kiwiki

kiwiki is a self-hosted Agent Harness and Markdown knowledge base for humans and AI agents. Notes stay as regular files on disk, the web UI gives you a searchable wiki, and the built-in MCP server lets tools such as Claude Code, Codex, OpenCode, Cursor, ChatGPT, OpenClaw, Hermes, and any MCP client read and write the same knowledge base in parallel.

No hosted account is required. Your Markdown files are the source of truth.

Visit kiwiki.xyz for the project website, FAQ, agent integration guides, and hosting options.

Screenshots

Related MCP server: Bruin

Features

• Multi-Agent Harness — Central knowledge base for Claude Code, Codex, OpenCode, Cursor, ChatGPT, OpenClaw, Hermes, and any MCP-compatible agent. All agents read and write the same wiki in parallel.

• Markdown files with YAML frontmatter — Your file system is the source of truth. No proprietary database, no vendor lock-in, simple to back up.

• 100 % privacy — Self-hosted on your infrastructure. No cloud, no telemetry, no vendor lock-in.

• Per-user isolated wiki folders under /data/<username>/ with role-based access (read / write / admin).

• SQLite FTS5 full-text search — Search thousands of Markdown files in milliseconds, including via MCP from your AI. Special prefix tag:<value> filters by tag.

• Responsive web UI — FastAPI with Jinja2, HTMX, and Toast UI Editor. Works from 4K desktop to mobile, with WCAG 2.2 AA-oriented accessibility (see docs/ui-accessibility.md).

• Streamable HTTP MCP endpoint at /mcp (legacy HTTP/SSE at /mcp/sse) with OAuth 2.1 authorization-code flow for ChatGPT-style connectors.

• Docker Compose and Helm — One command to start. Kubernetes-ready.

Web UI highlights (v2.2)

• Click-to-create new note: The "Neue Notiz" button prompts for a filename instead of overwriting notes/neue-notiz.md.

• Tag search: Click any tag chip in a note to run tag:<value> search across your workspace.

• Unsaved-changes guard: Editor warns before navigation when content has changed.

• Keyboard-friendly tree: role="tree" with aria-level/aria-expanded, double-click to rename, inline edit.

• Mobile UX: 44 px touch targets, IOS-zoom-safe 16 px inputs, sidebar Esc + focus return, action-sheet context menu.

• Skip-Link: Visible on first Tab press — jumps straight to the main content area.

• Reduced-motion safe: All animations honor prefers-reduced-motion — including login and editor.

Quick Start

git clone https://github.com/natorus87/kiwiki.git
cd kiwiki
cp .env.example .env
docker compose up -d

Open the web UI:

http://localhost:8082

Use the API key configured in KIWIKI_USERS.

Configuration

All runtime configuration is done through environment variables.

Example:

KIWIKI_DATA_DIR=/data
KIWIKI_USERS=admin:<admin-api-key>:admin,writer:<writer-api-key>:write,reader:<reader-api-key>:read
KIWIKI_BASE_URL=https://kiwiki.example.com
KIWIKI_TRUST_PROXY=true
KIWIKI_CORS_ORIGINS=https://kiwiki.example.com
KIWIKI_OAUTH_TOKEN_SECRET=<random-token-signing-secret>

Generate strong keys with:

openssl rand -hex 24

Do not commit real API keys, OAuth secrets, .env files, or local wiki data.

User Model

Each user gets a separate wiki root:

/data/admin/
/data/alice/
/data/bob/

Users cannot read or search another user's files. Web UI requests, REST API calls, search, and MCP tools all run in the authenticated user's namespace.

Roles:

ChatGPT and MCP

kiwiki exposes MCP over Streamable HTTP:

https://kiwiki.example.com/mcp

For ChatGPT custom connectors, configure the MCP endpoint as the connector URL. If OAuth is enabled, ChatGPT discovers:

/.well-known/oauth-protected-resource/mcp
/.well-known/oauth-authorization-server/mcp

The OAuth flow uses:

• Authorization code with PKCE

• Signed access tokens

• Refresh tokens

• The resource parameter expected by MCP clients

• Client ID Metadata Document style client IDs used by ChatGPT

For public deployments, set a stable KIWIKI_OAUTH_TOKEN_SECRET. This keeps connector tokens valid across container restarts while still allowing API-key rotation to revoke access.

Direct bearer-token access is also supported:

curl https://kiwiki.example.com/mcp \
  -H "Authorization: Bearer <api-key>" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

Usage as AI Memory

Configure your AI tools to use kiwiki as persistent memory. The following instruction works for ChatGPT (Custom Instructions → Personalization), Claude (Personalization), and coding agents:

Use kiwiki as my persistent memory. When asked about projects, decisions, recurring topics, personal preferences, or work context, first briefly check kiwiki. Use existing notes as context.

