first projects to focus on

Below is an **updated, popularity‑and‑impact–weighted** priority list for AgentHub launch. I weighted three things:

* **Popularity signals** (Stack Overflow 2025 survey usage/admiration, GitHub stars/“used by”, etc.). Example: Node/React/Next dominate web usage; PostgreSQL is the most admired/desired DB; GPT‑4o/Claude/Gemini lead model usage. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
* **LLM‑savvy density** (how many of a project’s users are already building with LLMs).
* **AgentHub fit / impact** (how much a high‑quality YAML agent would reduce friction: auth, tool‑calling schemas, pagination, streaming, errors).

For context, AgentHub is a **public registry + spec** of peer‑reviewed YAML agent files; the website and repo show the structure (“Agent Specs”, tutorials). ([[Agent Hub](https://agent-hub-1.netlify.app/)][2], [[GitHub](https://github.com/FIL-Builders/agent-hub)][3])

---

## Tier A — Ship first (1–20)

1. **OpenAI API (Chat/Assistants/Structured Output)** — Most‑used model family (GPT‑4o) + universal target; high leverage for schema/tool‑calling recipes. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
2. **Anthropic Claude** — #2 model usage; strong enterprise pull; tool‑use patterns worth encoding. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
3. **Google Gemini** — #3 model usage; breadth of modalities & Google Workspace tie‑ins. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
4. **Azure OpenAI** — Enterprise gateway to OpenAI; auth/deployment nuances → high impact.
5. **AWS Bedrock** — Model marketplace + enterprise controls; consistent tool schema helps.
6. **Ollama** — Massive OSS traction for local dev; clear OpenAI‑compatible patterns to codify. (150k★) ([[GitHub](https://github.com/ollama/ollama)][4])
7. **vLLM** — De‑facto high‑throughput OpenAI‑compatible hosting; agent spec clarifies params & batching. (55k★) ([[GitHub](https://github.com/vllm-project/vllm)][5])
8. **LangChain** — Largest LLM framework community; agent/tool patterns benefit from canonical YAML. (113k★) ([[GitHub](https://github.com/langchain-ai/langchain)][6])
9. **LangGraph** — Agent workflow graphs; encode best‑practice tool schemas & state handling. ([[GitHub](https://github.com/langchain-ai/langgraph)][7])
10. **LlamaIndex** — Data‑over‑LLMs focus; lots of RAG users need precise connector recipes. (43.7k★) ([[GitHub](https://github.com/run-llama/llama_index)][8])
11. **Next.js + Vercel AI SDK** — Dominant web app stack for AI UIs; 134k★ and “Used by 5m” repos → huge reach. ([[GitHub](https://github.com/vercel/next.js)][9])
12. **FastAPI** — Python API workhorse for AI backends; streaming & auth patterns to standardize. (88k★) ([[GitHub](https://github.com/tiangolo/fastapi)][10])
13. **PostgreSQL (+ pgvector)** — Most admired/desired DB; encode embeddings schema/migrations & RAG recipes. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
14. **Supabase** — Postgres + auth + edge; popular with indie AI builders. (87k★) ([[GitHub](https://github.com/supabase/supabase)][11])
15. **Pinecone** — Managed vector DB used across tutorials; upsert/query/namespace gotchas to fix in YAML.
16. **Qdrant** — 25k★ open‑source vector DB; clear clients, hybrid search to capture. ([[GitHub](https://github.com/qdrant/qdrant)][12])
17. **Chroma** — 21.6k★; “used by” 86k repos; encode collection defaults & persistence pitfalls. ([[GitHub](https://github.com/chroma-core/chroma)][13])
18. **Weaviate** — 14k★; popular with LangChain/LlamaIndex users; GraphQL/REST differences to nail. ([[GitHub](https://github.com/weaviate/weaviate)][14])
19. **Hugging Face (Transformers & Inference API)** — Model hub + simple inference; rate‑limit & model‑param templates help.
20. **Groq API** — Low‑latency inference; streaming/tool‑call quirks deserve a crisp spec.

