use[s] Retrieval-Augmented Generation (RAG)" exposed as tool calls (Claude Help Center). The practical default is retrieval first (vector, keyword or hybrid), long context for the top results or for a single long document, and agentic search over an index for open-ended questions. Prompt length still drives cost and latency, so retrieval keeps mattering even as context windows grow.

Memory layers compete on governance, and most of it is unpublished

Open-source memory engines: Mem0 leads on stars and disclosed funding

A memory layer stores facts about users, customers and the organisation so an assistant or agent can use them in later sessions. The open-source field is crowded: Mem0 has 66,978 stars, Cognee 32,012, Graphiti 31,642, Supermemory 31,185, Letta 25,102 and MemOS 11,813. Commercially there are two patterns. One is a venture-funded managed service (Mem0, Zep). The other is a deploy-anywhere open-source engine that sells enterprise support, BYOC or air-gapped installs (Cognee, Supermemory).

Funding is verified only for Mem0. TechCrunch reported $24M in total, a $3.9M seed plus a $20M Series A led by Basis Set Ventures, with Kindred, Y Combinator, Peak XV and GitHub Fund, announced on 28 Oct 2025 (TechCrunch). Aggregators disagree on the total: VCBacked lists $20.0M (VCBacked: Mem0), and Caplight lists about $24.4M with a $150M valuation estimate (Caplight: Mem0). TechCrunch's primary report is preferred because it gives the round breakdown. The same article carries company-stated traction at the time: 41k+ stars, 13M+ Python downloads, 80k+ cloud signups, and a claim to be the "exclusive memory provider" for AWS's new Agent SDK (company claims). Letta raised a $10M seed led by Felicis, dated to September 2024 by one listing (secondary) (Pulse 2.0). Supermemory's seed is reported as about $2.6M (Dealroom) or $3M (Let's Data Science), and the amount is disputed. Smaller 2026 rounds include xmemory (London), reported at $4.0M pre-seed in March 2026 (secondary, unverified) (VCBacked: xmemory). No funding was found for Zep or Cognee, and a 2026 round reported for MemOS's developer MemTensor gave no clear amount (secondary, unverified). In the disclosed sample, rounds run from about $2.6M to $24M, which suggests few players have the capital for heavy enterprise-governance engineering. That is an inference from a small sample.

Native assistant memory: four vendors, four admin models

The assistants that companies already license have their own memory, and the admin models differ in ways that matter to a compliance officer. Claude has the most detailed published admin model among the pages read (Claude Help Center). On Team and Enterprise, owners switch memory on for the organisation and each member stays off until they opt in. Owners cannot view or edit individual memories, and switching org memory off deletes all memory for all users immediately. The audit log records only the org-level switch. Microsoft 365 Copilot takes the opposite approach. Memory is on by default under a tenant "Enhanced personalization" control. Memories are stored in a hidden folder in each user's Exchange mailbox, and admins can search, export and delete them through Purview eDiscovery and Graph. But admins cannot restrict what is saved, and Purview retention policies do not apply to memory (Microsoft Learn). Gemini's personal-context features "are not available while signed in to a work, school, or supervised account" (Gemini Apps Help). ChatGPT's Business and Enterprise help pages returned HTTP 403, so its controls are known only from search snippets, below. Model-level detail is in Claude, OpenAI and Gemini.

Governance controls side by side

The table below compares what each vendor page says about the controls an enterprise buyer typically asks for. "Not read" means the page did not describe the control, not that the product lacks it.

Three conclusions follow. First, compliance badges are table stakes on enterprise tiers: SOC 2 and HIPAA BAAs appear across Zep, Supermemory and Portkey's Enterprise gateway tier, so US regulated buyers will expect them. Second, the controls buyers most need are the least documented: no vendor page read describes retention periods, PII handling, admin audit of individual memory writes or data residency in detail. Third, the native models force a trade-off. Claude protects employee privacy from administrators, which a regulated buyer may find hard to accept. Microsoft gives administrators discovery but no control over what is remembered and no retention-policy coverage. For Indian deployments, memory stores hold personal data about employees and customers, so the DPDP Act and Rules apply. Consent-manager provisions take effect on 13 Nov 2026 and the substantive obligations on 13 May 2027, under the Rules notified as G.S.R. 846(E) (MeitY DPDP Rules 2025). See Governance and India.

