Wiki Knowledge Sources for AgentsMake project wikis readable by coding agents.
Run llmwiki-serve beside an existing LLMWiki, Markdown, or Obsidian folder. Agents retrieve project context over HTTP and MCP. Bridge and chat stay optional.
Skip to content
Problem first
Your project context already lives in local files.
Personal and team workflows often already have useful project knowledge in LLMWiki, Markdown, or Obsidian folders. The first problem is not orchestration; it is making that existing folder readable by an agent without moving private files into a new hosted system.
The default path is intentionally small: run llmwiki-serve beside one folder, verify the cited source view, and connect it to Codex, Claude Code, Copilot-style agents, IDE extensions, or scripts. Bridge, bridge-start, and chat are later modules for specific hurdles after the source works.
Start here when you want to...
- Reuse an existing project wiki without moving the files.
- Give a coding agent project-specific memory on demand.
- Keep evidence, citations, and graph context visible during agent work.
- Decide later whether multi-source bridge, guided setup, or chat UI is worth adding.
Use the folder you already haveKeep project knowledge in its current LLMWiki, Markdown, or Obsidian folder. Serve it read-only instead of migrating it into a hosted RAG app.Understand the source layerSee what
llmwiki-serve reads, what it serves, how redaction and freshness work, and where bridge, chat, RAG, and vector DB boundaries sit.Let agents pull focused contextCoding agents and scripts can query page, search, context, graph, HTTP, and MCP surfaces before planning edits or answering.Add orchestration only when neededUse the bridge for multi-source evidence and runtime-backed answers. Use chat to test connections, citations, traces, and graph context.Minimum useful pathInstall
llmwiki-serve from PyPI, point it at an existing folder or the tiny QuickStart sample, run manifest, query, source-refs, and source-bundle, then serve 127.0.0.1:8765. Use ls or status --json when you need local instance discovery.Public-preview installUse the published, registry-verified packages first:
llmwiki-serve==0.2.10, npx llmwiki-bridge-start@latest resolving to 0.0.3, npx llmwiki-agent-bridge@latest resolving to 0.4.0, and the static llmwiki-chat@0.1.6 artifact. Source checkouts are for bundled fixtures and development.Protocol postureSource access is HTTP/MCP first. A2A source compatibility is opt-in, and bridge runtime surfaces are described as A2A-style and MCP-style compatibility.
Conceptual architecture demo
See the folder-to-agent flow before you install.
Upstream workflows create compatible Markdown or wiki files. llmwiki-serve projects that folder read-only, and agents, bridge, or chat consume the served Knowledge Source. The committed media is conceptual until screenshots are regenerated for the latest QuickStart and logging controls.
First Run At A Glance
Step 1Start your source
Install the PyPI package, then point llmwiki-serve at an existing wiki folder or the tiny QuickStart sample.
Call /health, /manifest, /source-refs, /source-bundle, and /query. Stop here if direct retrieval is enough.
Give the agent the source URL, call POST /query first, and cite returned page, path, and source-ref fields.
Use bridge-start for guided setup, agent-bridge for source fan-out or runtime-backed artifacts, and chat for human inspection.
Start With Source; Add Modules Later
Start herellmwiki-serve
Serves one existing LLMWiki, Markdown, or Obsidian folder as read-only HTTP/MCP context, search, page, graph, and manifest APIs.
Later hurdle: guided setupllmwiki-bridge-startDiscovers existing wiki folders, validates startable sources, starts loopback source servers, and optionally registers them with the bridge.
Later hurdle: companion endpointllmwiki-agent-bridgeConnects one or more sources, gathers cited evidence, optionally delegates synthesis to a runtime, and returns normalized artifacts.
Later hurdle: human previewllmwiki-chatLets humans configure sources and bridge runtimes, ask questions, and inspect citations, graph context, artifacts, and traces.