---

## Tier B — High‑probability traction (21–40)

21. **Mistral API** — Open‑weights + hosted; include function‑calling and RAG settings.
22. **Cohere (Embed/Rerank/Chat)** — Common RAG building blocks; unify embed dims & rerank contracts.
23. **OpenRouter** — Multi‑provider routing; normalize model names, output schemas.
24. **Together AI** — Hosted OSS models; parameters & tokenization differences documented once.
25. **Replicate** — Model‑as‑a‑service; async jobs, polling, webhooks patterns.
26. **Fireworks.ai** — High‑performance serverless inference; OpenAI‑compat plus extras.
27. **Cloudflare Workers AI** — Edge inference/vectorize; durable objects/limits captured in YAML.
28. **Haystack** — Python RAG framework; node configs & evals into a clean agent file.
29. **PydanticAI** — Type‑first agents; showcase structured I/O + tool schemas.
30. **DSPy** — Programmatic prompting; encode signature/optimizer recipes.
31. **CrewAI** — Multi‑agent tasks; tool boundaries & memory spelled out.
32. **AutoGen (Microsoft)** — Multi‑agent framework; roles, safety rails, tool scopes.
33. **Dify** — OSS LLM app platform; connectors & prompt flows standardized.
34. **Langfuse** — LLM observability; logging/traces API spec for quick drop‑ins.
35. **TruLens** — Evals; unify feedback functions & storage.
36. **Arize Phoenix / OpenInference** — Tracing/metrics; schema for spans and attributes.
37. **promptfoo** — Prompt testing; batch/run configs set via YAML.
38. **Temporal** — Durable workflows; best‑practice activities for agent loops.
39. **Dagster** — Data/LLM pipelines; solid IO‑manager & asset patterns.
40. **Ray** — Distributed inference; serve configs, batching & concurrency templates.

---

## Tier C — Retrieval, data, search, and app glue (41–70)

41. **Elasticsearch** — Hybrid/vector search; index mappings, query DSL examples.
42. **Redis (Redis‑Stack/RedisVL)** — Vector + cache; TTL/backfill patterns.
43. **Vespa** — Large‑scale ranking/vector; schema + query profiles.
44. **Milvus** — Vector DB; index types, partitions, filters.
45. **Typesense** — Simple search; collections & synonyms for RAG.
46. **Meilisearch** — Lightweight search; filters/typo‑tolerance patterns.
47. **Neon** — Serverless Postgres; pooling and pgvector perf defaults.
48. **PlanetScale** — Serverless MySQL; read/write splitting for agents.
49. **DuckDB** — Local analytics for evals; file‑backed workflows.
50. **SQLite (+ sqlite‑vss)** — Embedded vector store; persistence & limits.
51. **MongoDB Atlas (Vector Search)** — Embed dims, ANN options, auth.
52. **S3 (AWS)** — Chunk storage; multipart, presigned URLs, lifecycle.
53. **Cloudflare R2** — S3‑compat; auth differences & costs.
54. **Google Drive / Docs** — OAuth scopes, export formats for ingest.
55. **Notion API** — Rich text blocks; pagination & delta sync.
56. **Slack API** — Event & Web API; bot scopes, rate limits, threads.
57. **GitHub REST/GraphQL** — Search/issues/PRs; GraphQL pagination patterns.
58. **Linear API** — Issue/project mgmt; webhook & triage agents.
59. **Confluence API** — Enterprise docs; permissions & content export.
60. **Discord API** — Bot commands, threads, file uploads.
61. **Twilio** — SMS/voice agents; webhook security + rate caps.
62. **Algolia** — Doc search; indices, synonyms, filters.
63. **Firecrawl** — Crawl → clean text; anti‑bot & retry patterns.
64. **Tavily Search** — LLM‑friendly web search; source attribution contract.
65. **Exa (semantic search)** — Research agents; citation payloads.
66. **SerpAPI** — Search results; engine selection & quotas.
67. **Playwright** — Headless browse/extract; auth & anti‑bot recipes.
68. **Browserless/Browserbase** — Hosted headless; job lifecycle.
69. **PDF parsing (Unstructured, LlamaParse)** — Costs/limits & clean text output.
70. **Docstores (Typesense Cloud/Meili Cloud)** — Hosted variants; auth & backups.