Observability, evaluation and guardrails changed owners in 2026

Langfuse keeps an MIT core under ClickHouse; Phoenix is only source-available

Consolidation is the story of 2025 and 2026. ClickHouse acquired Langfuse, announced on 16 Jan 2026 alongside ClickHouse's own $400M Series D at a $15B valuation led by Dragoneer; the purchase price was not disclosed (ClickHouse blog; SiliconANGLE). Langfuse says its roadmap is unchanged and that it remains committed to open source and self-hosting (Langfuse discussion #11593). Its LICENSE, re-read on 11 Oct 2026, now reads "2023-2026 ClickHouse, Inc.", with the core under the MIT (Expat) licence and ee/ directories under a separate licence. Helicone joined Mintlify under a company post dated 3 Mar 2026; its services stay live "in maintenance mode", with security updates, new models and fixes still shipping (Helicone blog). OpenAI agreed to acquire promptfoo, Promptfoo's own blog said on 9 Mar 2026, adding that "the open-source project will continue". The LICENSE on main is still MIT (Promptfoo blog; promptfoo LICENSE). Traceloop (OpenLLMetry) is "joining ServiceNow", per its own blog in March 2026; no value was stated (Traceloop blog). CoreWeave acquired Weights & Biases on 5 May 2025 for $1.0B aggregate purchase consideration (CoreWeave 10-Q; primary), and the W&B pricing URL now redirects to CoreWeave (W&B pricing URL).

OpenAI is also retiring its own hosted evaluation product. Existing OpenAI Evals become read-only on 31 Oct 2026, and the dashboard and API shut down on 30 Nov 2026 (OpenAI deprecations). The same page names Promptfoo, which OpenAI agreed to acquire (above), as the migration path. Teams that rely on it have about seven weeks to move. Arize Phoenix is marketed as "Phoenix OSS", but its Elastic License 2.0 states, "You may not provide the software to third parties as a hosted or managed service" (Phoenix). LangSmith is a hosted product, and both its hybrid and fully self-managed deployments are Enterprise-only (LangSmith pricing). Among developers building agents, the most-used observability tools in Stack Overflow's 2025 survey were general-purpose ones: Grafana and Prometheus (43%) and Sentry (32%) (Stack Overflow 2025 AI survey). That makes the vendor-neutral OpenTelemetry GenAI semantic conventions (spec; the semantic-conventions repository's latest release is v1.44.0 of 4 Aug 2026) the safest instrumentation choice: traces then survive any change of observability vendor. The conventions have moved to github.com/open-telemetry/semantic-conventions-genai (the opentelemetry.io page now says "Moved") and are marked "Development", not stable (GenAI spans spec).

On price, Langfuse Core at $29 a month (100k units included, then $8 per 100k) is the cheapest paid hosted tier found. One million units a month would cost about $101 (computed) (Langfuse pricing). Helicone Pro ($79) and Braintrust Pro ($249) meter in different units (requests, GB and per-1,000 scores), so they cannot be compared directly (Helicone pricing; Braintrust pricing). Two maintenance signals matter as well. Ragas has had no release since 13 Jan 2026 and no push since 24 Feb 2026. Helicone's last GitHub release is from 21 Aug 2025, although the repository was still being pushed on 10 Oct 2026. Helicone's own post (3 Mar 2026) confirms that it is in maintenance mode after the Mintlify acquisition (above).

AI security start-ups are being absorbed by security incumbents

Independent AI-security vendors are being bought by established security companies. Check Point announced it would acquire Lakera in September 2025, and the deal closed on 22 Oct 2025, as reported in Check Point's Q3 2025 results (search summary; not read directly). Lakera brings runtime protection, continuous red teaming and multilingual defences into Check Point's Infinity platform, and Zurich becomes Check Point's Global Centre of Excellence for AI Security (CXOToday release; Check Point Q3 2025). Lakera's site footer now carries Check Point's copyright (Lakera). Snyk acquired Invariant Labs, dated 24 Jun 2025 by Hacker News (Snyk news; HN item). Invariant had disclosed MCP "tool poisoning" attacks on 1 Apr 2025 and released MCP-Scan on 11 Apr 2025 (Invariant Labs). SentinelOne acquired Prompt Security: the agreement was signed on 5 Aug 2025 and completed on 5 Sep 2025, for total consideration of about $180 million in cash and shares (SentinelOne 8-K; primary). The higher price given on Wikipedia is not supported by the filing (Wikipedia: Prompt Security). HiddenLayer's homepage shows a "$100M Series B" banner, undated, and no acquisition (HiddenLayer). Palo Alto Networks completed its acquisition of Protect AI on 22 Jul 2025, for total purchase consideration of $634.5 million (Palo Alto Networks FY2025 10-K; primary).

On the open side, Meta's guard models are the cheapest self-hosted screening option on licence cost. Llama Guard 4 (12B) classifies inputs and outputs, and Llama Prompt Guard 2 (86M) detects prompt injection and jailbreaks. Both are released under the Llama 4 Community License. Anyone distributing them, or a product containing them, must display "Built with Llama", include the Notice file and follow Meta's Acceptable Use Policy. Derived models must have names beginning "Llama", and licensees above 700 million monthly active users need a separate licence from Meta (Llama 4 licence; Llama Guard 4; Prompt Guard 2). Gateways now embed guardrails too: Cloudflare bills guardrails as Workers AI tokens on Llama Guard 3 8B, and Portkey includes deterministic guardrails on its free tier. No tool here makes prompt injection a solved problem. OWASP's 2025 list ranks Prompt Injection as LLM01, and its LLM06 (Excessive Agency) mitigations are least-privilege tools, downstream authorisation and human approval for high-impact actions. Those are design controls that no guard model replaces (OWASP LLM Top 10). See Governance and, for agent-level security, Agent frameworks.