When To Add Later Modules
| Need | Recommended path |
|---|---|
| One coding agent should read one wiki folder while it works | Run llmwiki-serve and connect the agent directly. |
| You want guided handoff after the source smoke passes | Run npx llmwiki-bridge-start@latest --path /path/to/your/wiki, then keep the direct source URLs or add the bridge. Use status or ls for started-source and bridge-registration state. |
| Several wiki folders must be searched together | Run one llmwiki-serve per folder and connect them through npx llmwiki-agent-bridge@latest; inspect registered sources with sources --probe --json. |
| A service should gather evidence and call a model runtime for a cited answer | Use npx llmwiki-agent-bridge@latest in delegated-runtime or hybrid mode. |
| A human needs to test setup, inspect evidence, or debug traces | Host the static llmwiki-chat@0.1.6 artifact as the browser workbench. |
| You only need package status, release support, or protocol details | Read Release Status, Protocols, and API Reference. |
Optional Module Map
flowchart LR wiki["Wiki folder"] start["llmwiki-bridge-start"] serve["llmwiki-serve"] agent["Host agent"] bridge["Agent bridge"] runtime["Runtime"] chat["llmwiki-chat"] wiki -->|discover/start| start start -->|project| serve start -->|optional register| bridge serve -->|direct retrieval| agent bridge -->|query sources| serve bridge -->|delegate synthesis| runtime chat -->|configure runtime| bridge chat -->|inspect source| serve
| Module | Owns | Does not own |
|---|---|---|
llmwiki-serve | File projection, manifest, context packs, search, read, graph, HTTP, MCP, optional A2A source compatibility. | Wiki authoring, ingestion jobs, model calls, answer synthesis, browser UI. |
llmwiki-bridge-start | Guided discovery, local source startup, optional bridge registration, and smoke-test handoff for existing wiki folders after the source layer works. | Wiki compilation or ingestion, runtime hosting, answer synthesis, replacing the bridge. |
llmwiki-agent-bridge | Source fan-out, runtime profile config, OpenAI-compatible chat completions call, normalized answer artifact, citations, graph, trace. | Reading local files directly, hosting a model, browser source selection UI. |
llmwiki-chat | Source setup, bridge/runtime setup, graph inspection, answer review, run details, citation selection. | Serving wiki files, storing provider secrets, production answer quality. |
llmwiki-docs | Cross-repo first-run path, architecture, protocol posture, release status, operations references. | Package runtime behavior or upstream LLM Wiki specification ownership. |
Choose Your Path
Default pathRun the QuickStart
Exercise manifest, query, source-refs, source-bundle, HTTP, and MCP Streamable HTTP before adding bridge or chat.
Best when the agent can retrieve context and do its own planning, editing, and synthesis.
Guided local setupUse bridge-startBest when you want discovery, source startup, optional bridge registration, and smoke checks from existing local folders.
Hermes, DeepAgents, local runtimesAdd the Agent BridgeUse the bridge when one service should gather source evidence and return a model-backed answer artifact.
Human previewHost the static chat workbenchUse chat when a person needs to test source readiness and inspect citations, graph context, artifacts, or traces.
Private or team networkCheck exposure boundariesReview loopback defaults, private HTTP, CORS, source policy, bearer tokens, and TLS before sharing endpoints.
What To Read Next
| Goal | Page |
|---|---|
| Run the first local path | QuickStart |
| Understand the source layer | llmwiki-serve |
| Understand the data flow | Demo and Data Flow |
| Understand the vocabulary | Core Concepts |
| Decide direct source vs bridge vs chat | Architecture and Runtime Adapters |
| Connect Codex, Claude Code, Copilot-style IDE agents, or scripts | Direct Agent Integrations and AI Tool Support |
| Understand HTTP, MCP, and A2A-style compatibility surfaces | Protocols and API Reference |
| Expose endpoints beyond loopback | Network & Security and Deployment |
| Check package/public-preview status and evidence | Release Status & Compatibility and Evidence |
| Prepare endpoint operations | Deployment, Network & Security, and Troubleshooting |
| Check public-preview support status | Release Status & Compatibility and FAQ |
Public preview
This is independent community tooling for LLM Wiki-style Markdown knowledge folders and agent-readable context. It is not an official project from Andrej Karpathy or any upstream producer named in compatibility examples, and it does not claim certified MCP or A2A conformance.