workspace-agent-relay-mcp

Local relay + dashboard for ChatGPT Workspace Agents: the agent reports plan, progress, tool calls, and results back to your machine in real time.

Why: the Workspace Agent trigger API is fire-and-forget (202 with no body). This relay gives the agent an MCP callback channel and you a live dashboard.

Architecture

you (browser)     your machine                         ChatGPT
┌──────────┐    ┌─────────────────────────┐        ┌─────────────────┐
│ dashboard│◀── │ workspace-agent-relay-mcp│ ◀─MCP── │ Workspace Agent │
│  :8799   │    │  (this repo)            │        └─────────────────┘
└──────────┘    └───────────┬─────────────┘
                            │ tool-trace POST
                            ▲
                   ┌────────┴────────────┐
                   │ notion-local-ops-mcp │  optional: files / shell / git
                   └─────────────────────┘

• This repo — reporting MCP (record_plan, record_progress, record_result, …) + dashboard + SSE.

• notion-local-ops-mcp (optional) — execution tools; bind_relay_run mirrors tool calls here.

• Dashboard IA — conversations are organized by Workspace (working_directory) or the built-in 无目录 space. Workspace Agent entries are execution backends selected in Settings, not the primary sidebar object.

The active code path is intentionally relay-only: MCP callback events plus local tool traces. Older cloud-side readback experiments have been removed from the working tree; use Git history if you need to study them.

Quick start

python3.11 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env   # set WORKSPACE_AGENT_RELAY_AUTH_TOKEN at minimum
workspace-agent-relay-mcp

Open http://127.0.0.1:8799/, enter the relay access password (WORKSPACE_AGENT_RELAY_AUTH_TOKEN).

Tunnel (optional): ./scripts/dev-tunnel.sh

Local development (recommended)

For day-to-day UI work you do not need pnpm run build. The Python server serves a static frontend/dist bundle on :8799 — that path is production-like and requires a rebuild after every change.

Instead, run the backend and the Vite dev server separately:

# Terminal 1 — API + MCP + SSE (launchd/dev-tunnel is fine too)
.venv/bin/python -m workspace_agent_relay_mcp.server

# Terminal 2 — dashboard with HMR
./scripts/dev-dashboard.sh

Open http://127.0.0.1:5173 for the dashboard. ChatGPT MCP stays on http://127.0.0.1:8799/mcp (or your tunnel URL).

dev-dashboard.sh checks that the relay is healthy, then starts Vite. It reads WORKSPACE_AGENT_RELAY_AUTH_TOKEN from the repo .env and seeds the browser token automatically — no need to paste it again on :5173.

Run cd frontend && pnpm run build only before CI or when you want the single-port :8799 dashboard without Vite.

ChatGPT Workspace Agent setup (required)

Without these four steps the dashboard stays empty.

1. Register an execution backend in the relay dashboard

Settings → Add backend (or use .env for a single default backend):

For multiple ChatGPT accounts, add one backend row per account. Pick the Current backend in Settings; new threads use that backend automatically.

2. Configure workspaces

Use the sidebar workspace selector or Settings → Workspaces:

• 无目录 is built in and sends no cwd.

• A workspace has a name and an absolute working_directory.

• New threads inherit the current workspace.

• Every run stores a working_directory snapshot and injects it into the Workspace Agent trigger, so later workspace path edits do not rewrite old runs.

3. Connect MCP: relay (required)

In the Workspace Agent → MCP connectors:

Confirm tools: record_plan, record_progress, record_result, ask_user, get_run_context, server_info.

4. Connect MCP: local ops (recommended)

Add notion-local-ops-mcp as a second connector so Tool calls stream in the dashboard. See its README for URL/auth.

5. Paste Agent Instructions (easy to miss)

Open docs/agent-instructions.md:

1. Replace placeholders <YOUR_RELAY_MCP> and <YOUR_LOCAL_OPS_MCP> with your connector names (defaults: workspace-agent-relay-mcp and notion-local-ops-mcp).

2. Copy the fenced 指令正文 block.

3. ChatGPT → edit Workspace Agent → Instructions → paste (append or replace any old relay section).

Expected workflow:

record_plan  →  bind_relay_run  →  batch record_progress  →  record_result

Without this paste, the agent may work only inside ChatGPT and the operator sees nothing on the relay.

6. Smoke test

Select a workspace, then send a short task from the dashboard (e.g. inspect the current directory or create/read a file under /tmp). You should see plan → tool traces (if local-ops connected) → result, and the trigger should include working_directory for non-无目录 workspaces.

MCP tools (relay)

Dashboard

Per run: your message → plan checklist → tool traces (from local-ops) → progress notes → ask_user (if any) → final result. Updates via SSE (no refresh).

Pairing with notion-local-ops-mcp

1. Agent calls record_plan on this relay.

2. Agent calls bind_relay_run on notion-local-ops-mcp with request_id (+ conversation_key) from the trigger (no relay_url needed when configured locally).

3. Traced tools POST to /internal/tool-trace on this relay (shared bearer) → dashboard.

Details: notion-local-ops-mcp → Relay Bridge.

Project layout

src/workspace_agent_relay_mcp/   server, MCP tools, API, SQLite store, trigger client
frontend/                        React dashboard
docs/agent-instructions.md       paste block for ChatGPT Instructions
scripts/dev-tunnel.sh            cloudflared supervisor
scripts/dev-dashboard.sh         Vite HMR dashboard (proxies /api → relay)
tests/

Security

• Never commit .env, tunnels, or *.sqlite*.

• WORKSPACE_AGENT_RELAY_AUTH_TOKEN — dashboard + /mcp bearer, and the shared bearer for /internal/tool-trace (the notion-local-ops bridge). When unset, /internal/tool-trace is disabled.

• Workspace Agent access tokens — stored in relay DB or .env; rotate if leaked.

• MCP tool writes route by request_id + conversation_key and are rejected once a run is terminal (done/blocked/failed/superseded).

• Dashboard sends default to queue/new request: Enter creates a fresh request_id and does not close the current active run. Explicit steer/guidance reuses the selected active run's request_id.

• Workspace paths are plain absolute local paths. They are copied into each run as working_directory_snapshot; old runs are not rewritten when a workspace is edited or deleted.

Status

Local-first prototype for learning the Workspace Agent callback gap — not a product.