---

## Tier D — Crypto + DevOps + product infra (71–100)

71. **Foundry (Ethereum)** — Testing/forking scripts; RPC config patterns.
72. **Hardhat** — Tasks/plugins; multi‑chain deploy flows.
73. **viem** — Typed EVM client; chain configs & account mgmt.
74. **wagmi** — React wallet hooks; connect/sign flows.
75. **WalletConnect** — Session lifecycle; mobile handoff.
76. **ethers.js** — ABI encode/decode; provider/signers.
77. **The Graph** — Subgraphs; query pagination & indexing.
78. **Alchemy** — RPC & data APIs; rate‑limit/backoff.
79. **QuickNode** — Multi‑chain RPC; tracing & websockets.
80. **Tenderly** — Simulation & debugging; TX preview agents.
81. **thirdweb** — Contracts SDK; auth & storage.
82. **Coinbase Cloud (Nodes/Onchain)** — RPC/auth idiosyncrasies.
83. **Kubernetes API** — Job/cron/Ingress recipes for agents.
84. **Docker Engine API** — Build/run workers; security flags.
85. **Helia (IPFS JS)** — Content APIs; pins & blocks.
86. **IPFS HTTP API (Kubo)** — Add/cat/pin; gateway caveats.
87. **web3.storage** — Simple IPFS/Filecoin; auth & CAR files.
88. **NFT.storage** — NFT content pinning; metadata format.
89. **Estuary** — Filecoin pinning; deals pipeline.
90. **Pinata** — IPFS pinning; gateway & limits.
91. **Stripe** — Payments; idempotency keys & webhooks.
92. **Clerk** — Auth; JWT/session patterns for agents.
93. **Auth0** — Enterprise auth; scopes, M2M flows.
94. **Sentry** — Error/event capture; trace links from agents.
95. **PostHog** — Product analytics; event schema for agents.
96. **Trigger.dev** — Background jobs; schedules & retries.
97. **Modal** — Serverless Python; secrets, mounts, concurrency.
98. **Fly.io** — App deploy; volumes & regions.
99. **Railway** — PaaS; environment templates for agents.
100. **Hugging Face Spaces/Inference Endpoints** — Deploy simple tools; queue/timeout patterns.

---

### Why these first?

* **Framework & model gravity.** Node/React/Next lead web frameworks; Python stacks (FastAPI) power AI backends; Postgres is the most admired/desired DB; GPT‑4o/Claude/Gemini dominate usage. Focusing here maximizes adoption at launch. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
* **Vector DB reality.** Qdrant (≈25k★), Chroma (≈21.6k★), and Weaviate (≈14k★) plus Pinecone cover most developer choices; one YAML per store reduces dozens of “how do I upsert/query?” issues. ([[GitHub](https://github.com/qdrant/qdrant)][12])
* **Local & hosted inference.** Ollama (≈150k★) and vLLM (≈55k★) are ubiquitous for OpenAI‑compatible endpoints; a solid spec lets devs swap providers without breaking tools. ([[GitHub](https://github.com/ollama/ollama)][4])
* **Where AgentHub fits.** The repo/site position AgentHub as a **curated, open spec + registry**—exactly what teams need to avoid “LLM hallucinated API calls” and to standardize auth, pagination, function schemas, and streaming behavior. ([[GitHub](https://github.com/FIL-Builders/agent-hub)][3], [[Agent Hub](https://agent-hub-1.netlify.app/)][2])

---

## Suggested launch plan

* **Publish 20–30 Tier‑A agents** (models + LangChain/LangGraph/LlamaIndex + Next.js/FastAPI + 3–4 vectors + Postgres/pgvector). This aligns with the **highest‑usage surfaces** from 2025 SO survey and the **largest OSS repos** above. ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1])
* **Add 10 Tier‑B “glue” agents** (OpenRouter, Replicate, Cohere, Cloudflare Workers AI, Temporal, Langfuse, etc.) to round out real‑world workflows.
* **Keep the registry crisp**: one YAML per API with opinionated “Do/Don’t”, **tool‑calling JSON schema**, **auth scopes**, **rate‑limit/backoff**, **working examples**, and **DESIGN\_NOTES.md**—mirroring the tutorial and existing pages. ([[Agent Hub](https://agent-hub-1.netlify.app/tutorials/getting-started-loading-your-first-agent-file)][15])

If you want, I can **turn the top 25 into ready‑to‑merge YAMLs** following the current AgentHub tutorial format (auth, params, streaming, errors, examples) and open PRs against the repo. ([[Agent Hub](https://agent-hub-1.netlify.app/tutorials/getting-started-loading-your-first-agent-file)][15], [[GitHub](https://github.com/FIL-Builders/agent-hub)][3])

*Notes on evidence:* Popularity signals come primarily from the Stack Overflow 2025 survey (framework usage, DB sentiment, LLM model usage) and from GitHub repo stats for high‑traffic AI infra (Next.js, LangChain, LlamaIndex, Ollama, vLLM, vector DBs). ([[Stack Overflow](https://survey.stackoverflow.co/2025/technology/)][1], [[GitHub](https://github.com/vercel/next.js)][9])

Would you like this list **as a CSV** with columns for “Popularity signal”, “Impact rationale”, and “Owner links” so the team can track outreach and PR status?