Automation surfaces: workflow engines, voice, WhatsApp and low-code builders

n8n has the largest community and the tightest licence

n8n is the most-starred workflow tool in this comparison, at 206,996 stars (n8n@2.42.6 on 9 Oct 2026), but it is source-available rather than open source. Its Sustainable Use License says: "You may use or modify the software only for your own internal business purposes or for non-commercial or personal use" (n8n LICENSE). It allows distribution or provision to others only "free of charge for non-commercial purposes". Files with .ee. in their name, or .ee in their directory name, are outside that licence and need an n8n Enterprise License (licence re-read on 12 Oct 2026). Running n8n as a paid managed service for third parties therefore needs a commercial agreement with n8n (inference from the text; no n8n FAQ found). If a partner proposes to host n8n for you as a paid managed service, ask whether it has that agreement, and get a lawyer's reading where it matters. A direct read of n8n's pricing page on 12 Oct 2026 (annual billing) shows Starter at €20 a month (2,500 executions, one shared project, 1,600 AI-assistant credits), Pro at €50 (10,000) and Business at €667 (40,000, "for companies with < 100 employees") (n8n pricing). Business is labelled both "Self-hosted only" and "Hosted by n8n?", so its hosting option is ambiguous. An earlier summarised read of the same page adds Business overage of €4,000 per 300,000 extra executions (secondary; not re-checked on 12 Oct 2026). Secondary trackers give €24 and €60 on monthly billing and Business at $800 (Toolradar), and one says the Cloud free plan ended on 14 Apr 2026, which is unverified (AgentDeals).

The cleaner licences when a partner hosts and manages the workflow tool for you are Activepieces (MIT core with an ee/ carve-out; 25,000 stars; around 400 MCP servers per its repository description) and Windmill (Apache-2.0 and AGPLv3 dual-licensed, with some proprietary components; 18,162 stars). For durable, retry-safe, long-running agent workflows, Temporal (MIT; Cloud at $50 per million actions, $150 of free credits for 90 days, active storage at $0.042/GBhr) gives a predictable cost model (Temporal pricing). Prefect (Apache-2.0) and Airflow (Apache-2.0, 47,146 stars) remain the data-pipeline choices. Among hosted tools, Zapier's MCP product connects to 9,000+ apps (Zapier MCP), and Make offers AI Agents (beta) on all plans plus an on-prem agent for local networks. RPA-plus-AI combinations (UiPath, Power Automate) were not researched.

Voice agents cost two to six times their headline platform rate

Voice cost has to be modelled per minute across five legs: platform, LLM, speech-to-text, text-to-speech and telephony. Retell's own page shows "$0.07–$0.31 / min", and its worked example of $0.11 a minute splits into $0.04 for the LLM, $0.055 for voice infrastructure and $0.015 for TTS, with telephony at $0.00 (Retell pricing). Secondary comparisons put a tuned Retell stack near $0.10 a minute and a current-generation model past $0.23 (Morph). Vapi's own page lists $0.05 a minute for Vapi hosting, with transcriber, model and voice billed as separate components (Vapi pricing; primary). Secondary comparisons put the all-in cost at $0.13–$0.31 once LLM, TTS, STT and Twilio are added (AI Agent Rank). Against that fee, the all-in range is 2.6 to 6.2 times higher, and Retell's example is double its infrastructure rate (computed). Vapi's own page lists Enterprise from a $999 monthly minimum and Twilio pass-through at $0.008 (inbound) and $0.014 (outbound) a minute (Vapi pricing). Bland bundles the models, at $0.14 a minute on Start and $0.12 on a $299 monthly plan (Bland pricing). Its blog adds a Scale tier at $0.11 on $499 a month, with telephony billed separately (vendor-authored) (Bland blog). Synthflow's enterprise contracts "start at $30,000 annually" (Synthflow pricing).

