LLM Wiki Serve

llmwiki-serve is the read-only Knowledge Source server for the LLMWiki toolchain. Point it at an existing Markdown, Obsidian-style, or LLMWiki folder and it exposes the same local projection as CLI commands, HTTP endpoints, MCP-style JSON-RPC tools, an MCP Streamable HTTP endpoint, and optional A2A-style compatibility endpoints.

Use it when:

• You want local, agent-readable context from files you already own.

• A coding agent, IDE agent, script, or workbench can retrieve cited evidence and do its own planning or answer synthesis.

• You need the source layer that llmwiki-agent-bridge and llmwiki-chat can call.

In the toolchain, llmwiki-serve is the first layer: it reads your wiki and serves evidence. llmwiki-agent-bridge is optional answer-synthesis escalation, and llmwiki-chat is the browser workbench for source, graph, runtime, and trace inspection.

It is not a full-stack RAG app: it does not crawl the web, build a hosted vector store, run a model, synthesize final answers, or mutate your wiki.

Examples | Architecture | OpenAPI contract | Release checklist | Docs portal | Release status | Contributing | Security | Support | Changelog

Public-preview note: source-checkout usage is the supported first-run path. Package-install commands apply after the first PyPI release is published.

Demo

Watch the docs demo to see llmwiki-serve project an already-existing LLMWiki, Markdown, or Obsidian-style folder as a read-only Knowledge Source.

Related MCP server: personal-llm-wiki-kit

10-Minute Quick Start

Install uv and use Python 3.11 or newer:

uv --version
uv python install 3.11

Run the bundled sample wiki from a source checkout:

git clone https://github.com/knowledge-bridge-labs/llmwiki-serve.git
cd llmwiki-serve
uv sync --extra dev

uv run llmwiki-serve manifest ./examples/sample-wiki
uv run llmwiki-serve query ./examples/sample-wiki "release readiness"
uv run llmwiki-serve serve ./examples/sample-wiki --host 127.0.0.1 --port 8765

In another terminal, query the local server:

curl -s http://127.0.0.1:8765/manifest

curl -s http://127.0.0.1:8765/query \
  -H 'content-type: application/json' \
  -d '{"query":"release readiness","limit":4}'

On Windows PowerShell, use curl.exe explicitly or Invoke-RestMethod with a PowerShell object body. Plain curl may resolve to PowerShell's Invoke-WebRequest alias and handle JSON quoting differently.

You have succeeded when manifest returns page/source metadata and query returns a context pack with cited pages from examples/sample-wiki. Point the same commands at your own Markdown folder when you are ready:

uv run llmwiki-serve query /path/to/wiki-folder "what should an agent know?"
uv run llmwiki-serve serve /path/to/wiki-folder --host 127.0.0.1 --port 8765

After the first package release is published, install the CLI with one of:

uv tool install llmwiki-serve
# or
pipx install llmwiki-serve

What It Serves

llmwiki-serve is a protocol layer over a local knowledge folder. It builds a read-only projection from Markdown pages, links, headings, tags, front matter, source references, and optional sidecar graph facts.

Use it for Obsidian, Logseq, Foam, Dendron, Quartz, native LLMWiki folders, and generated Markdown knowledge bases that fit the documented folder contract.

Do not use it as a wiki compiler, crawler, authoring tool, hosted RAG application, vector database, model runtime, answer synthesizer, or certified MCP/A2A implementation.

How It Fits

flowchart LR
  wiki["existing Markdown or LLMWiki folder"]
  serve["llmwiki-serve<br/>read-only Knowledge Source"]
  direct["Codex / Claude Code / Copilot<br/>direct context use"]
  bridge["llmwiki-agent-bridge<br/>runtime synthesis escalation"]
  chat["llmwiki-chat<br/>graph and trace workbench"]

  wiki --> serve
  serve --> direct
  serve --> bridge
  serve --> chat

Direct Agent Skill

Coding agents can use llmwiki-serve directly when a trusted local server is already running. Set the server URL in your project instructions or local environment:

export LLMWIKI_SERVE_URL=http://127.0.0.1:8765

Then instruct Codex, Claude Code, Copilot, or another local client to:

1. Call POST /query first for the task-specific context pack.

2. Use /search, /read/{page_id}, /graph, /mcp, or /mcp/stream only for follow-up inspection.

3. Treat returned pages as source evidence, not as generated final answers.

Do not hard-code private hosts, ports, credentials, or bearer tokens in committed agent instructions. Reusable Codex, Claude Code, and Copilot direct client examples live in the llmwiki-agent-bridge repository under integrations/; after public transfer, start with https://github.com/knowledge-bridge-labs/llmwiki-agent-bridge/tree/main/integrations. In local sibling checkouts, use ../llmwiki-agent-bridge/integrations/README.md.

Escalate to llmwiki-agent-bridge when the agent should call a single model-backed answer endpoint instead of managing source retrieval itself. Escalate to llmwiki-chat when a human needs to inspect graph context, runtime selection, traces, and cited answers interactively.

Serving Surface

All entry points use the same read-only service behavior.

Agents should call llmwiki_context first for a single grounded question. Agents that coordinate host-owned RAG or multi-source orchestration should also inspect llmwiki_source_bundle to discover the stable source identity, projection signature, raw-origin metadata, and opaque source references. Search, read, graph, and source-ref tools are follow-up tools for focused inspection.

