docs(integrations): copy-paste "connect in 10 lines" recipes for LlamaIndex / CrewAI / AutoGen / LangGraph

Context

Capability goal: 易接入 (easy to integrate) — lower the barrier to "paste a snippet and you're connected" before native per-framework adapters land.

Today only two agent surfaces have first-class integration docs: Claude Code (hooks) and Codex (MCP). For the broader Python agent ecosystem — LlamaIndex, CrewAI, AutoGen, LangGraph — there is no "how do I connect Hebb Mind?" page, even though the generic levers already work today: the MCP stdio server (hebb-mcp) and the REST API (/api/v1, search + ingest + memories). The one example we ship for a non-native framework, examples/05_langchain_adapter.py, is an unimplemented skeleton.

This is a doc-only task on existing surfaces — no new library code required.

Current state

• src/hebb/mcp/server.py exposes four MCP tools over the hebb-mcp stdio entrypoint:
  • write_memory(content, tags, importance) — src/hebb/mcp/server.py:54
  • search_memory(query, top_k) — src/hebb/mcp/server.py:90
  • consolidate() — src/hebb/mcp/server.py:141
  • ingest_conversation(content, format_hint, importance) — src/hebb/mcp/server.py:160
  • hebb-mcp console entrypoint → hebb.mcp.server:main — pyproject.toml:65
• The MCP server is a thin wrapper over the REST API; default service URL is http://localhost:8321 (src/hebb/mcp/server.py:1-10).
• REST endpoints (all mounted under /api/v1, src/hebb/server/app.py:153-160):
  • POST /api/v1/search — src/hebb/server/routers/search.py:16; request body is MemoryQuery with query: str and top_k: int (default 10, 1–100) — src/hebb/models/memory.py:74-99. Response is SearchResponse (results: list[MemorySearchResult], related: list[Memory]) — src/hebb/models/memory.py:113-117.
  • POST /api/v1/memories — write a memory — src/hebb/server/routers/memories.py:74
  • POST /api/v1/ingest — ingest a conversation — src/hebb/server/routers/memories.py:145
• examples/05_langchain_adapter.py is a non-functional skeleton: HebbChatMessageHistory and HebbRetriever both raise NotImplementedError (examples/05_langchain_adapter.py:46, :50, :54, :69). examples/README.md lists it as WIP and names LlamaIndex / CrewAI / AutoGen as wanted-but-missing.
• Existing per-framework docs in the site sidebar (repo_pages/.vitepress/config.mts:314-316) cover only Claude Code, Codex, and a generic MCP Integration page (repo_pages/guide/mcp-integration.md + zh mirror repo_pages/zh/guide/mcp-integration.md). No LlamaIndex / CrewAI / AutoGen / LangGraph page exists.

Proposed approach

Add one docs guide page (EN + zh mirror, per the CLAUDE.md per-language parity rule) titled "Use Hebb Mind from <framework>", with a ~10-line copy-paste snippet for each of:

• LlamaIndex
• CrewAI
• AutoGen
• LangGraph

Each snippet wraps an existing surface — either the MCP tools (search_memory / write_memory via the framework's MCP-tool adapter against the hebb-mcp stdio server) or the REST API (POST /api/v1/search and POST /api/v1/memories) using httpx/requests. Prefer whichever is the lower-friction path per framework; the snippet must be self-contained and runnable against a locally running hebb service (http://localhost:8321).

Then link the new page from:

• the README integrations area (next to the Claude Code / Codex lines around README.md:107-108), and
• examples/README.md (the "Pick your starting point" / contributing section that currently only points at the WIP LangChain skeleton).
• the docs sidebar (repo_pages/.vitepress/config.mts) alongside the existing Claude Code / Codex / MCP entries (EN + zh).

Acceptance criteria

• New EN docs page under repo_pages/guide/ with a copy-paste snippet (~10 lines each) for LlamaIndex, CrewAI, AutoGen, and LangGraph.
• zh mirror page under repo_pages/zh/guide/ with the same four snippets (no mixed-language content per CLAUDE.md).
• Each snippet is runnable against a local hebb service on http://localhost:8321 and uses only existing surfaces (MCP tool names search_memory/write_memory, or REST POST /api/v1/search with {"query": ..., "top_k": ...} and POST /api/v1/memories).
• New page added to the VitePress sidebar (repo_pages/.vitepress/config.mts) for both EN and zh.
• Linked from README.md (integrations section) and examples/README.md.
• No new runtime library code introduced (documentation-only).

Scope / out of scope

• In scope: documentation + sidebar links for the four frameworks against the already-shipped MCP and REST surfaces.
• Out of scope: native Python adapter classes (e.g. completing examples/05_langchain_adapter.py or shipping a langchain/llamaindex integration module). Those are follow-ups; this issue is the low-cost "paste a snippet" bridge that ships first.

References

• reports/audit/core-system-audit-2026-06-07.md
• reports/audit/newuser-experience-audit-2026-06-08.md
• reports/design/capability-gap-roadmap-2026-06-11.md

Filed from the capability-gap roadmap (reports/design/capability-gap-roadmap-2026-06-11.md).