[1]: https://survey.stackoverflow.co/2025/technology/ "Technology | 2025 Stack Overflow Developer Survey"
[2]: https://agent-hub-1.netlify.app/ "Agent Hub"
[3]: https://github.com/FIL-Builders/agent-hub "GitHub - FIL-Builders/agent-hub"
[4]: https://github.com/ollama/ollama "GitHub - ollama/ollama: Get up and running with OpenAI gpt-oss, DeepSeek-R1, Gemma 3 and other models."
[5]: https://github.com/vllm-project/vllm "GitHub - vllm-project/vllm: A high-throughput and memory-efficient inference and serving engine for LLMs"
[6]: https://github.com/langchain-ai/langchain "GitHub - langchain-ai/langchain:  Build context-aware reasoning applications "
[7]: https://github.com/langchain-ai/langgraph "GitHub - langchain-ai/langgraph: Build resilient language agents as graphs."
[8]: https://github.com/run-llama/llama_index "GitHub - run-llama/llama_index: LlamaIndex is the leading framework for building LLM-powered agents over your data."
[9]: https://github.com/vercel/next.js "GitHub - vercel/next.js: The React Framework"
[10]: https://github.com/tiangolo/fastapi "GitHub - fastapi/fastapi: FastAPI framework, high performance, easy to learn, fast to code, ready for production"
[11]: https://github.com/supabase/supabase "GitHub - supabase/supabase: The Postgres development platform. Supabase gives you a dedicated Postgres database to build your web, mobile, and AI applications."
[12]: https://github.com/qdrant/qdrant "GitHub - qdrant/qdrant: Qdrant - High-performance, massive-scale Vector Database and Vector Search Engine for the next generation of AI. Also available in the cloud https://cloud.qdrant.io/"
[13]: https://github.com/chroma-core/chroma "GitHub - chroma-core/chroma: Open-source search and retrieval database for AI applications."
[14]: https://github.com/weaviate/weaviate "GitHub - weaviate/weaviate: Weaviate is an open-source vector database that stores both objects and vectors, allowing for the combination of vector search with structured filtering with the fault tolerance and scalability of a cloud-native database."
[15]: https://agent-hub-1.netlify.app/tutorials/getting-started-loading-your-first-agent-file "Getting Started: Loading Your First Agent File | Agent Hub"

## 0) Cross‑provider template (what every AgentHub entry should include)

**A. “Doc Finder” facets (power the site UI + README block in each spec)**

1. **Provider surface**: Direct (OpenAI) / Azure / Google (AI Studio vs Vertex) / AWS Bedrock.
2. **Capability**: Responses/Chat, Tools/Function‑calling, Embeddings, Image/Multimodal, Streaming.
3. **Runtime**: Server (Node/TS, Python), Serverless/Edge (Vercel/Cloudflare), Enterprise (Java, .NET).
4. **Auth**: key header vs cloud credentials (tenant/subscription/profile/role).
5. **Compliance/Region**: sovereign regions, private networking/VPC.
6. **Version selector**: API version string (if applicable), SDK version, model family.

**B. Languages to support (initial)**

* **Primary**: **TypeScript/Node** and **Python** (largest LLM‑dev share).
* **Enterprise add‑ons** (where it moves the needle): **Java** (Azure/AWS), **Go** (Google).
* **Always include cURL** for the canonical REST shape.

**C. Canonical source of truth**

* **REST reference** is the ground truth; SDK snippets are convenience mirrors.
* Each spec page shows: **REST > TS > Python** (plus Java/Go where relevant).

**D. Version & drift controls (baked into AgentHub)**

* Every spec has a `versions:` block (see skeleton below), pinning:

  * `apiVersion` (e.g., `2024-10-21` for Azure OpenAI),
  * `sdkVersions` (per language),
  * `models` (aliases → concrete model IDs),
  * `status` (ga | preview | deprecated | retired),
  * `sunset` (ISO date if known),
  * `docs` (deep links to vendor references).
* A **registry index** resolves `latest` → pinned version for each provider, with **branch‑per‑major** in the repo and **directory‑per‑version** on disk.

**E. Repo layout (multi‑version)**

```
/providers/
  openai/
    v1/           # REST is not date-versioned; we use semantic slices
      agent.yml
      DOCMAP.yml
  anthropic/
    2023-06-01/
      agent.yml
      DOCMAP.yml
    2025-xx-yy/
      ...
  google-gemini/
    v1/           # AI Studio (GenAI SDK)
    vertex-v1/    # Vertex AI
  azure-openai/
    2024-06-01/
    2024-10-21/
  aws-bedrock/
    2024-xx-xx/   # Bedrock Runtime surface stays stable; model IDs vary
```

**F. Spec skeleton (shared)**

```yaml
name: openai
slug: openai
capabilities: [responses, tools, embeddings, vision, streaming]
latest: v1
versions:
  - id: v1
    apiVersion: null           # date or null if not used by provider
    status: ga
    sunset: null
    docs:
      rest: https://platform.openai.com/docs/api-reference
      sdk:
        ts: https://platform.openai.com/docs/api-reference?lang=node.js
        py: https://platform.openai.com/docs/api-reference?lang=python
    models:
      default: gpt-4o          # example alias
      # …
    sdkVersions:
      ts: ">=4.0.0"
      py: ">=1.0.0"
routing:
  choose:
    - facet: runtime
    - facet: capability
    - facet: language
```

---

## 1) OpenAI API — **Responses + Structured Outputs** first

* **Why**: “Responses API + Structured Outputs” is the most robust, schema‑safe way to build tools; deprecation policy exists for models, so we pin models and track the deprecation page. ([[OpenAI Platform](https://platform.openai.com/docs/guides/structured-outputs?utm_source=chatgpt.com)][1])

**Doc Finder (what users see)**

* **Choose capability** → *Responses* (recommended), *Chat Completions* (legacy), *Embeddings*, *Images*.
* **Choose language** → Node/TS, Python (show SDK code), then cURL (REST).
* **Show links**:

  * Structured Outputs guide (JSON Schema) → **OpenAI docs**. ([[OpenAI Platform](https://platform.openai.com/docs/guides/structured-outputs?utm_source=chatgpt.com)][1])
  * API reference (“Responses”, “Embeddings”, “Images”). ([[OpenAI Platform](https://platform.openai.com/docs/guides/structured-outputs?api-mode=responses&utm_source=chatgpt.com)][2])
  * **Model deprecations** (live schedule) → prompt to check before pinning. ([[OpenAI Platform](https://platform.openai.com/docs/deprecations?utm_source=chatgpt.com)][3])

**Client vs Server**

* **Both**: REST is canonical; SDKs (TS/Python) mirror parameters 1:1.

**Version drift strategy**

* OpenAI doesn’t use a dated `api-version` header; we manage **model churn**:

  * Maintain a `models:` table with `status` and `sunset` dates sourced from **Deprecations**. ([[OpenAI Platform](https://platform.openai.com/docs/deprecations?utm_source=chatgpt.com)][3])
  * When “Responses” changes materially, snapshot as `openai/v1` → `openai/v1r2` (minor), update `latest`.

---

## 2) Anthropic Claude — **Messages API** with required API version header

* **Why**: Anthropic requires the `anthropic-version` header; some features use `anthropic-beta`. We’ll expose those choices and pin safe defaults. ([[Anthropic](https://docs.anthropic.com/en/api/messages?utm_source=chatgpt.com)][4])

**Doc Finder**

* **Choose capability** → Messages (tool use, JSON output), Embeddings (if/when applicable).
* **Choose language** → Node/TS, Python; show SDK + cURL.
* **Show links**:

  * **Messages API reference** incl. headers. ([[Anthropic](https://docs.anthropic.com/en/api/messages?utm_source=chatgpt.com)][4])
  * **Versioning guidance** (the `anthropic-version` header). ([[Anthropic](https://docs.anthropic.com/en/api/versioning?utm_source=chatgpt.com)][5])
  * **Examples** page (tool use, streaming). ([[Anthropic](https://docs.anthropic.com/en/api/messages-examples?utm_source=chatgpt.com)][6])

**Client vs Server**

* **Both**: REST is canonical (headers visible), SDKs follow as convenience.