The self-host path is LiveKit Agents (Apache-2.0, 14,694 stars) or Pipecat (BSD-2-Clause, 16,356 stars). LiveKit's cloud charges $0.01 per agent-session minute beyond the included minutes (Ship $50/month for 5,000 minutes; Scale $500 for 50,000), and its calculator's example stack totals $0.0479 a minute with models billed separately (summarised read) (LiveKit pricing). For Indian languages, Sarvam publishes INR rates: ₹30 an hour for real-time or batch STT (₹0.50 a minute, computed), ₹45 an hour for batch with diarization, and ₹3 per 1,000 characters for streaming TTS (Sarvam API pricing). Pricing the speech leg in INR removes currency exposure on that leg, but Sarvam's quality against English-first vendors was not benchmarked independently. Indian-language speech-to-text, text-to-speech and a leg-by-leg rupee voice-agent minute are covered in Indian languages; the India document quotes the same Sarvam rates. Plivo's India page lists calls at ₹0.38 and ₹0.25 a minute (rows not labelled by direction), a domestic number at ₹200 a month and transcription at ₹0.81 a minute (Plivo India); its global page lists a Voice AI Agent at $0.03 a minute excluding telephony (Plivo). Bolna sells voice agents at ₹5.52 (6.00 US cents) a minute on its $15 pilot plan and 4.51 cents at a higher tier (Bolna pricing; vendor-reported). Exotel's pricing page is sales-led (Exotel); its blog reportedly puts Indian-number calls at about ₹0.80–1.00 a minute outbound and ₹0.30–0.50 inbound (Exotel blog; search summary, not read directly). Reseller figures of ₹4–6 a minute for Exotel AI agents conflict with one another and are not Exotel's own price (edesy; reseller, search summary). Twilio India (Twilio India voice) and Exotel rate cards remain unresolved.

TRAI's Direction of 4 May 2024 moved voice calling onto DLT (140-series within 60 days; a 160-series plan) (TRAI Direction). The draft TCCCPR Third Amendment Regulations 2026 add an A2P call definition, pre-declaration and termination charges for A2P calls, a lower complaint threshold before a sender is suspended, and operator "AI-flagging" (comments on the draft hosted by TRAI). As drafted, a fully automated voice agent appears to fall within A2P (inference). The final rules were reported as notified on 18 Sep 2026 (secondary); the official text was not read, and no explicit AI-voice disclosure rule is confirmed. Indian languages covers the draft in more detail.

WhatsApp India billing must move to INR by 31 December 2026

WhatsApp-based order and lead handling is among the likely high-demand automations for Indian SMEs (an inference in the India research, not a survey result), and Meta's own pricing page sets the rules (Meta WhatsApp pricing). Template messages are charged per message by category and country calling code. All non-template messages are free inside an open customer-service window, and so are utility templates delivered inside one. The binding date for eligible India-billed accounts is 31 Dec 2026. Billing localisation launched on 1 Jan 2026 for partners and directly integrated businesses whose Sold-To country is India. Eligible customers "must ensure all WABAs in their business portfolio are migrated to INR by December 31, 2026", because "as of January 1, 2027 Meta will no longer deliver the messages of non-INR WABAs of eligible customers" (page re-read on 11 Oct 2026 through a summarising reader). Meta's rate-card files (PDF and CSV) were not read. A vendor page gives India rates per message, excluding 18% GST, "effective 1 Oct 2026": marketing ₹0.8631, utility ₹0.1150, authentication ₹0.1150, authentication-international ₹2.4971 (ChatMaxima; vendor-reported). Gallabox's INR tab shows exactly 1.15 times those (marketing ₹0.9926; others ₹0.1323), reason not stated (Gallabox). At ₹0.8631, 10,000 marketing templates cost about ₹8,631 before GST and partner markups (computed). Rupee per-conversation costs for Indian-language bots are worked through in Indian languages. Business-solution-provider (BSP) markups were not researched, and for small businesses they may matter more than Meta's rate. The workflow tools above supply the automation logic. Because these flows process customer personal data, they fall under the DPDP timeline in Governance.

Flowise's archival leaves Dify and Langflow as the open builders

Low-code builders let non-developers assemble agents, but two of the three big open projects carry warnings. Flowise was archived by its owner on 13 Aug 2026 (repository banner re-read on 11 Oct 2026), after an announcement and code freeze on 29 Jul 2026; end of life was 31 Aug 2026, and its npm packages and Docker images are deprecated (Flowise discussion #6727). Existing deployments are now unsupported and need migration plans. Dify (158,150 stars, 1.17.1 on 10 Sep 2026) uses a modified Apache-2.0 licence. It bars running a multi-tenant environment "unless explicitly authorized by Dify in writing", where "one tenant corresponds to one workspace", and it bars removing or modifying the logo or copyright in the console or applications whenever Dify's frontend is used (licence re-read on 11 Oct 2026) (Dify). That makes it suitable for a deployment dedicated to one organisation, but a shared hosted product serving many businesses needs Dify's written authorisation. Langflow (MIT, 155,498 stars, v1.12.5 on 6 Oct 2026) is the permissive alternative. Hosted builders can also disappear: OpenAI's visual Agent Builder shuts down on 30 Nov 2026, and OpenAI's migration guidance points to the Agents SDK or ChatGPT Workspace Agents (OpenAI deprecations). Microsoft Learn lists tenant-wide packs of 25,000 Copilot Credits at $200 a pack a month ($0.008 a credit, computed). A pay-as-you-go meter billed through Azure also exists; a rate of $0.01 per credit is reported in secondary sources, not read directly (Microsoft Learn: Copilot Studio billing).

