agentboard

agentboard

Source: https://github.com/edgeetech/agentboard

Local agent management portal as a Claude Code plugin. A kanban board where PM / Worker / Reviewer agents — spawned as headless claude -p subprocesses — move tasks through predefined workflows. Only the Human (Product Owner) approval step is manual.

One SQLite DB per project, one local web UI, zero cloud dependencies.

Install

In any Claude Code session:

/plugin marketplace add edgeetech/agentboard
/plugin install agentboard@agent-board-local
/reload-plugins

Then in any Claude Code session:

/agentboard:open     # boot server + open UI in browser
/agentboard:doctor   # health checklist
/agentboard:stop     # stop the local server

No post-install build. Plugin ships the prebuilt UI bundle.

Requirements

• Node ≥ 22 or Bun ≥ 1.x on PATH (server uses node:sqlite, built-in).
• claude CLI ≥ 2.0.0 on PATH with either ANTHROPIC_API_KEY env set or an active OAuth session (claude /login).
• Windows, macOS, or Linux.

Workflows

Pick one per project at creation:

• WF1 (default): Todo → Agent Working → Agent Review → Human Approval → Done
• WF2: Todo → Agent Working → Human Approval → Done (skip Reviewer)

Flow is autonomous: task creation auto-dispatches PM; state transitions auto-chain Worker and (WF1) Reviewer. Human approval is the only manual gate.

Roles

Role	Does
Project Manager	Enriches the raw task into a tight brief + 3–7 acceptance criteria, hands off to Worker. Read-only over the repo.
Worker	Edits files under repo_path. No commits, no branches — leaves working tree dirty for the human to review.
Reviewer (WF1)	Checks work against AC, approves to Human or bounces back to Worker with REWORK: reason.
Product Owner	You. Final approval / rejection in the UI.

Rework loop caps at 3 iterations — after that the task stalls and surfaces a "Retry from Worker" button.

Cost tracking

Every run parses its stream-json usage events and stamps agent_run.cost_usd using the hardcoded pricing table (agent-board-core/src/pricing.mjs). Task cards show cost chips; project header shows all-time / 7d / 30d totals. Unknown model → $0 with an "uncosted" flag (never silent wrong numbers). Pricing table carries a PRICING_VERSION; when Anthropic prices change, a v1.1 reprice skill will recompute historical runs.

Data location

Outside the repo, never touched by plugin upgrades:

~/.agentboard/             (%USERPROFILE%\.agentboard on Windows)
  projects/<code>.db       one SQLite per project
  logs/<run_id>.jsonl      headless Claude stream-json output
  logs/<run_id>.err.log    captured stderr
  run-configs/<id>.json    tmp MCP config per run (deleted on exit)
  config.json              port, pid, token, server_id, active project
  server.lock              single-instance file lock

Security

• Binds 127.0.0.1 only; DNS-rebinding blocked via Host-header guard (421 on mismatch).
• 32-byte hex Bearer token in ~/.agentboard/config.json (0600 on Unix, owner-only ACL on Windows via os.userInfo().username).
• Spawned agents receive a whitelisted env — AWS / GitHub / SSH / cloud-SDK secrets in your shell are not passed to claude -p subprocesses. Only ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, and OS basics (PATH, HOME, LANG, Windows APPDATA/LOCALAPPDATA, etc.). See src/child-env.mjs.
• run_token is issued to an agent exactly once at claim_run, to the process that wins the queued→running CAS. No re-issue on retry.
• REST /api/tasks/:id/transition always attributes to by_role='human' — agent transitions go through the HTTP MCP endpoint.
• UI HTML carries a per-request CSP nonce; inline scripts without the nonce are blocked. Token delivered via Set-Cookie: ab_token=…; HttpOnly; SameSite=Strict + inline script for fetch. X-Content-Type-Options: nosniff, Referrer-Policy: no-referrer also set.
• /alive is intentionally unauthenticated — returns only {ok, server_id, plugin_version} for the plugin's ensure-server probe. No token is ever reflected in any endpoint response.
• ~/.agentboard/logs/ contains full Claude stream-json transcripts (prompts + tool I/O). Served via /api/logs/:run_id (Bearer-protected, ULID-gated). Treat the data dir as sensitive.

v1.1 planned: per-run rate limiting on MCP mutations, stricter UTF-8 body validation.

Development

Repo layout: