bureau
Provides an Outline-compatible API shim supporting documents.* and collections.* RPC methods, Outline envelopes, revision conflict handling, editMode patching, and markdown export, plus an importer that preserves document UUIDs and URL IDs so existing links continue to resolve.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Agent Bureau
Self-hosted tasks + docs for agent fleets. One app, one database, one MCP endpoint.
Agent Bureau merges what Vikunja (kanban/tasks) and Outline (knowledge base) each do half of into a single lightweight service designed for teams where most of the writers are agents.
flowchart LR
H["Humans<br/>browser / CLI"] --> API
A["Agents<br/>MCP clients"] --> MCP["/mcp"]
S["Other services"] --> API["REST /api/v1"]
MCP --> APP
API --> APP["Agent Bureau<br/>one process"]
APP --> DB[("SQLite<br/>WAL")]
APP -->|signed webhooks<br/>retries + replay| W["Your automations,<br/>relays and bots"]Why another tool?
Because we ran Vikunja + Outline with a fleet of coding agents, and three things kept breaking:
Webhook delivery was fire-and-forget. Vikunja retries for a few seconds in memory, then drops. A bot mid-restart means a lost card wake, silently. Agent Bureau writes an event to a transactional outbox in the same commit as the change, then delivers with retries over hours, a full delivery log, and byte-stable replay.
Events didn't say who did it. A bot reacting to its own write is the classic agent echo loop. Every Agent Bureau write carries
actor: {id, type: human | bot | system}— and webhook subscriptions can exclude actors entirely.Tasks and docs lived in different apps, with two auth paths, two MCP servers, and no link between a card and the spec it implements. In Agent Bureau a card↔doc link is a foreign key, and one MCP endpoint serves both.
Everything is append-only auditable: every field change lands in a status history with actor, transport (via: rest | mcp | ui | cli | rule) and rule id.
Related MCP server: KittyClaw
What works today (v0.2)
Tasks: projects, board views with buckets (WIP limits, default/done buckets), labels, assignees, comments, priorities, due dates,
percent_done. Moving a card into the done column marks it done — and syncs every other board view of the project. Safe partial updates (PATCH) that never wipe fields you omit.Docs: collections, nested markdown documents, revisions on every save with optimistic concurrency (409 on stale revision), slug stability with 301 redirects for renamed docs.
Search: SQLite FTS5 (BM25) across tasks and documents with snippets.
Card↔doc links: both directions, with
link.created/link.deletedevents.Agent surfaces: one MCP endpoint (
/mcp, Streamable HTTP, 18 tools) and a clean REST API withbureau_-prefixed tokens (read/writescopes). Bots are first-class users with owners.Webhooks: native signed envelopes (
X-Bureau-Signature), event/project/actor filtering, retry schedule (8 attempts ≈ 5h42m), auto-disable, delivery log, replay.Vikunja compat shim: existing bots keep their Vikunja v1 API calls and
X-Vikunja-Signaturewebhooks — only the base URL and token change. Upstream quirks (full-object-POST field wipes,task.comment.editednaming, numeric error codes) are reproduced and contract-tested.Outline compat shim:
documents.*/collections.*RPC methods with Outline envelopes,lastRevision409s,editModepatching, and markdown export.Importers:
bureau import vikunja|outlinereads the source systems' APIs, preserves ids (tasks keep numeric ids; docs keep uuid+urlId so every old URL resolves), records anid_map, and verifies counts against the source.Automation rules: "when a card enters bucket X, assign user Y / add label Z" — server-side, cascade-free, attributed to
systemwith the rule id.Deploy: one Docker image, one SQLite file. Backups are
cp(plus WAL checkpoint).