Licence traps to check when you or a partner runs a tool for you

Four kinds of trap appear across the stack. Use-restriction licences (n8n's SUL, Phoenix's ELv2, Dify's multi-tenant clause, LM Studio's app terms) allow self-hosting for your own business but restrict hosting the tool for several customers at once, which is how a managed-service or implementation partner may propose to run it for you. Directory-split open core (LiteLLM enterprise/, Langfuse and Activepieces ee/, n8n's .ee. files, Flowise's enterprise carve-out, and Elasticsearch and Weaviate by directory) makes a feature's licence depend on which folder its code lives in, so it has to be re-checked at each upgrade. Copyleft terms (AGPL-3.0 and SSPL-1.0 in Elasticsearch, AGPLv3 in Windmill, GPL-3.0 in Neo4j Community) matter when code is embedded or hosted for others. The same applies to the AGPL-licensed browser agent Skyvern, covered in Agent frameworks. Status traps are not licence clauses but behave like them: archived projects (TGI, Flowise), ownership changes, and stalled release cadences (Ragas, Helicone, Chroma, TensorRT-LLM). Across vendors, SSO and audit logs are gated to top tiers (LiteLLM, Mem0, Portkey, Prefect Team, and n8n Enterprise per secondary sources). The GitHub API returns NOASSERTION for several licences that the LICENSE files settle. Dynamo and NeMo Guardrails are Apache-2.0. The MCP servers repository is mixed: new contributions Apache-2.0, documentation CC-BY-4.0, un-relicensed contributions MIT. This is not legal advice: items marked "counsel" need a lawyer's reading before you rely on them in a contract with a vendor or partner.

Three reference stacks: budget SME, privacy-first self-hosted, enterprise and regulated

These stacks are recommendations drawn from the comparisons above, not tested deployments. They assume a business that runs its automations itself or through an implementation partner. Prices are list prices in USD, and INR billing and GST were not checked.

Stack A, SME on a budget, uses hosted models with minimal operations and keeps software subscriptions between about $0 (everything self-hosted) and about $70 a month: Langfuse Core $29 plus Mem0 Starter $19 plus Activepieces' $20 tier, excluding model usage, hosting and staff (computed). Stack B, privacy-first self-hosted, suits businesses that need data kept in their own or in-country infrastructure and DPDP-aligned retention. Licence fees are close to zero, and GPUs and operations dominate cost. Stack C, enterprise and regulated, suits US companies that require SOC 2, HIPAA BAAs, SSO and audit trails. Its fixed floor runs to thousands of dollars a month before quotes, for example Langfuse Enterprise at $2,499 plus a Pinecone Enterprise minimum of $500, or $2,999 together (computed).

What this means for your business

Start with permissive defaults and keep hosted services swappable. Each layer now has a permissively licensed default (MIT, Apache-2.0, BSD or the PostgreSQL License): vLLM, LiteLLM, pgvector or Qdrant, Docling, Langfuse's core, Activepieces or Temporal, and LiveKit or Pipecat. Hosted alternatives mostly add cost at volume, at enterprise tiers or through monthly minimums. Put a gateway in front of every model call and use OpenTelemetry GenAI traces, so that switching a model, router or observability vendor is a configuration change rather than a rebuild. Nine of the observability, evaluation and security vendors in this document have been acquired or agreed to be acquired since May 2025. The same gateway can route Indian personal data to in-country options (Bedrock's in. profiles for the previous Claude generation, Indian managed endpoints or self-hosted vLLM), since first-party APIs and the newest models process outside India.

Budget in all-in units and plan for currency risk. Costs arrive per minute (voice, $0.07–$0.31 all-in), per page (parsing, $0.00125–$0.04), per message (WhatsApp, INR rate cards), as a percentage of model spend (5–8% on hosted routers) or as monthly minimums (Pinecone, Turbopuffer). Software floors run from about $0–$70 a month for the budget SME stack to about $3,000 a month before quotes for the enterprise stack (reference stacks above). Almost every vendor prices in USD. Sarvam, Plivo India and Bolna publish AI-service or telephony prices in INR (vendor-reported), and Meta now requires INR billing for eligible WhatsApp customers whose Sold-To country is India. If your revenue is in INR and a contract passes USD costs through to you, agree an FX clause or keep a buffer. Costs has the wider pricing models.

Put the dated deadlines in your plan. OpenAI Evals goes read-only on 31 Oct 2026; Evals and Agent Builder shut down on 30 Nov 2026; DPDP consent-manager provisions start on 13 Nov 2026; eligible India-billed WABAs must migrate to INR billing by 31 Dec 2026; the substantive DPDP obligations follow on 13 May 2027. If you run TGI or Flowise (end of life 31 Aug 2026), you already need a migration.

Test Indian-language quality yourself. No embedding model, OCR tool or memory layer in this research was tested on Hindi, other Indic scripts or code-mixed text, and vendors' multilingual claims do not settle the question. Pilot on your own documents, scans and call recordings before accepting any accuracy figure (Indian languages sets out a test plan). Memory layers that hold employee or customer data need retention, deletion and audit controls that the vendors themselves do not document.

Questions to ask any vendor or implementation partner

• Which licence covers each component, and does the way it will be run for you fit it? Hosting n8n, Dify, Phoenix or LM Studio for several customers can breach their terms (see the licence table above).
• Is the instance owned by your business, and does the contract record each component's licence, the self-host commitment and an export path?
• Is any component archived, end-of-life or in maintenance mode (TGI, Flowise, Helicone), or slow to release (Ragas)?
• Where are your data processed and stored, and which tier includes SSO, audit logs and retention controls?

Open questions and unverified items

Sources

All sources were read or queried on 11 October 2026, or on 12 October 2026 where the text says so, unless a different date is given in the text.

1. GitHub: vllm-project/vllm release v0.31.0
2. Wikipedia: vLLM (secondary)
3. GitHub: huggingface/text-generation-inference repository and README
4. GitHub REST API: huggingface/text-generation-inference latest release
5. Hugging Face: Text Generation Inference documentation
6. Spheron: migrating from TGI to vLLM or SGLang, 2026 (secondary)
7. Markaicode: Llama 3.3 on A100 80GB latency benchmark (secondary)
8. NVIDIA investor press release: NVIDIA enters production with Dynamo
9. NVIDIA Developer Blog: NVIDIA Dynamo accelerates llm-d community initiatives
10. Stack Overflow 2025 Developer Survey: AI section
11. GitHub REST API: vllm-project/vllm repository metadata
12. GitHub REST API: vllm-project/vllm latest release
13. GitHub REST API: sgl-project/sglang repository metadata
14. GitHub REST API: sgl-project/sglang latest release
15. GitHub REST API: NVIDIA/TensorRT-LLM repository metadata
16. GitHub REST API: NVIDIA/TensorRT-LLM licence file
17. GitHub REST API: NVIDIA/TensorRT-LLM latest release
18. GitHub REST API: ai-dynamo/dynamo repository metadata
19. NVIDIA Dynamo LICENSE
20. GitHub REST API: ai-dynamo/dynamo latest release
21. GitHub REST API: llm-d/llm-d repository metadata
22. GitHub REST API: llm-d/llm-d latest release
23. GitHub REST API: ollama/ollama repository metadata
24. GitHub REST API: ollama/ollama latest release
25. Ollama pricing
26. GitHub REST API: ggml-org/llama.cpp repository metadata
27. GitHub REST API: ggml-org/llama.cpp latest release
28. GitHub REST API: mudler/LocalAI repository metadata
29. GitHub REST API: mudler/LocalAI latest release
30. GitHub REST API: ml-explore/mlx repository metadata
31. GitHub REST API: ml-explore/mlx-lm repository metadata
32. GitHub REST API: ml-explore/mlx latest release
33. GitHub REST API: ml-explore/mlx-lm latest release
34. LM Studio homepage
35. LM Studio desktop app terms (23 Aug 2026)
36. GitHub REST API: lmstudio-ai/lmstudio-js repository metadata
37. GitHub REST API: huggingface/text-generation-inference repository metadata
38. GitHub REST API: BerriAI/litellm licence file
39. GitHub REST API: BerriAI/litellm contents, enterprise
40. LiteLLM pricing
41. TKMX AI: LiteLLM alternatives in 2026 (competitor marketing)
42. AWS Bedrock: model support by Region (read 12 Oct 2026)
43. GitHub REST API: BerriAI/litellm repository metadata
44. GitHub REST API: BerriAI/litellm latest release
45. OpenRouter pricing
46. GitHub REST API: Portkey-AI/gateway repository metadata
47. GitHub REST API: Portkey-AI/gateway latest release
48. Portkey pricing
49. Kong AI Gateway product page
50. GitHub REST API: Kong/kong repository metadata
51. GitHub REST API: Kong/kong latest release
52. Kong pricing
53. Cloudflare AI Gateway pricing documentation
54. Pinecone pricing
55. Chroma pricing
56. Qdrant pricing
57. Turbopuffer pricing
58. GitHub REST API: elastic/elasticsearch licence file
59. GitHub REST API: weaviate/weaviate licence file
60. OWASP GenAI: Top 10 for LLM Applications 2025
61. GitHub REST API: pgvector/pgvector repository metadata
62. GitHub REST API: pgvector/pgvector licence file
63. GitHub REST API: qdrant/qdrant repository metadata
64. GitHub REST API: qdrant/qdrant latest release
65. GitHub REST API: milvus-io/milvus repository metadata
66. GitHub REST API: milvus-io/milvus latest release
67. Zilliz pricing
68. GitHub REST API: weaviate/weaviate repository metadata
69. GitHub REST API: weaviate/weaviate latest release
70. Weaviate pricing
71. GitHub REST API: chroma-core/chroma repository metadata
