AgentDock

Self-hosted control plane for AI agents, automations, approvals, and failure triage.

AgentDock is an operator cockpit for developers who run multiple AI agents, cron jobs, MCP tools, and personal/team automations. It answers the daily operational questions that generic dashboards do not:

• Which agents/jobs ran recently?
• What failed, and what should I check next?
• Which risky actions are waiting for approval?
• Which connectors are degraded or stale?
• What should the operator review today?

Why AgentDock

AI agents are becoming part of everyday development workflows, but their operational state is usually scattered across logs, GitHub, cron, chat threads, observability tools, and local scripts. AgentDock focuses on the missing control-plane layer: visibility, triage, approvals, connector health, and auditability.

This is not a generic personal dashboard. AgentDock is built around agent operations:

Surface	Purpose
Today Cockpit	Prioritized operator queue for blocked actions, approvals, failures, and events
Agent Registry	Agent/job status, risk level, last run, and last output
Pending Approval	Risky actions that must not run without operator review
Failure Triage	Probable cause and suggested next check for failed automations
Connectors	Read-only health view of cron/GitHub/calendar/log connectors
Activity Timeline	Recent agent and automation events

Current MVP

AgentDock v0.1 is intentionally read-only. It demonstrates the control-plane shape without requiring production credentials, OAuth, or external write actions.

• Dependency-free Node.js server.
• Static dashboard UI.
• JSON-backed sample read model.
• Redacted public state and summary APIs.
• Fail-closed action boundary: every POST /api/* returns 405 with read_only: true.
• Contract tests for summary, redaction, fail-closed POSTs, and UI surfaces.
• GitHub Actions CI workflow.

Quickstart

Requirements:

• Node.js 22+

git clone https://github.com/j-yoon08/agentdock.git
cd agentdock
npm run check
npm test
npm start

Open:

http://127.0.0.1:4177

API

Endpoint	Method	Description
/api/health	GET	Liveness and product mode
/api/summary	GET	KPIs, overall health, today queue, risk counts
/api/state	GET	Redacted public read model
/api/*	POST	Blocked in v0.1, returns 405 and read_only: true

Example:

curl -sS http://127.0.0.1:4177/api/summary

Safety model

AgentDock starts from a conservative boundary:

1. Read-only connectors first.
2. Approval queue before any external write.
3. Fail-closed POST/action routes.
4. Redacted event and failure summaries.
5. Local/self-hosted by default.

Future action execution should require a durable approval request, exact action fingerprint, operator receipt, policy revalidation, append-only audit log, and replay protection.

Project structure

data/sample-state.json      sample agent/approval/failure/event/connector state
src/control-plane.js        normalize, redact, summarize read model
src/server.js               GET-only API and static file server
public/                     dashboard UI
test/                       node:test contract tests
docs/                       architecture, roadmap, product brief, submission materials

Roadmap

• v0.2: read-only connector interface, cron connector, GitHub connector, local log connector.
• v0.3: SQLite persistence, append-only event ledger, approval request/receipt schema.
• v0.4: approval detail page, acknowledge/snooze/edit receipt flow, GitHub PR/report integration.
• v1.0: self-hosted agent operations console with approval-gated action bridges.

See docs/roadmap.md for details.

Documentation

• Product brief
• Architecture
• Roadmap
• Submission packet
• Demo script
• OpenAI/Codex OSS application draft

License

MIT — see LICENSE.