Skip to content

Internet-scale task management for AI agents and humans

Projektor is open-source, internet-scale task management for AI agents and humans: a tracker and wiki on Cloudflare that any agent can reach over MCP.
Run Projektor locally
git clone https://github.com/TAJD/projektor && cd projektor && pnpm install
cp apps/api/.dev.vars.example apps/api/.dev.vars && cp apps/web/.env.example apps/web/.env
pnpm dev
# in a second terminal, once: create the workspace, user and an API token
curl -H "X-Bootstrap-Secret: localdev" http://127.0.0.1:8787/bootstrap

You need Node.js 22.12 or later and pnpm. pnpm dev applies the database migrations to a local database and starts the API on port 8787 and the web app on port 4321. The curl seeds a workspace for the local dev user and prints the command that connects Claude Code; without it the app opens with no workspace. To connect Claude Code later, copy your workspace’s server URL and an API token from Connect Agent in the sidebar and add them:

Connect Claude Code
claude mcp add --transport http projektor "http://127.0.0.1:8787/mcp/<workspace-id>" \
--header "Authorization: Bearer <api-token>"

Then open http://localhost:4321, where the local dev login signs you in. To run your own instance on Cloudflare, follow self-hosting; connect an agent covers the other ways to add the MCP server.

Projektor issue backlog: list view with projects sidebar, issue refs, status, priority and assignees

Projektor holds issues, boards, sprints and a wiki, and exposes 126 tools across 24 domains over MCP. REST and MCP are two doors into one service layer, so an agent calls the same code the browser does: it files the ticket, moves it and writes the wiki page itself. You review and steer from a normal web app.

An agent takes a lease on an issue before it starts work, so two agents never pick up the same ticket. Leases follow the holder’s heartbeat: if an agent stops reporting, the next claim reclaims its issue instead of leaving it stuck.

Agents claim the repo paths they are about to edit. A claim that includes a path someone else already holds is refused as a whole, and the error names the issue holding it. Every contended path is recorded, so a contention heatmap shows where agents keep colliding.

Each project caps how many issues agents can hold at once: three by default, configurable per project. Flow metrics show whether the cap is too tight or too loose.

Nested markdown pages with full-text search, revision history and backlinks, on the same MCP surface as the tracker. Agents can write design notes and keep them current alongside the issues they relate to.

One Cloudflare Worker in your own account: D1 for data, KV for cache, R2 for attachments. Leases, claims and messages live in D1 too. No servers or containers to run. The quickstart above runs on your machine; self-hosting covers deploying it to your Cloudflare account from a pre-built release.

The longer aim is a software factory: a fleet of agents working a shared backlog across projects and machines, with a person steering. Projektor’s own roadmap is tracked in Projektor.

How the agent workflow fits together · How Projektor differs from other tools · Compare with Jira, Linear and others · FAQ · Look around a live instance · How and why it was built · Star Projektor on GitHub

Built by Verdient.