**Version drift strategy**

* Directory per **version header** (e.g., `anthropic/2023-06-01/`).
* Spec pins `anthropic-version: 2023-06-01` (or later) and lists any `anthropic-beta` flags separately with warnings and off‑by‑default behavior. ([[Anthropic](https://docs.anthropic.com/en/api/messages?utm_source=chatgpt.com)][4])

---

## 3) Google Gemini — **AI Studio (GenAI SDK)** and **Vertex AI** as two surfaces

* **Why**: Two official entry points: **Gemini API (AI Studio)** with the **Google GenAI SDK** and **Vertex AI** for GCP‑managed production; each has different auth and model naming. ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs/libraries?utm_source=chatgpt.com)][7], [[Google Cloud](https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference?utm_source=chatgpt.com)][8])

**Doc Finder**

* **Choose surface** → *AI Studio (Gemini API)* or *Vertex AI*.
* **Choose capability** → text/chat, multimodal/vision, structured output, long‑context. ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs?utm_source=chatgpt.com)][9])
* **Choose language** → Node/TS, Python (GenAI SDK), plus cURL.
* **Show links**:

  * **GenAI SDK libraries** (official, GA). ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs/libraries?utm_source=chatgpt.com)][7])
  * **Gemini API overview/models** (AI Studio). ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs?utm_source=chatgpt.com)][9])
  * **Vertex AI inference reference** (request/response shapes). ([[Google Cloud](https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference?utm_source=chatgpt.com)][8])

**Client vs Server**

* **AI Studio**: prioritize **SDK** (Google recommends it), mirror REST via cURL. ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs/libraries?utm_source=chatgpt.com)][7])
* **Vertex**: document **REST + SDKs (Python/Node)**; call out project/region setup.

**Version drift strategy**

* Maintain **two version tracks**: `google-gemini/v1` (AI Studio) and `google-gemini/vertex-v1`.
* Pin **model variants** (e.g., “2.5 Pro/Flash”) in the spec’s `models:` with a `lastVerified` date; link to model catalog. ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs/models?utm_source=chatgpt.com)][10])

---

## 4) Azure OpenAI — **date‑versioned REST** (must pass `api-version`)

* **Why**: Azure OpenAI **requires** `api-version` (YYYY‑MM‑DD). Latest GA is **2024‑10‑21** as of July 2025; also track model retirements. ([[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/reference?utm_source=chatgpt.com)][11])

**Doc Finder**

* **Choose capability** → Chat/Completions (Azure surface), Assistants/Responses (if supported), Embeddings, On‑Your‑Data.
* **Choose language** → TS/Node, Python, **Java** (enterprise prevalent), cURL.
* **Show links**:

  * **REST reference & versioning** (api‑version param). ([[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/reference?utm_source=chatgpt.com)][11])
  * **Latest GA API version** page. ([[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/api-version-lifecycle?utm_source=chatgpt.com)][12])
  * **Model retirements/availability**. ([[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/model-retirements?utm_source=chatgpt.com)][13])

**Client vs Server**

* **Both**: REST is canonical; Azure SDK snippets (TS/Python/Java) are included with explicit `api-version` examples.

**Version drift strategy**

* Directory per API version (e.g., `azure-openai/2024-06-01/`, `azure-openai/2024-10-21/`).
* `latest` alias in registry points at newest GA.
* Spec includes a “**deployment vs model**” explainer (Azure binds to **deployments**, not raw model IDs).

---

## 5) AWS Bedrock — **Bedrock Runtime (InvokeModel / streaming)**

* **Why**: One API surface across many models/providers; users choose by **modelId** and region. Bedrock has consistent **InvokeModel** and **InvokeModelWithResponseStream** calls via REST and SDKs. ([[AWS Documentation](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html?utm_source=chatgpt.com)][14])

**Doc Finder**

* **Choose capability** → text/chat, tool use (where model supports), embeddings, image/multimodal.
* **Choose language** → Python (boto3), Node/TS (AWS SDK v3), **Java**/.NET (enterprise), cURL.
* **Show links**:

  * **Bedrock Runtime InvokeModel** (REST). ([[AWS Documentation](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html?utm_source=chatgpt.com)][14])
  * Representative **SDK reference** (e.g., .NET or v3 clients) to emphasize parity. ([[AWS Documentation](https://docs.aws.amazon.com/sdkfornet/v3/apidocs/items/BedrockRuntime/MBedrockRuntimeInvokeModelInvokeModelRequest.html?utm_source=chatgpt.com)][15])
  * **Model ID catalog** (link to Bedrock model IDs and region support) surfaced in the spec (user picks model by drop‑down).

