Dispatch

A self-hosted, air-gapped issue tracker where MCP is a first-class interface for coding agents.

Quick start · Connect an agent · MCP tools · Architecture · Security

One Next.js app + PostgreSQL. No external services, no telemetry, no CDN, no remote fonts or images: it works with zero internet access.

Features

• 📋 Board and list views, labels, relations, Markdown comments, per-issue activity

• ⌨️ Command palette (⌘/Ctrl + K), keyboard shortcuts (press ?), PostgreSQL-backed search

• 🤖 Built-in MCP server at http://localhost:15000/mcp (20+ tools) and an agent skill

• 🔍 Traceability: structured JSON logs on stdout plus an audit log of every UI and MCP change (Settings → Audit log)

• 👥 No accounts. Two fixed actors: you (the UI) and the agent (everything that arrives over MCP). Issues can be assigned to either

Related MCP server: jira-mock-mcp

Quick start

Docker (one command)

docker run -d --name dispatch -p 15000:3000 -v dispatch-data:/data ghcr.io/iwandejong/dispatch

Open http://localhost:15000. No configuration needed: the image bundles PostgreSQL and applies migrations on start. Data lives in the dispatch-data volume.

Docker Compose (separate Postgres container)

git clone https://github.com/iwandejong/dispatch.git && cd dispatch
./scripts/init-env.sh        # writes .env with a random database password
docker compose up -d

Open http://localhost:15000, create a project, and start filing issues.

Postgres lives on the internal compose network only (no host port), with a persistent pgdata volume and a health check. The app waits for it and applies migrations on start. No demo data is created.

⚠️ There is no authentication. The app is published on 0.0.0.0:15000. See SECURITY.md before exposing it to a network.

Connect a coding agent

claude mcp add --transport http dispatch http://localhost:15000/mcp

Then install the skill so the agent knows how to use the tools well:

ln -s "$PWD/skills/dispatch" ~/.claude/skills/dispatch

Details, other clients and the tool list: MCP.md.

Architecture

Browser ─┐                       ┌─ Server Actions ─┐
         ├─ Next.js (one app) ───┤                  ├─ lib/services/* ─ Prisma ─ PostgreSQL
Agent ───┘   UI · auth · /mcp    └─ MCP tools ──────┘
                                   └── audit + JSON logs

All business rules live in lib/services/; UI actions and MCP tools are thin adapters. See ARCHITECTURE.md.

Configuration

Logs and audit

• docker compose logs -f app shows one JSON line per operation: ts, level, msg, requestId, source (ui|mcp), actor (HUMAN|AGENT), target, ok, error, durationMs. Secret-looking argument keys are redacted and long text truncated.

• Every mutation and failed operation is also stored in the AuditLog table and shown under Settings → Audit log. Reads (search, get, list) are logged to stdout only.

• Per-issue history (status, assignee, priority, labels, comments) is under each issue's Activity.

Data, backup, upgrade

• Data lives in the pgdata Docker volume. docker compose down keeps it; docker compose down -v deletes it.

• Backup: docker compose exec -T postgres pg_dump -U dispatch dispatch > backup.sql

• Restore into a fresh stack: docker compose exec -T postgres psql -U dispatch dispatch < backup.sql

• Upgrade: git pull && docker compose up -d --build (migrations run automatically on start).

Limitations

Single-user by design (you + one agent, no accounts or permissions). Search uses PostgreSQL trigram indexes and is meant for thousands to tens of thousands of issues. Project pages load up to 200 issues. There is no realtime sync between browser tabs.

Development, contributing, security

DEVELOPMENT.md · CONTRIBUTING.md · SECURITY.md · MIT License