Local Chrome side-panel for Claude Code, Cursor, Codex, and Antigravity sessions.
Three pieces:
- Chat — sessions in the side panel, streamed from your local CLI
- Browser — bind a tab; Claude can read and click through MCP
- Connection — a loopback-only Python bridge on
127.0.0.1:7878
The bridge launches your already-installed CLI/IDE adapters (claude,
cursor-agent, codex, or Antigravity headless mode). Nothing is sent to a
DevScope cloud — there isn’t one. Auth is a local token
(~/.dev-bridge/token, chmod 0600). The process refuses to bind anywhere
except loopback.
The side panel is the surface. The goal underneath it is a standing colleague — an assistant and developer that works on its own initiative, on your machine, with your accounts, and comes back with work already done instead of waiting to be told what to do.
A chat tool answers when spoken to. Turning that into a colleague takes three things, and all three are in this repo today:
| Property | What it means | Where it lives |
|---|---|---|
| A heartbeat | Something wakes the agent on a timer, so it can act with nobody typing | orchestrator.py — a tick loop driving a Manager-of-Managers → Domain Manager → Worker hierarchy |
| Durable state | Work outlives a single conversation, a restart, or a crashed process | task_store.py — a task board with parent/child trees, logs, and stuck-worker reconciliation |
| Explicit limits | Autonomy you can leave running, because the dangerous actions stop and ask | agent_policies.py — workers must call task_request_approval before sending, buying, deleting, or spending |
Reading, researching, navigating, and filling forms without submitting are autonomous. Sending, publishing, deleting, and anything touching money pause for you. That line is the whole design: the agent should be able to run all day unattended without any single step being one you'd want to undo.
Working today: the tick loop, the task hierarchy, the approval gate, per-site
playbooks that remember selectors and workflows across runs
(schedule_playbook.py), and an
autonomous-employee layer that can report to Notion, WhatsApp, or email
(employee_config.py — off by default).
Not there yet, and the difference between "runs unattended" and "genuinely doesn't wait for instructions":
- Self-originating work. Tasks still mostly arrive from you. The agent can already read mail, messages, and calendar through MCP; what's missing is the step that turns an observation into a proposed task on its own.
- Memory of you. Playbooks remember websites. Nothing yet remembers your preferences, your past decisions, or how you like things done.
- Judgement about when to interrupt. A colleague who asks about everything is as tiring as one who asks about nothing.
Local-first, permanently. The bridge binds loopback only and refuses anything
else. The agent is the claude CLI you already installed and logged into —
this project ships no model, no account, and no server of its own. Autonomy
here means your machine doing more on your behalf, not your work moving
somewhere you can't see it.
| Thing | Why |
|---|---|
| macOS or Linux, Python 3.11+ | Runs the bridge |
Claude CLI logged in (claude login) |
Claude sessions |
| Chrome | Hosts the side panel |
| Node.js 18+ | Builds the extension once |
| Cursor CLI (optional) | Cursor sessions |
| Codex CLI (optional) | Codex sessions and app-server resume/steering |
| Antigravity IDE/CLI (optional) | Antigravity headless sessions |
Verify the adapters you plan to use are available:
which claude && claude --version
which cursor-agent && cursor-agent --version # optional
which codex && codex --version # optionalFrom this folder:
python3 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e .That puts devscope-bridge on PATH inside the venv.
devscope-bridge
# or: python -m devscope_bridge.mainIt prints:
Dev Bridge started. Token: <hex>
and writes the same value to ~/.dev-bridge/token (chmod 0600). The token is
reused across restarts — you only paste it into the extension once.
Keep this terminal open, or install the macOS service in step 4.
Check:
curl http://127.0.0.1:7878/health
# {"ok":true,"version":"0.1","uptime_s":…}| Variable | Default | Purpose |
|---|---|---|
BRIDGE_HOST |
127.0.0.1 |
Bind address. Loopback only — other hosts are rejected. |
BRIDGE_PORT |
7878 |
TCP port |
BRIDGE_EXTENSION_ID |
placeholder | Chrome origin added to CORS, e.g. chrome-extension://abcdef… |
CURSOR_BIN |
macOS Cursor.app path | Cursor CLI binary |
Copy .env.example → devscope_bridge/.env.local if you prefer a file (gitignored).
Do not run the bridge with uvicorn --workers 2+. PTY terminals and
WebSockets live in one process. workers=1 is required.
cd extension
npm install
npm run buildIn Chrome:
chrome://extensions- Enable Developer mode (top right)
- Load unpacked → select
extension/dist/(the built folder, notsrc/) - Copy the ID shown on the DevScope card (32 characters)
Pin the icon if you want. Click it to open the side panel.
