AgentDock
AgentDock lets MCP clients (such as ChatGPT) inspect, edit, and verify a local developer workspace through a bounded, mode-controlled toolset.
Workspace management: open the configured/current workspace or a chosen root, list/close workspaces, and set the default workspace.
File exploration: generate file trees, search text/regex or by symbol/reference/impact intent, and read text files in bounded, paged chunks.
Repository analysis: build a repository map with languages, project types, entrypoints, symbols, relationships, related tests, and change-risk signals.
Image access: get image info, safe previews, rectangular crops, and tiled high-detail reads from workspace-local JPEG/PNG/WebP images.
File modification: write files, apply targeted edits, or apply unified patches, returning diffs while enforcing workspace/path guards.
Verification: run allowlisted verification commands (tests/build/lint/typecheck) and review Git changes with
show_changes.Skills: discover workspace/user/plugin skills and load their bounded SKILL.md contents.
Diagnostics/config: view server config, safety modes, limits, and blocked paths; run a controlled self-test; dispatch to registered tools via the
local-workspace-bridgesupertool.
Provides Git-aware review of workspace changes, including status, diff stats, and unified diffs via the show_changes tool.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@AgentDock Open this workspace and summarize the repo architecture without making changes."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
AgentDock
Languages: English · 简体中文 · 日本語
AgentDock is a local MCP gateway that gives ChatGPT bounded access to a developer workspace. The model remains the planner; AgentDock supplies repository evidence, local files and images, direct edits, command verification, and Git-aware review.
Status: early public release (
0.2.x). Review the security model before exposing a local workspace through a public tunnel.
0.2.0 agent tools
Persistent commands with incremental logs, stdin, cancellation and HTTP reconnects; isolated local-browser interactions and native screenshots; bounded batch reads and SHA-256 checked edits. The latest Control Center startup, hot settings and PyMuPDF fixes are retained. ChatGPT remains the planner.
See agent tools, ChatGPT instructions and changes. Browser tools require an installed Edge, Chrome or Chromium and target loopback development pages. General scripts and development servers require full bash mode.
Related MCP server: CodexPro
Why AgentDock
ChatGPT can reason about code, but a web session normally cannot inspect or modify arbitrary files on your computer. AgentDock bridges that gap through MCP while keeping workspace boundaries explicit.
The primary loop is intentionally simple:
ChatGPT model
├─ inspect/search repository evidence
├─ read text or images
├─ edit/write/apply patch
├─ run bounded verification commands
└─ review the resulting Git changes
↓
local workspaceAgentDock does not add a second server-side planning or execution layer. Planning stays with the model that is already driving the MCP tools.
Features
Workspace isolation — open one or more explicitly allowed roots; path traversal, blocked files, and symlink escapes are guarded.
Repository analysis — bounded inventory, language/project detection, symbols, relationships, impact hints, related tests, and change-risk signals.
Text tools — tree, search, paged reads, exact edits, writes, and unified patches.
Image tools — image metadata, safe preview, rectangular crop, and tiled high-detail reads for JPEG/PNG/WebP files.
Verification — safe/full/off command modes plus Git-aware
show_changesreview.HTTP + stdio transports — local MCP clients can use stdio; ChatGPT can connect through an HTTPS tunnel.
OAuth for public HTTP mode — OAuth discovery/PKCE is available when AgentDock has a stable public URL.
Multiple workspaces — open, list, close, and switch the default workspace inside an allowed root set.
Optional local Codex history reads — disabled by default and only enabled explicitly.
Linux service support — saved stable profiles can run as a per-workspace
systemd --userservice.Windows Control Center — a bundled desktop UI can change Default Root, Allowed Roots, Bash Mode, Tool Mode, and Write Mode without restarting the MCP server.
Requirements
For the recommended Windows + ChatGPT + ngrok setup, install:
Git — used to clone AgentDock and inspect workspace changes.
Node.js 20+ —
node -vshould report 20 or newer.Python 3 — used by the image/PDF helper workers.
PyMuPDF — required for PDF text/page rendering and image processing:
python -m pip install pymupdf.ngrok — recommended for a stable public HTTPS address that ChatGPT can reach.
A ChatGPT account/workspace that exposes Developer Mode / custom MCP app creation. Availability and UI labels vary by plan and rollout; see the current OpenAI help page linked below.
You do not need to buy a domain from Namecheap for the recommended path. A free ngrok account currently includes one account-assigned development domain. If you already own a custom domain, the Cloudflare named-tunnel route is an optional alternative.
Windows Control Center
On Windows 10/11, AgentDock includes a desktop Control Center for the five most frequently changed workspace and permission settings:
Default Root
Allowed Roots
Bash Mode
Tool Mode
Write Mode
After cloning the repository:
npm ci
npm run control-center:installThe install command builds AgentDock, creates an AgentDock Control Center desktop shortcut, and installs a background supervisor at Windows sign-in.
