OpenWiki generates a linked Markdown wiki from a codebase and keeps it current as the code changes. Coding agents read the wiki as context to understand a repository, answer questions, and find information across related projects, and people can read the same pages or browse them as an interactive graph. The wiki lives in the repository as plain Markdown, alongside the grounding and run metadata that OpenWiki maintains.
There are two ways to run it. Inside a coding agent such as Codex, Claude Code, OpenCode, Cursor, Kiro, Pi, or Oh My Pi, OpenWiki uses that agent's model session and repository tools, and the agent writes each page through an MCP lifecycle while OpenWiki owns the page queue, validation, and finalization. As a standalone CLI, it runs its own documentation agent built on Deep Agents, with a choice of model provider, and also supports scheduled updates and a personal wiki mode.
Each repository page tracks its facts as Grounded Claims, each pointing at a versioned piece of source evidence. On an update, OpenWiki checks which evidence changed or disappeared and asks only for the affected facts to be confirmed, rewritten, or retired. A clean update skips model work and leaves the wiki untouched. Generation runs as a queue of independently durable page jobs, so an interrupted run resumes where it stopped.
In personal mode, OpenWiki ingests knowledge from connected sources (Notion, Gmail, Slack, X, web search, Hacker News, local Git repositories, and any MCP server) and synthesizes it into a local wiki. A LangSmith connector instead enriches a code wiki with runtime traces, so the docs reflect how the code behaves when it runs.
Features
- Coding-agent integrations: runs inside Codex, Claude Code, OpenCode, Cursor, GitHub Copilot CLI, IBM Bob, Kiro, Antigravity, Pi, and Oh My Pi using the host's model session
- Grounded Claims: each fact on a repository page points to versioned source evidence, so updates revise only the facts whose evidence changed
- Wiki search for agents: local, read-only MCP tools return ranked results and full sections without making model calls
- Linked workspaces: groups wikis from related repositories (for example a control plane, data plane, and infrastructure) so agents search across a whole service
- Graph visualizer: renders the wiki as an interactive node graph beside a live Markdown reader, and exports a static version for any static host
- Scheduled updates: example CI jobs for GitHub Actions, GitLab CI, and Bitbucket Pipelines open a docs pull request when the wiki changes
- Resumable generation: pages are durable units of work, and native CLI runs can write up to eight pages in parallel, lowering concurrency on provider rate limits
- Mermaid diagrams: adds sequence, ER, state, and flowchart diagrams where they clarify a concept, validates them, and repairs broken ones on the next update
- Open Knowledge Format: emits OKF v0.2 bundles with provenance and verification metadata, so the wiki is portable to OKF-aware tools
- Agent context files: maintains an AGENTS.md and refreshes an existing CLAUDE.md so agents find the wiki
- Personal wiki connectors: ingests Notion, Gmail, Slack, X, web search, Hacker News, local Git repositories, and custom MCP servers
- Ignore rules: a gitignore-style file keeps private or generated paths out of what the agent reads
