English | 中文

Workspace Viewer is a read-only local workspace viewer for ChatGPT Apps and MCP.

It lets ChatGPT inspect user-approved local workspaces through a local Agent, without giving the cloud service direct access to the user's filesystem.

The first version focuses on a conservative safety model: discover registered workspaces, list file trees, inspect text files, search paths or file contents, and batch multiple read-only checks. It does not execute shell commands, write files, delete files, or access directories that the user has not explicitly registered.

Website: https://workspace-viewer.slhaf.work

ChatGPT is useful for code understanding, debugging, documentation, and refactoring discussions, but it usually cannot inspect a user's local project directly.

Workspace Viewer provides a narrow bridge:

• the user runs a local Agent on their own machine;
• the user registers only the local directories they want to expose;
• ChatGPT calls MCP tools through the hosted Worker;
• the Worker relays read-only requests to the user's local Agent;
• the Agent reads from approved workspace roots and returns bounded text results.

This keeps local filesystem access under the user's control while still making project inspection available to ChatGPT.

Workspace Viewer is intentionally limited.

It can:

• list approved workspaces;
• describe a workspace at a high level;
• list directory trees inside a workspace;
• inspect text files by path and line range;
• search file paths or text content;
• run a batch of read-only workspace checks.

It cannot:

• execute shell commands;
• write files;
• modify files;
• delete files;
• access unregistered directories;
• access paths outside a registered workspace root;
• directly browse the user's machine from the hosted Worker.

The local Agent is the final filesystem boundary. It validates workspace IDs, normalizes paths, rejects path traversal, applies ignore rules, and limits output size.

Workspace Viewer uses a hosted Worker plus a local Agent.

Main components:

• apps/worker: Cloudflare Worker, OAuth flow, D1 metadata, Durable Object relay, and MCP endpoint.
• apps/agent: local Node.js Agent that connects to the Worker and serves read-only workspace requests.
• packages/protocol: shared schemas, limits, tool names, and protocol types.

The Worker does not mount, sync, or directly access local files. Actual filesystem reads happen only inside the local Agent process.

The v1 MCP tool set includes:

Install dependencies and build all packages:

Run type checks and tests:

Run the Worker locally:

Run the Agent locally:

For local-only development without OAuth, set DEV_AUTH_BYPASS_ENABLED=true in a local Wrangler environment. Do not enable this in production.

The Agent stores local configuration under:

Example shape:

Workspace root paths are stored locally. The hosted Worker stores only workspace metadata needed for routing and discovery.

The normal flow is:

1. Sign in through the ChatGPT App OAuth flow.
2. Ask ChatGPT to create an Agent pairing code.
3. Run the local Agent pairing command.
4. Register one or more local workspaces.
5. Start the Agent.
6. ChatGPT can now inspect approved workspaces through MCP tools.

Example commands:

To revoke local Agent credentials while keeping local workspace entries:

Production deployments require OAuth and Cloudflare resources.

Required Worker secrets:

Optional reviewer login for ChatGPT Apps review:

REVIEW_PASSWORD_HASH should be formatted as:

Do not commit real secrets, real reviewer credentials, or production-only tokens.

Development seed SQL, if present, is kept outside the production migration path.

Do not apply development seed data to production D1 databases. In particular, avoid fixed development Agent IDs, workspace IDs, or tokens in any production environment.

Workspace Viewer is currently an early version intended for ChatGPT Apps / MCP publishing and review.

Implemented:

• hosted Worker MCP endpoint;
• OAuth-protected access;
• local Agent pairing;
• Durable Object based Agent relay;
• D1-backed metadata;
• read-only workspace tools;
• bounded output handling;
• reviewer login support for app review.

Planned or possible future work:

• richer project summaries;
• symbol-level navigation;
• diagnostics integration;
• LSP-assisted code understanding;
• better installation and onboarding flow.

Licensed under the Apache License 2.0. See LICENSE.