A local MCP server that gives ChatGPT Web persistent shells, direct file editing, Computer Use, browser-backed subagents, webpage tools, and reusable skills on macOS.
Quick start · Operations · Security · Maintainer wiki
Caution
Shellby MCP runs with the full permissions of your local macOS user. An authorized ChatGPT caller can run commands, edit files, fetch webpages, and control supported applications.
| Capability | Description |
|---|---|
| Persistent shells | Named login shells retain cwd, environment, processes, and command history across MCP calls. |
Native apply_patch |
Native ChatGPT apply_patch binary used to edit files independently of shell state. |
| Computer Use | Focused macOS observation and interaction backed by Peekaboo. |
| Browser subagents | Detached ChatGPT Web conversations with follow-up context and concurrent result retrieval. |
| Web and images | Rendered webpage extraction, bounded document pagination, and native image transport. |
| Dynamic skills | Reusable workflows loaded from <workspace>/skills/*/SKILL.md. |
- macOS on Apple Silicon or Intel
- Node.js 22.13.0 or newer
- npm
- An ngrok account and CLI
- A ChatGPT Plus or Higher account with Developer Mode turned on
Google Chrome is optional and is used for browser-backed subagents. Computer Use is optional and uses the Peekaboo package installed with this repository.
skills/install-shellby-mcp/SKILL.md
Tip
The manual install process should be ran from Terminal.app for the best macOS permission context.
-
Install ngrok, clone the repository, and install dependencies:
brew install --cask ngrok git clone https://github.com/Serbyte-Development/shellby-mcp.git cd shellby-mcp npm ci -
Authenticate ngrok:
ngrok config add-authtoken <your-token>
Get an authtoken from the ngrok dashboard if needed.
-
Create the active Shellby config:
npm run setup -- --config-only
Review
.shellby/config.tomland edit any values you want to customize. The generated file shows every supported setting: defaults are active, and settings without defaults are commented examples. -
Run guided setup:
npm run setup
Setup checks the machine, prepares the workspace, and builds Shellby MCP. It also checks Computer Use and prepares the dedicated ChatGPT Chrome profile when those tool groups are enabled.
-
Run the first managed start from Terminal.app:
npm start
This creates or reuses the repository-local PM2 runtime, starts Shellby MCP and ngrok, launches the configured ChatGPT browser when agent tools are enabled, waits for local health, and prints the public
/mcpURL. Starting from Terminal.app gives the managed process tree the intended macOS permission context for Computer Use. -
In ChatGPT Developer Mode, create a custom MCP app with the printed
https://.../mcpURL and select No Auth.
Important
The first trusted remote tool call binds the installation to that ChatGPT subject. Use npm run auth:reset only when you intend to clear that binding.
npm run status
curl -fsS http://127.0.0.1:3333/healthz
npm run print-urlThe local MCP endpoint is http://127.0.0.1:3333/mcp.
Computer Use
Shellby ships a package-local Peekaboo CLI build that matches its Computer Use adapter. Check or grant permissions with:
npm run setup:computerScreen Recording enables observation. Accessibility and Event Synthesizing enable actions. Shellby always uses its bundled Peekaboo executable so the adapter and CLI stay on the tested version together.
See Computer Use for runtime details.
Browser-backed ChatGPT subagents
Run the dedicated browser setup when Chrome was unavailable during initial setup or when you want to configure it later:
npm run setup:chatgptThis creates a dedicated Chrome profile under ~/.shellby/chatgpt-chrome and attaches over CDP at 127.0.0.1:9222. Sign into ChatGPT once in that profile. Future npm start runs launch it automatically while clone or subagent tools are enabled.
Conversation URL and turn count are persisted for reused agent_id values. Use npm run reset-agents to forget those local mappings.
