Dashboard

The local web dashboard: chat with the agent, watch runs stream, replay past runs event-by-event, browse memory, and inspect scheduler jobs — all in the browser, all offline.

agentkai dashboard            # default http://127.0.0.1:8931/
agentkai dashboard --port 9000

Security posture

• Binds 127.0.0.1 only — passing any other host is refused. It is not reachable from your LAN or the internet, by design.
• A fresh random token is generated per launch, printed once by the CLI, and required on every data request (query param ?token= or Authorization: Bearer header). Wrong token → 401. The token is never logged (uvicorn access log is disabled).
• The SPA is fully offline: no CDN, no external fonts, no analytics. The only network the page touches is the dashboard itself.

Pages

• Chat — send a prompt, watch the answer stream token-by-token over SSE, with tool calls shown inline as they happen.
• Runs — every run's history with a full event timeline (messages, tool calls, results, approvals). Replay any finished run.
• Memory — browse SOUL/USER/MEMORY.md and daily logs.
• Scheduler — inspect jobs from the real scheduler database.

Under the hood

The backend is FastAPI; the frontend is a small dependency-free SPA in src/agentkai/dashboard/static/. Runs emit an append-only event log (~/.agentkai/runs/<run_id>/events.jsonl); the dashboard replays and live-streams those events over Server-Sent Events.

Honest v1 note: the chat page currently streams from a stubbed run simulation while the real agent-core wiring lands — the API contract (src/agentkai/dashboard/API.md) is stable and won't change when the real loop is plugged in. Scheduler and memory pages already read the real stores.

API

The full endpoint contract is documented in src/agentkai/dashboard/API.md (installed with the package): auth, event shapes, and every /api/* endpoint, plus the planned additions (memory editing, approval cards). Scheduler job management (create, delete, manual run) shipped in v0.2.0 and is documented there too.