**Client vs Server**

* **Both**: Emphasize **SDKs** for auth/regions, but keep cURL REST for parity.

**Version drift strategy**

* Bedrock’s API is steady; **models rotate**. Maintain a versioned **model map**:

  * `models: { claude-*, mistral-*, meta-*, amazon-* }` with region lists and last‑verified date.
  * Expose **guardrails/prompt‑resource** pointers in the spec where supported (kept optional). ([[AWS Documentation](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html?utm_source=chatgpt.com)][14])

---

## 6) Site UX: “Doc Finder” block (component spec)

* A collapsible **wizard** on each provider page (and a global finder) using the shared facets.
* Emits: **(a)** deep links to official docs **(b)** copy‑ready code snippet for the chosen language **(c)** the exact **AgentHub YAML** slice (tool schema) pre‑filled with version/model.
* Keep a `DOCMAP.yml` per provider to drive the wizard:

```yaml
surfaces:
  - id: rest
    label: REST (cURL)
    docs: https://...
  - id: sdk-ts
    label: Node/TypeScript
    docs: https://...
  - id: sdk-py
    label: Python
    docs: https://...
capabilities:
  - id: responses
    label: Responses / Messages
  - id: tools
    label: Tool use / Function calling
models:
  - alias: default
    id: gpt-4o
    status: ga
    sunset: null
```

---

## 7) Version governance (automatable, minimal toil)

**A. Pin + monitor**

* Pin `apiVersion`, `sdkVersions`, and `models` per provider/version directory.
* Add a nightly CI job to **diff canonical docs** (status codes only) and compare pinned values to vendor “what’s new / deprecations” pages:

  * OpenAI **Deprecations**. ([[OpenAI Platform](https://platform.openai.com/docs/deprecations?utm_source=chatgpt.com)][3])
  * Anthropic **Versioning**. ([[Anthropic](https://docs.anthropic.com/en/api/versioning?utm_source=chatgpt.com)][5])
  * Google Gemini **libraries/models**. ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs/libraries?utm_source=chatgpt.com)][7])
  * Azure **API version lifecycle** + **model retirements**. ([[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/api-version-lifecycle?utm_source=chatgpt.com)][12])
  * AWS **Bedrock Runtime** (check for param additions). ([[AWS Documentation](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html?utm_source=chatgpt.com)][14])

**B. Compatibility tests (per provider/version)**

* Local schema tests (no internet): validate our YAML tool schemas against **known request/response** shapes.
* “Smoke snippets” in examples folder (excluded by default) that developers can run with their keys.

**C. Deprecation flow**

* When upstream marks **deprecated/retired**, flip `status`, add `sunset`, and move `latest` pointer only after an example build turns green.
* Show a **deprecation banner** in the site for any spec whose status ≠ `ga`.

---

## 8) What we’ll ship for each of the five (deliverables checklist)

* **Versioned `agent.yml`** with tool schemas for the common capabilities.
* **`DOCMAP.yml`** powering the Doc Finder (per language, capability, surface).
* **Examples**: REST (cURL), TS, Python (and Java/Go where relevant).
* **Design notes**: Auth, rate limiting, pagination/streaming, error shapes, idempotency, retries.
* **Version notes**: What changed since the previous directory; migration tips.

---

### Quick references used above

* OpenAI Structured Outputs & Responses; Model deprecations. ([[OpenAI Platform](https://platform.openai.com/docs/guides/structured-outputs?utm_source=chatgpt.com)][1])
* Anthropic Messages API & versioning header. ([[Anthropic](https://docs.anthropic.com/en/api/messages?utm_source=chatgpt.com)][4])
* Google Gemini GenAI SDK & API/Vertex docs; model catalog. ([[Google AI for Developers](https://ai.google.dev/gemini-api/docs/libraries?utm_source=chatgpt.com)][7], [[Google Cloud](https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference?utm_source=chatgpt.com)][8])
* Azure OpenAI REST versioning; latest GA API (2024‑10‑21); retirements. ([[Microsoft Learn](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/reference?utm_source=chatgpt.com)][11])
* AWS Bedrock Runtime InvokeModel (+ sample SDK ref). ([[AWS Documentation](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html?utm_source=chatgpt.com)][14])

---

If you want, I can turn this into **five ready‑to‑commit folders** with `agent.yml`, `DOCMAP.yml`, and example snippets, matching the structure above.