See Browser ChatGPT Subagents for lifecycle details.
| Command | Purpose |
|---|---|
npm start |
Build and start or reload Shellby MCP, ngrok, and enabled supporting services. |
npm run restart |
Rebuild, clear the audit log, and reload services using the existing PM2 daemon. |
npm run restart -- --hard |
Rebuild and recreate Shellby's PM2 daemon from a healthy Terminal.app session. |
npm run status |
Show PM2 process state. |
npm run logs |
Follow PM2 logs. |
npm run pm2 -- <args> |
Run a PM2 command against Shellby's dedicated daemon. |
npm run print-url |
Print the active public /mcp URL. |
npm run stop |
Stop the managed Shellby MCP and ngrok processes. |
npm run auth:reset |
Clear the bound remote ChatGPT subject after confirmation. |
npm run reset-agents |
Forget persisted subagent conversation mappings. |
PM2 is installed as a repository dependency. All Shellby PM2 commands use ~/.shellby/pm2 for their daemon, sockets, logs, and process state. This is separate from the default ~/.pm2 daemon used by other projects; use npm run pm2 -- <args> for direct access to Shellby's daemon.
Use npm run restart for routine code or config changes, including from shell_run. It keeps the PM2 daemon, reloads ngrok, then reloads MCP. The initiating shell and MCP connection may close; after services return, make a fresh tool call. Build failures leave running services and the audit log intact. The authenticated ChatGPT Chrome profile is reused.
For macOS permission or service-context problems, run npm run restart -- --hard from a healthy Terminal.app session. This also recreates Shellby's dedicated PM2 daemon. Hard restart refuses to run inside Shellby because stopping that daemon would kill the command responsible for starting its replacement.
If Shellby was previously running under the default shared daemon, move it once from Terminal.app before using the new commands:
PM2_HOME="$HOME/.pm2" ./node_modules/.bin/pm2 delete shellby-mcp shellby-ngrok
npm run restartThe first command removes only Shellby's old app entries, leaving the shared daemon and other projects running. If an older installation still has a shellby-cursor-host entry in that daemon, delete that entry there too. The migration briefly interrupts Shellby; it does not erase authentication or browser profiles. Future restarts need only npm run restart.
git pull
npm ci
npm run setup -- --config-only
npm startShellby's public configuration is the gitignored .shellby/config.toml. npm run setup creates a config showing every supported setting for new installations, with defaults active and the optional ngrok.url shown as a commented example. Uncomment and customize that URL to use a reserved ngrok endpoint; leaving it commented lets ngrok assign the public URL. Existing files are left untouched, preserving their values, comments, and formatting. Missing settings use defaults automatically. Invalid values produce a warning and fall back individually; unknown keys warn and are ignored. Valid TOML formatting—including reordered sections, dotted keys, and inline tables—is accepted. Malformed TOML syntax still needs correction; errors show the location and never rewrite your file.
The TOML surface currently owns the workspace, shell path, ChatGPT CDP/project routing, MCP tool-output format, and startup-static tool groups. The generated config enables every tool group and defaults tool output to compact. Setting a group to false removes those tools from tools/list after Shellby restarts and skips its supporting runtime service where one exists. start_here is always published. For example, this customization disables browser-backed agents and Computer Use:
workspace = "~/Desktop/agent-workspace"
[shell]
path = "/bin/zsh"
rtk = false
[chatgpt]
cdp_endpoint = "http://127.0.0.1:9222"
project_url = "https://chatgpt.com/"
max_delegated_agents = 3
[mcp]
tool_output = "compact"
[tools]
review = true
shell = true
apply_patch = true
clones = false
subagents = false
web = true
skills = true
image = true
computer = falsechatgpt.max_delegated_agents sets the maximum number of delegated agent IDs per main-agent session, shared by subagents and clones. It defaults to 3; invalid values warn and fall back to 3. Set a positive integer to override it. Saved agents and in-flight creations count toward the limit; existing IDs remain reusable even if the limit is lowered. The per-call batch limit remains three. Older configs can omit this field. To override it, add it under [chatgpt], then restart Shellby after changing its value.