72. GitHub REST API: chroma-core/chroma latest release
73. GitHub REST API: lancedb/lancedb repository metadata
74. GitHub REST API: lancedb/lancedb latest release
75. GitHub REST API: opensearch-project/OpenSearch repository metadata
76. GitHub REST API: opensearch-project/OpenSearch latest release
77. GitHub REST API: elastic/elasticsearch repository metadata
78. GitHub REST API: elastic/elasticsearch latest release
79. Elastic pricing
80. Hugging Face: Qwen/Qwen3-Embedding-8B model card
81. Hugging Face: BAAI/bge-m3 model card
82. Hugging Face: MTEB leaderboard
83. Voyage AI pricing documentation
84. Cohere pricing
85. Jina AI embeddings
86. OpenAI API pricing
87. Unstructured pricing
88. Reducto pricing
89. LlamaIndex (LlamaParse) pricing
90. Mistral AI pricing
91. Mistral AI: Mistral OCR announcement
92. Microsoft Azure: Document Intelligence pricing
93. Azure Retail Prices API: Document Intelligence
94. Google Cloud Document AI pricing
95. AWS Textract pricing
96. GitHub REST API: docling-project/docling repository metadata
97. GitHub REST API: docling-project/docling latest release
98. GitHub REST API: Unstructured-IO/unstructured repository metadata
99. GitHub REST API: Unstructured-IO/unstructured latest release
100. GitHub REST API: run-llama/llama_parse repository metadata
101. GitHub REST API: run-llama/llama_parse latest release
102. GitHub REST API: deepseek-ai/DeepSeek-OCR repository metadata
103. GitHub REST API: microsoft/graphrag repository metadata
104. GitHub: HKUDS/LightRAG repository
105. GitHub REST API: neo4j/neo4j repository metadata
106. arXiv:2610.08463, UNREAL: Unifying Retrieval and Long-Context with a Single Model (preprint)
107. arXiv API query: "long context" AND RAG
108. Claude Help Center: Use Claude's chat search and memory to build on previous context
109. GitHub REST API: microsoft/graphrag latest release
110. GitHub REST API: HKUDS/LightRAG repository metadata
111. GitHub REST API: HKUDS/LightRAG latest release
112. GitHub REST API: getzep/graphiti repository metadata
113. GitHub REST API: getzep/graphiti latest release
114. TechCrunch (28 Oct 2025): Mem0 raises $24M from YC, Peak XV and Basis Set
115. VCBacked: Mem0 (aggregator)
116. Caplight: Mem0 (aggregator)
117. Pulse 2.0: Letta raises $10 million seed (secondary)
118. Dealroom: Supermemory secures $2.6M (secondary)
119. Let's Data Science: Supermemory raises $3M (secondary)
120. VCBacked: xmemory (aggregator)
121. GitHub REST API: mem0ai/mem0 repository metadata
122. GitHub REST API: mem0ai/mem0 latest release
123. Mem0 pricing
124. GitHub REST API: getzep/zep repository metadata
125. Zep README (Community Edition deprecated)
126. GitHub REST API: getzep/zep latest release
127. Zep pricing
128. GitHub REST API: letta-ai/letta repository metadata
129. GitHub REST API: letta-ai/letta latest release
130. GitHub REST API: topoteretes/cognee repository metadata
131. GitHub REST API: topoteretes/cognee latest release
132. Cognee pricing
133. GitHub REST API: supermemoryai/supermemory repository metadata
134. GitHub REST API: supermemoryai/supermemory latest release
135. Supermemory pricing
136. GitHub REST API: MemTensor/MemOS repository metadata
137. GitHub REST API: MemTensor/MemOS latest release
138. GitHub REST API: langchain-ai/langmem repository metadata
139. GitHub REST API: modelcontextprotocol/servers repository metadata
140. modelcontextprotocol/servers LICENSE
141. GitHub REST API: modelcontextprotocol/servers contents, src/memory/README.md
142. Microsoft Learn: Microsoft 365 Copilot personalization and memory
143. Gemini Apps Help: personal context and memory
144. OpenAI Help Center: Managing workspace settings in ChatGPT Enterprise (search snippet)
145. OpenAI blog: Memory and new controls for ChatGPT (search snippet)
146. OpenAI Help Center: Memory FAQ, Business version (search snippet)
147. MeitY: Digital Personal Data Protection Rules, 2025, G.S.R. 846(E)
148. ClickHouse blog: ClickHouse acquires Langfuse
149. SiliconANGLE (16 Jan 2026): ClickHouse raises $400M, acquires Langfuse
150. GitHub: Langfuse discussion #11593
151. Helicone blog: Helicone is joining Mintlify (3 Mar 2026)
152. Promptfoo blog: agreed to be acquired by OpenAI (9 Mar 2026)
153. promptfoo LICENSE
154. Traceloop blog: joining ServiceNow (March 2026)