[1]: https://platform.openai.com/docs/guides/structured-outputs?utm_source=chatgpt.com "Structured model outputs - OpenAI API"
[2]: https://platform.openai.com/docs/guides/structured-outputs?api-mode=responses&utm_source=chatgpt.com "structured outputs or responses"
[3]: https://platform.openai.com/docs/deprecations?utm_source=chatgpt.com "Deprecations - OpenAI API"
[4]: https://docs.anthropic.com/en/api/messages?utm_source=chatgpt.com "Messages"
[5]: https://docs.anthropic.com/en/api/versioning?utm_source=chatgpt.com "Versions"
[6]: https://docs.anthropic.com/en/api/messages-examples?utm_source=chatgpt.com "Messages examples"
[7]: https://ai.google.dev/gemini-api/docs/libraries?utm_source=chatgpt.com "Gemini API libraries | Google AI for Developers"
[8]: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference?utm_source=chatgpt.com "Generate content with the Gemini API in Vertex AI"
[9]: https://ai.google.dev/gemini-api/docs?utm_source=chatgpt.com "Gemini API | Google AI for Developers"
[10]: https://ai.google.dev/gemini-api/docs/models?utm_source=chatgpt.com "Gemini models | Gemini API | Google AI for Developers"
[11]: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/reference?utm_source=chatgpt.com "Azure OpenAI in Azure AI Foundry Models REST API ..."
[12]: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/api-version-lifecycle?utm_source=chatgpt.com "Azure OpenAI in Azure AI Foundry Models API lifecycle"
[13]: https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/model-retirements?utm_source=chatgpt.com "Azure OpenAI in Azure AI Foundry Model Retirements"
[14]: https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html?utm_source=chatgpt.com "InvokeModel - Amazon Bedrock - AWS Documentation"
[15]: https://docs.aws.amazon.com/sdkfornet/v3/apidocs/items/BedrockRuntime/MBedrockRuntimeInvokeModelInvokeModelRequest.html?utm_source=chatgpt.com "AmazonBedrockRuntimeClient.InvokeModel (InvokeModelRequest)"

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

first projects to focus on #4

Tier A — Ship first (1–20)

Tier B — High‑probability traction (21–40)

Tier C — Retrieval, data, search, and app glue (41–70)

Tier D — Crypto + DevOps + product infra (71–100)

Why these first?

Suggested launch plan

0) Cross‑provider template (what every AgentHub entry should include)

1) OpenAI API — Responses + Structured Outputs first

2) Anthropic Claude — Messages API with required API version header

3) Google Gemini — AI Studio (GenAI SDK) and Vertex AI as two surfaces

4) Azure OpenAI — date‑versioned REST (must pass `api-version`)

5) AWS Bedrock — Bedrock Runtime (InvokeModel / streaming)

6) Site UX: “Doc Finder” block (component spec)

7) Version governance (automatable, minimal toil)

8) What we’ll ship for each of the five (deliverables checklist)

Quick references used above

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Uh oh!

first projects to focus on #4

Description

Tier A — Ship first (1–20)

Tier B — High‑probability traction (21–40)

Tier C — Retrieval, data, search, and app glue (41–70)

Tier D — Crypto + DevOps + product infra (71–100)

Why these first?

Suggested launch plan

0) Cross‑provider template (what every AgentHub entry should include)

1) OpenAI API — Responses + Structured Outputs first

2) Anthropic Claude — Messages API with required API version header

3) Google Gemini — AI Studio (GenAI SDK) and Vertex AI as two surfaces

4) Azure OpenAI — date‑versioned REST (must pass api-version)

5) AWS Bedrock — Bedrock Runtime (InvokeModel / streaming)

6) Site UX: “Doc Finder” block (component spec)

7) Version governance (automatable, minimal toil)

8) What we’ll ship for each of the five (deliverables checklist)

Quick references used above

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions

4) Azure OpenAI — date‑versioned REST (must pass `api-version`)