The generated FastAPI OpenAPI contract is committed at docs/openapi.json. It covers the default HTTP and MCP-style JSON-RPC compatibility surface; the mounted MCP Streamable HTTP ASGI app is served at runtime, and A2A-style schemas are available only when the app is created with A2A compatibility enabled.

What It Reads

• Generic Markdown wikis with hot.md, index.md, overview.md, and topic pages.

• Obsidian-style wikilinks such as [[Process Page]] and Markdown links to other .md pages.

• YAML front matter fields such as id, title, status, review_state, source_refs, tags, and updated_at.

• Folder-level graph structure from pages, headings, links, tags, and source references.

• Optional sidecar graph facts from graph/graph.json.

• Source-bundle metadata that identifies one served knowledge source, its current projection signature, visible source refs, and metadata-only raw-origin hints. Raw files remain owned by the operator or host RAG layer; llmwiki-serve does not read or expose arbitrary binary source content.

Named producer repositories in the architecture guide are compatible Markdown output targets, not endorsed integrations or per-release support claims. llmwiki-serve reads their generated or stored Markdown when it matches the native folder contract or a supported adapter shape.

Compared With

Python API

The documented Python import surface is:

from llmwiki_serve import LlmWikiService, create_app

LlmWikiService owns manifest, source-bundle, source-ref, context, search, read, graph, and refresh behavior. create_app builds the FastAPI app for embedding the HTTP, MCP-style JSON-RPC, MCP Streamable HTTP, and optional A2A-style compatibility surfaces. Other package modules are implementation details unless documented here.

Safety Defaults

• The selected source folder is immutable input. No source Markdown is rewritten, normalized, migrated, annotated, or uploaded by the server.

• CLI manifest and query build a fresh projection for each process. Long-running serve instances cache an in-memory projection and refresh it on the next request when Markdown, Org, adapter marker/config, or graph/graph.json source files change.

• Draft and unpublished pages are withheld by default from read, search, context, and graph responses. Visibility blocks explicit non-serving markers: draft: true, published: false, publish: false, draft/proposed/needs_review review_state values, and status values such as draft, proposed, needs_review, blocked, unpublished, private, hidden, embargoed, confidential, internal, or withheld. Other lifecycle or maturity status values are served by default. HTTP and MCP tool include_drafts=true is honored only when --allow-drafts or create_app(..., allow_drafts=True) is used. A2A-style compatibility endpoints are disabled by default; when enabled, message:send always builds approved-only context.

• Network manifest responses omit the local wiki root path. The CLI manifest is local operator output and includes the root path.

• The default HTTP CORS policy allows local browser origins on localhost, 127.0.0.1, and IPv6 localhost [::1]; it is not a wildcard. Explicit --cors-origin values replace the default local allowlist.

• Symlinked Markdown/Org source files, symlinked adapter marker/config files, and symlinked graph/graph.json sidecars are ignored by default so the served source tree stays within the selected wiki root.

Review SECURITY.md before exposing a wiki beyond a trusted local environment. Use SUPPORT.md for issue routing and compatibility report expectations.

Repository Structure

Project Posture

llmwiki-serve is independent community tooling for LLM Wiki-style Markdown knowledge folders. It is Apache-2.0 licensed and is not an official project from Andrej Karpathy or any upstream producer named in compatibility examples.

This repository is in public source-checkout preview. Source-checkout usage is the supported path today. Package-install commands apply after the first PyPI release is published. Use the hosted docs and Release Status & Compatibility matrix for the current package and protocol posture.

The current protocol surface is HTTP plus MCP-style JSON-RPC, MCP Streamable HTTP, and opt-in A2A-style message shapes. The Streamable HTTP endpoint uses the official MCP Python SDK FastMCP transport; the compatibility endpoints are local agent and harness surfaces, not a claim of A2A certification, exhaustive runtime feature completeness, or upstream producer certification.

Validation

For a quick source-checkout smoke:

uv run python scripts/check_third_party_notices.py
uv run python scripts/export_openapi.py --check
uv run python scripts/release_smoke.py

For a release-oriented local gate:

uv run ruff format --check .
uv run ruff check .
uv run mypy src
uv run pytest -q
uv build
uv run python scripts/release_smoke.py --wheel dist/*.whl --sdist dist/*.tar.gz

The release smoke checks the bundled sample wiki through CLI, HTTP, MCP-style JSON-RPC, MCP Streamable HTTP, and opt-in A2A-style message shapes, including draft filtering, local-only CORS, MCP error redaction, source immutability, source distribution contents, OpenAPI contract freshness, and packaged wheel CLI installation.

Optional validation paths are documented in docs/release.md: real local-server curl checks, pinned public upstream sample snapshot smoke, and generated candidate sample artifacts. These are compatibility probes for the current serving contract. They do not certify upstream producer versions, full MCP/A2A protocol support, private wiki safety, live network deployment, authentication, TLS, or every application-specific Obsidian/Logseq/Foam/Dendron/Quartz feature.

Project Documents

• Docs portal

• Release Status & Compatibility

• Architecture

• Examples

• Release checklist

• Contributing

• Security

• Support

• Code of Conduct

• Changelog

License

Apache-2.0. See LICENSE.