Quickstart
git clone https://github.com/poonasor/agent-bureau && cd bureau
npm install
npm run dev # http://127.0.0.1:8080Create an admin and a token:
npx tsx src/main.ts users create --username alice --password "long-passphrase"
npx tsx src/main.ts users set-admin --username alice
npx tsx src/main.ts tokens create --user alice --title first --stdout > token.txt && chmod 600 token.txtcurl -s -H "Authorization: Bearer $(cat token.txt)" http://127.0.0.1:8080/api/v1/projectsDocker
export BUREAU_SECRET_KEY=$(openssl rand -base64 32)
docker compose up -d --build
curl -s http://127.0.0.1:8080/healthThen exec into the container to bootstrap (docker compose exec bureau node dist/main.js …).
Point an agent at it
Any MCP client (Claude Desktop, ZCode, your own agent):
{
"mcpServers": {
"bureau": {
"url": "http://your-host:8080/mcp",
"headers": { "Authorization": "Bearer bureau_…" }
}
}
}One endpoint, both domains: create_task, move_task, search, create_document, link_doc, …
The API in one minute
# projects come with a board, default buckets and a done column
curl -X POST …/api/v1/projects -d '{"title":"Launch"}'
# agents write with PATCH semantics — omitted fields are never wiped
curl -X PATCH …/api/v1/tasks/42 -d '{"percent_done":0.5}'
# moves apply done-bucket rules and fire task.moved webhooks
curl -X POST …/api/v1/tasks/42/move -d '{"view_id":1,"bucket_id":3}'Full reference: docs/api.md.
Coming from Vikunja + Outline?
Your bots don't have to be rewritten. Point them at the compatibility shims —
/compat/vikunja/api/v1 and /compat/outline/api — and import your data with
bureau import vikunja|outline --verify. Tasks keep their ids, document URLs keep resolving,
and Vikunja-signed webhooks keep their HMAC scheme. Details:
docs/compatibility.md and the
migration guide.
Design in 60 seconds
One process + SQLite (WAL,
synchronous=FULL). The outbox is a table; the delivery worker is a loop in the same process. No Redis, no queue. Postgres remains available as a future storage backend if you outgrow a single file (see ADR-0002).Transactional outbox. Change + event commit together or not at all; fan-out and delivery happen after commit. Deliveries store their exact signed bytes, so retries and replays are byte-identical.
Actor tagging everywhere.
human/bot/systemon every event, history row and webhook.Clean-room. No code was copied from Vikunja (AGPL) or Outline (BSL). Agent Bureau is Apache-2.0.
More: docs/architecture.md · all decisions: docs/adr/
Migrating from Vikunja / Outline
Import tooling and an API compatibility layer (existing bots keep working with a token swap) are the next milestones — see docs/roadmap.md and the migration guide.
Status
v0.1 — core platform, stable API surface forming. Expect breaking changes until 1.0, tracked in the roadmap. The UI is intentionally minimal right now (read-only doc view; full board/editor ships in M3) — the API and MCP are the product today.
Contributing
PRs welcome — see CONTRIBUTING.md. Good first issues are tagged
good first issue.
Run make test before pushing; CI runs lint + typecheck + tests on Node 22/24.
License
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Task & board management for AI agents + humans. Kanban, comments, digests via MCP.
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
Shared project wiki for AI agents: read and write pages, next actions, and activity logs over MCP.
Minimal issue tracker: projects, objectives and trackable plans, written by agents over MCP.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceMCP server that provides a real-time project management board with cards, sprints, comments, and docs, enabling coding agents to query and edit project state.2MIT
- AlicenseNot gradedqualityBmaintenanceSelf-hosted kanban board that dispatches AI agent fleets against tickets. The embedded Streamable HTTP /mcp endpoint exposes 7 tools to list projects, read boards and tickets, and create, move and comment tickets from any MCP client.30AGPL 3.0
- AlicenseAqualityCmaintenanceEnables AI agents to manage tasks on a local-first board via MCP, exposing task creation, updates, and queries through a thin adapter over the REST service.7MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents and humans to collaboratively manage kanban boards and Markdown documentation via MCP tools, with stable item keys, revision-safe editing, and full audit trails.AGPL 3.0