Cockpit
Self-hosted kanban/roadmap tracker built API-first for AI agents, evolved out of Mercury’s built-in kanban.
Mercury’s task-tracking was built as one piece of a bespoke agent-orchestration platform - custom agent SDKs, a purpose-built runtime, its own kanban wired into that runtime. As portable AI coding tools got capable enough on their own, that custom orchestration layer stopped earning its complexity.
Rather than keep maintaining a bespoke runtime just to get a task board, the tracking piece was rebuilt standalone: a tool-agnostic HTTP API any agent runtime can talk to, not something wired into one custom platform. Recognising when your own infrastructure has become the overhead is as much the story here as the tool itself.
A Python HTTP API (Phase → Milestone → Task hierarchy plus first-class Decision records) with optimistic concurrency via updated_at tokens, transactional bulk updates, and an append-only audit event log per item - self-described via a hand-written OpenAPI document so any agent can discover the contract without custom integration work.
Cockpit is now the shared task home across active projects - Train Mate, Mondoro, and Cockpit’s own development among them - replacing Mercury’s built-in, platform-specific kanban with one tracker any project or agent can point at. A mobile-friendly installable PWA and a thin CLI sit on top of the same API.
- Phase → Milestone → Task → Decision hierarchy with a canonical relationship model (related / blocks / depends_on)
- Optimistic concurrency via updated_at tokens; transactional bulk updates (all-or-nothing across up to 100 items)
- Append-only, per-item audit event log plus a project-wide event feed
- Decision workflow: lock with rationale, reopen with a reason, spin off follow-up work - never silently converted
- Installable PWA with a desktop launcher and Tailscale HTTPS access for phone review
- Self-describing OpenAPI contract so agents discover routes without re-deriving them from docs
- Deliberate separation of human notes/feedback from agent notes/feedback on every ticket
Evolved out of Mercury’s built-in kanban once portable AI tooling made a bespoke orchestration runtime unnecessary for task tracking. Actively used as the shared task home for Train Mate, Mondoro, and its own roadmap. Currently local-only with no auth layer by design; a hosted Vercel deployment is planned next.