Restart the bridge so CORS trusts this install:
export BRIDGE_EXTENSION_ID="chrome-extension://PASTE_THE_ID_HERE"
devscope-bridgeIn the side panel: gear (Settings) → paste:
cat ~/.dev-bridge/tokeninto Bridge token → Save / Test connection.
The connection pip in the header turns green when the WebSocket is live.
First chat: New chat → pick a project folder (your code) → send a message. Claude runs in that folder. For browser tools, bind the current tab with the globe control in the composer.
| Symptom | Fix |
|---|---|
/health times out or connection refused |
Bridge not running — start it, or see ~/.dev-bridge/bridge.log |
403 |
Token mismatch — paste cat ~/.dev-bridge/token again |
| Pip grey, health 200 | CORS — restart bridge with BRIDGE_EXTENSION_ID=chrome-extension://<id> |
claude not found |
which claude; LaunchAgent PATH is small — use scripts/install-macos-service.sh |
503 Browser client not connected |
Open the side panel and bind a tab |
After the venv is installed:
chmod +x scripts/install-macos-service.sh
./scripts/install-macos-service.shThis writes ~/Library/LaunchAgents/com.devscope.bridge.plist, starts at login,
keeps the process alive, and raises the open-files limit.
# logs
tail -f ~/.dev-bridge/bridge.log
# stop
launchctl bootout "gui/$(id -u)/com.devscope.bridge"
# start again
launchctl bootstrap "gui/$(id -u)" "$HOME/Library/LaunchAgents/com.devscope.bridge.plist"If you rebuild the Python package, kickstart:
launchctl kickstart -k "gui/$(id -u)/com.devscope.bridge"On Linux, run devscope-bridge under systemd/user or a tmux session. There is
no Linux unit in this snapshot.
Claude only sees browser tools if an MCP config is loaded.
Option A — this repo as the chat’s project folder
.mcp.json already lists browser-control. In DevScope → New chat →
Project = this directory.
Option B — your own repo
Copy .mcp.json into that project, then set the chat’s project path to that repo.
Option C — register globally for the Claude CLI
source .venv/bin/activate
python -m devscope_bridge.setup_mcp
claude mcp listUse the same Python that has devscope_bridge installed. If .mcp.json
says python3 -m devscope_bridge.…, that python3 must resolve to the venv
(activate it, or edit .mcp.json to .venv/bin/python).
Bind a tab with the globe control in the composer. Without a bound tab,
browser tools return 503.
Beyond chat + browser + connection, the repo ships the pieces described in Where this is going. All of it is off by default — a fresh install behaves exactly like a side panel with a chat in it.
| Piece | What it does | Settings |
|---|---|---|
| Orchestrator | Tick loop, 3-tier agent hierarchy, worker dispatch, concurrency cap, quota throttling | ~/.dev-bridge/orchestrator_config.json |
| Task board | Persistent tasks, subtasks, logs, approval requests | ~/.dev-bridge/*.db |
| Schedules | Recurring runs with a learned per-site playbook | ~/.dev-bridge/*.db |
| Autonomous employee | Board sync (Notion) and reporting over panel / WhatsApp / email | ~/.dev-bridge/employee_config.json |
| Cockpits | WhatsApp, Gmail, Calendar, Meta Ads — read/act through MCP | OAuth tokens in ~/.dev-bridge/ |
Turn the orchestrator on and it starts waking agents on its own. Read
agent_policies.py first — it is the
document that defines what an unattended worker may and may not do without
asking you. WORKER_SAFETY_POLICY is the default; the _E2E variant relaxes
sending and posting and is a deliberate choice, not a default.
OAuth tokens (if you use the cockpits) land in ~/.dev-bridge/ and must
never be committed.
devscope_bridge/ Python package (FastAPI bridge + MCP stdio servers)
main.py App, lifespan, loopback bind
session_http.py REST: sessions, transcript, browser actions
ws_session.py Chat WebSocket
pty_ws.py Interactive terminal
session_manager*.py Claude/Cursor/Codex/Antigravity process adapters
orchestrator*.py Tick loop + pure decision logic for the agent hierarchy
agent_policies.py System prompts and the worker approval policy
task_store.py Persistent task board
extension/ Chrome MV3 side panel (Vite)
scripts/ macOS LaunchAgent installer
.mcp.json Claude Code MCP (this repo)
Data on disk (all local):
~/.dev-bridge/token~/.dev-bridge/sessions.json~/.dev-bridge/bridge.log~/.dev-bridge/chat-transcripts/~/.dev-bridge/*.db
source .venv/bin/activate
pip install pytest
pytestEmpirical tests that spawn a real claude process stay skipped unless
RUN_EMPIRICAL=1.
Step-by-step setup guide: docs/SETUP.md (English) · docs/SETUP.he.md (עברית)
PRs welcome — see CONTRIBUTING.md. To report a vulnerability, see SECURITY.md. Setup help: SUPPORT.md.
MIT — see LICENSE.