Save new important information in kiwiki when it might be useful later: preferences, decisions, project knowledge, workflows, important facts, and open items. Prefer to update existing files rather than creating new ones. Organize according to the existing structure: /projects, /decisions, /notes, /shared and /user. Write short Markdown notes with frontmatter. Do not delete anything without explicit instruction. Also ask in longer chats whether you should save something.

For coding agents (Claude Code, Codex, OpenCode, Cursor), add this instruction to your project's AGENTS.md, CLAUDE.md, or equivalent configuration file.

Agent Harness Setup

When using kiwiki as the Agent Harness for your project, instruct coding agents to connect to kiwiki via MCP. Add the following block to your project's AGENTS.md or CLAUDE.md:

This project uses kiwiki as its Agent Harness and persistent memory.

MCP connection — Connect to the kiwiki MCP server using the appropriate command for your tool:

• Claude Code: claude mcp add kiwiki http://localhost:8082/mcp --header "Authorization: Bearer <api-key>"

• Codex: codex mcp add kiwiki http://localhost:8082/mcp --header "Authorization: Bearer <api-key>"

• OpenCode: configure MCP server in opencode.json

• Cursor: configure MCP server in .cursor/mcp.json

Once connected, use kiwiki as persistent memory (see usage instruction above).

This ensures every coding agent working on the project automatically connects to the shared knowledge base.

MCP Tools

The MCP server exposes tools for common wiki workflows, grouped by required role:

Read (any role)

read_file · read_many · read_lines · fetch · search · grep · find · list_files · list_all_files · file_info · read_index · recent_files · backlinks · related_files · search_status · whoami

Write (write role or admin)

write_file · write_many · chunked_write · edit · append_file · create_note · upsert_note · update_frontmatter · preview_edit · replace_many · build_index · reindex_all · tag_index · move_file

For autonomous agents, prefer write_many when updating several files and chunked_write when a large file or flaky client payload limit makes a single write_file / append_file call unreliable. Ordinary create, update, append, and index-refresh operations do not require confirmation; clients should ask before deleting files or running destructive reorganizations. chunked_write keeps temporary chunks in process memory for KIWIKI_MCP_UPLOAD_TTL_SECONDS (default 3600) and limits staged uploads with KIWIKI_MCP_MAX_UPLOAD_BYTES (default 10485760) and KIWIKI_MCP_MAX_UPLOAD_CHUNKS (default 1000).

Admin (admin role only)

validate_wiki · delete_file · sort

Keyboard Shortcuts

See docs/ui-accessibility.md for the full accessibility model.

Architecture

The kiwiki frontend is intentionally framework-free — server-rendered Jinja2 templates plus HTMX for partial swaps, plus a small vanilla-JS layer (app/static/kiwiki.js) for interactive widgets (sidebar, tree, dialogs, toasts). All styles live in a single app/static/kiwiki.css with one :root token source. See docs/architecture.md for the layout, request flow, namespaces, and helper conventions before contributing frontend changes.

Local Development

Create a virtual environment and install dependencies:

python3.12 -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements-dev.txt
npm ci

Run the app:

KIWIKI_DATA_DIR=./data \
KIWIKI_USERS="admin:dev-key:admin" \
KIWIKI_BASE_URL="http://127.0.0.1:8080" \
KIWIKI_TRUST_PROXY=false \
uvicorn app.main:app --host 127.0.0.1 --port 8080 --reload

Build the frontend motion bundle:

npm run build:motion

Run checks:

ruff check app tests
pytest -q
npm audit --audit-level=high
docker build -t kiwiki:test .

Docker

The default Compose file builds the local image and serves the app on port 8082:

docker compose up -d
docker compose logs -f kiwiki

Persistent data is mounted at:

./data:/data

For public deployments, move real secrets into .env or your secret manager.

Helm

A Helm chart is available under charts/kiwiki.

Install example:

helm upgrade --install kiwiki ./charts/kiwiki \
  --set env.KIWIKI_USERS="admin:<admin-api-key>:admin" \
  --set env.KIWIKI_BASE_URL="https://kiwiki.example.com" \
  --set env.KIWIKI_OAUTH_TOKEN_SECRET="<random-token-signing-secret>"

Review charts/kiwiki/values.yaml before deploying to production.

Repository Hygiene

The repository includes:

• GitHub Actions CI for Ruff, Pytest, frontend build, and Docker build

• Dependabot configuration for Python, npm, and GitHub Actions

• Issue and pull request templates

• Security policy

• MIT license

Ignored local artifacts include .venv/, node_modules/, data/, Python caches, test caches, and local agent configuration.

Security

See SECURITY.md.

Important operational rules:

• Use strong random API keys.

• Set KIWIKI_TRUST_PROXY=true behind HTTPS.

• Restrict KIWIKI_CORS_ORIGINS in production.

• Set KIWIKI_OAUTH_TOKEN_SECRET for public MCP/OAuth deployments.

• Do not publish local wiki data or deployment secrets.

License

MIT. See LICENSE.