mcp.tool_output controls the representation used for ordinary tool results. compact is optimized for model context and omits public output schemas; structured preserves each tool's structured result and output schema for MCP clients that use them. Computer Use and image_view keep their native MCP content in either mode. Changing this setting requires a Shellby restart.
shell.rtk defaults to false, so RTK is not required to install or run Shellby. To enable transparent RTK command rewriting, install RTK Token Killer with brew install rtk, set shell.rtk = true, and restart Shellby. Shellby resolves that executable from startup PATH, then uses the resolved absolute path for supported rewrites regardless of the shell command's cwd. Unsupported rewrites and RTK failures fall back to the original command. Shellby keeps the caller's original command for request identity, auditing, and command previews, and disables RTK's separate failure tee, telemetry, and persistent history for Shellby-launched commands while still loading normal RTK filtering/exclusion configuration.
Shellby does not use a repository .env file. User-configurable Shellby settings come only from .shellby/config.toml; external tools use their normal machine-level configuration. In particular, ngrok is resolved from PATH and authentication is configured with ngrok config add-authtoken. Chrome is discovered in the normal macOS application locations, and Shellby uses its bundled Peekaboo build. Host, port, runtime limits, and other non-configurable settings remain code-owned in src/config.ts.
If the endpoint stays offline after restarting, check npm run status and npm run logs from Terminal.app. Run npm start if services are stopped. npm run print-url reports the active ngrok endpoint; an offline reserved URL does not by itself mean your config changed. Routine restart through shell_run may disconnect its own call while PM2 brings MCP back.
Setup or startup fails
Run:
npm run preflight
npm run status
npm run logspreflight checks the supported macOS/Node environment, local dependencies, ngrok installation, and ngrok authentication. If /healthz does not become available after startup, inspect PM2 status and logs first.
Computer Use permissions are missing
Run npm run setup:computer from Terminal.app and follow Peekaboo's permission guidance. Keep the managed Shellby process associated with the same intended Terminal permission context.
The PM2 daemon needs to be recreated
Run npm run restart -- --hard from a newly opened Terminal.app session. This recreates Shellby's dedicated PM2 daemon as well as Shellby MCP and ngrok. A stale macOS session inherited by PM2 can cause Chromium to abort before navigation and DNS lookups to fail even while MCP remains reachable. Ordinary restart retains the daemon's session and cannot repair that context. Daemon isolation does not make PM2 independent of Terminal's macOS session; keep Terminal.app running until PM2 is managed through a macOS LaunchAgent.
More startup and recovery details are in Configuration and Startup.
- The checked-in ngrok traffic policy exposes the local MCP endpoint to ChatGPT.
- Direct localhost MCP access is unauthenticated. Do not expose the local endpoint through another untrusted proxy.
- Trusted remote tool calls are bound to the first ChatGPT subject stored in
~/.shellby/auth.json. - The dedicated authenticated Chrome profile is part of the trust boundary for browser subagents.
agent-commands.yamlcan contain sensitive tool inputs. It is gitignored and permission-restricted and should be treated as private.
See SECURITY.md for reporting and scope.
npm run dev
npm run ui:dev
npm run lint
npm run typecheck
npm test
npm run build
npm run ui:buildRun npm run ui:install once after cloning to install the dashboard dependencies. Use npm run inspect for the MCP inspector and npm run schemas to print the published tool schemas. Authenticated browser tests are excluded from CI. See Build and Test.
The maintainer wiki contains implementation and operational details:
- Project Overview
- Architecture Map
- Configuration and Startup
- Computer Use
- MCP Tool Surface
- Build and Test
- Open Questions and Risks
Read CONTRIBUTING.md before opening a pull request. Run the development validation commands above for code changes.
MIT. The vendored apply_patch binary retains its upstream OpenAI Codex license and notices under vendor/apply-patch/.
Shellby MCP is created and maintained by Serbyte Development.