155. CoreWeave 10-Q, Q2 2025 (Weights & Biases acquisition)
156. Weights & Biases pricing URL (redirects to CoreWeave)
157. OpenAI API deprecations
158. GitHub REST API: Arize-ai/phoenix repository metadata
159. LangChain (LangSmith) pricing
160. OpenTelemetry: semantic conventions for generative AI
161. OpenTelemetry semantic-conventions-genai: GenAI spans
162. Langfuse pricing
163. Helicone pricing
164. Braintrust pricing
165. GitHub REST API: langfuse/langfuse repository metadata
166. Arize pricing
167. GitHub REST API: braintrustdata/braintrust-sdk repository metadata
168. GitHub REST API: Helicone/helicone repository metadata
169. GitHub REST API: wandb/weave repository metadata
170. Datadog pricing
171. GitHub REST API: promptfoo/promptfoo repository metadata
172. promptfoo homepage
173. GitHub REST API: confident-ai/deepeval repository metadata
174. GitHub REST API: explodinggradients/ragas repository metadata
175. GitHub REST API: open-telemetry/semantic-conventions repository metadata
176. CXOToday: Check Point completes acquisition of Lakera
177. Check Point Q3 2025 results (press release via Seeking Alpha)
178. Lakera homepage
179. Snyk news: Snyk acquires Invariant Labs
180. Hacker News item 44370620 (Algolia API), Snyk and Invariant Labs
181. Invariant Labs: MCP security notification, tool poisoning attacks (1 Apr 2025)
182. SentinelOne 8-K (Prompt Security acquisition)
183. Wikipedia: Prompt Security (secondary)
184. HiddenLayer homepage
185. Palo Alto Networks FY2025 10-K (Protect AI acquisition)
186. Meta: Llama 4 Community License
187. Hugging Face: meta-llama/Llama-Guard-4-12B model card
188. Hugging Face: meta-llama/Llama-Prompt-Guard-2-86M model card
189. NVIDIA docs: NeMo Guardrails library overview
190. NVIDIA NeMo Guardrails LICENSE.md
191. GitHub REST API: guardrails-ai/guardrails repository metadata
192. GitHub REST API: invariantlabs-ai/mcp-scan repository metadata
193. GitHub: n8n-io/n8n LICENSE.md (Sustainable Use License)
194. n8n pricing
195. Toolradar: n8n pricing 2026 (secondary)
196. AgentDeals: n8n (tracker, secondary)
197. Temporal pricing
198. Zapier MCP
199. GitHub REST API: n8n-io/n8n repository metadata
200. Zapier pricing
201. Make pricing
202. Pipedream pricing
203. Activepieces pricing
204. Windmill pricing
205. GitHub REST API: temporalio/temporal repository metadata
206. GitHub REST API: PrefectHQ/prefect repository metadata
207. Prefect pricing
208. GitHub REST API: apache/airflow repository metadata
209. Retell AI pricing
210. Morph: best AI voice agent platforms 2026 (secondary)
211. Vapi pricing
212. AI Agent Rank: Vapi vs Retell vs Bland 2026 (secondary)
213. Bland AI pricing
214. Bland AI blog: voice AI pricing for enterprise (vendor-authored)
215. Synthflow pricing
216. LiveKit pricing
217. Sarvam AI API pricing
218. Plivo: voice pricing for India (vendor-reported)
219. Plivo pricing
220. Bolna pricing (vendor-reported)
221. Exotel pricing
222. Exotel blog: migrating from Twilio (search summary; not read directly)
223. edesy: Exotel AI voice agent (reseller; search summary)
224. Twilio: voice pricing for India
225. TRAI Direction of 4 May 2024 (DLT-based voice calling)
226. Comments on the draft TCCCPR Third Amendment Regulations 2026 (hosted by TRAI)
227. GitHub REST API: livekit/agents repository metadata
228. GitHub REST API: pipecat-ai/pipecat repository metadata
229. ElevenLabs pricing
230. Deepgram pricing
231. AssemblyAI pricing
232. Cartesia pricing
233. Meta for Developers: WhatsApp Business Platform pricing
234. ChatMaxima: WhatsApp API pricing in India (vendor-reported)
235. Gallabox: WhatsApp pricing overview (vendor-reported)
236. Authkey: WhatsApp API pricing update (secondary)
237. TelecomTalk: WhatsApp Business price reveal (secondary)
238. GitHub: FlowiseAI/Flowise discussion #6727, The Future of Flowise
239. GitHub REST API: langgenius/dify repository metadata
240. Microsoft Learn: Copilot Studio billing and licensing
241. Dify pricing
242. GitHub REST API: langflow-ai/langflow repository metadata
243. GitHub REST API: FlowiseAI/Flowise repository metadata
244. Botpress pricing
245. Voiceflow pricing
246. Stack AI pricing
247. Relevance AI pricing
248. E2E Networks: TIR inference endpoints (vendor-reported)
249. Yotta Shakti Cloud: AI endpoints (vendor-reported)