For reboot-safe startup, install a normal system/user Node.js 20+ runtime first. The installer validates Node candidates, rejects temporary/cache runtimes such as Codex .cache/codex-runtimes, verifies the runtime dependencies, and writes a stable Node executable into the generated shortcuts. The repository itself contains no hard-coded Node path or user home path.
npm run control-center:doctor verifies the selected Node runtime, npm availability, AgentDock build, runtime dependencies, saved workspace profile, Windows sign-in shortcut, and the local Control Center/AgentDock endpoints.
When AgentDock is already running, the Control Center hot-reloads the five core settings through the local runtime configuration endpoint. The AgentDock process, tunnel, port, and MCP session stay online, so an existing ChatGPT Web conversation can continue using the same connection after the settings change.
If this is a fresh machine with no saved AgentDock workspace profile yet, run the normal setup once before relying on the Control Center to start AgentDock:
node scripts/local-workspace-bridge.mjs setupThen verify reboot-safe startup:
npm run control-center:doctorSee control-center/README.md for install, uninstall, startup-supervisor, and verification details.
Zero-to-working deployment: Windows + ngrok + OAuth
This section is intentionally verbose. It is written so that a person—or a coding agent such as Codex/Claude Code—can follow the deployment from a clean machine without guessing missing steps.
Step 0 — Know what will happen
The final connection looks like this:
ChatGPT
↓ HTTPS + OAuth
your ngrok development domain
↓
AgentDock HTTP/MCP server on your computer (default port 8787)
↓
one explicitly allowed local workspaceThere are three different credentials/identifiers that are easy to confuse:
Item | Where it comes from | Where it is used | Keep private? |
ngrok authtoken | ngrok Dashboard |
| Yes |
ngrok dev domain | ngrok Dashboard → Domains | AgentDock | No, it is a public hostname |
AgentDock OAuth approval key | AgentDock setup/start output | AgentDock consent page opened during ChatGPT OAuth | Yes |
The ngrok authtoken and AgentDock approval key are not the same thing.
Step 1 — Clone AgentDock
Open PowerShell, Windows Terminal, or another normal terminal:
git clone https://github.com/Nippori709/AgentDock.git
cd AgentDockIf the repository already exists locally, update it instead:
git pullStep 2 — Install Node/Python dependencies and build
From the AgentDock repository:
npm install
python -m pip install pymupdf
npm run buildThe build must finish without TypeScript errors. You can run the full local smoke suite as an optional sanity check:
npm run smokeFor easier daily use, you may also install this local checkout as a global CLI:
npm install -g .That exposes the command local-workspace-bridge. If you skip this optional step, use node scripts/local-workspace-bridge.mjs ... in the commands below.
Step 3 — Create a free ngrok account and get your two ngrok values
Create/sign in to an ngrok account.
Open the ngrok Dashboard.
Find your authtoken. This is a long secret string used only to authenticate the ngrok agent running on your computer.
Open Domains and copy the development domain assigned to your account.
Do not invent the domain or blindly copy the suffix from this README. ngrok has changed domain suffixes over time. Depending on the account, you may see something such as:
example-name.ngrok-free.appor an older assigned domain such as:
example-name.ngrok-free.devCopy exactly what the ngrok Dashboard shows for your account. Free plans generally provide an assigned development domain; choosing a custom name or bringing your own custom domain may require a paid ngrok plan.
Namecheap is not part of this recommended free path. Namecheap only becomes relevant if you separately buy/own a custom domain and choose an advanced custom-domain deployment.
Step 4 — Install and authenticate ngrok on Windows
ngrok currently recommends the Microsoft Store/WinGet path on Windows. One option is:
winget install ngrok -s msstoreYou can also install it from ngrok's official download page. After installation, verify:
ngrok versionThen add the ngrok authtoken from Step 3:
ngrok config add-authtoken "YOUR_NGROK_AUTHTOKEN"Do not paste this token into README files, source code, screenshots, issues, or ChatGPT prompts.
Step 5 — Run the AgentDock setup wizard
The easiest and least error-prone path is the interactive wizard:
node scripts/local-workspace-bridge.mjs setupIf you installed the global CLI in Step 2, this is equivalent to:
local-workspace-bridge setupThe wizard asks several questions. For the normal ChatGPT + ngrok setup, use the following answers.
5.1 Where is your project located?
Enter the folder that you actually want ChatGPT to inspect/edit, for example:
D:\Projects\MyAppIf you are intentionally exposing AgentDock itself, the AgentDock repository path is fine. Otherwise, do not accidentally expose the AgentDock source folder when you meant to expose another project.
5.2 Which local port should LocalWorkspaceBridge use?
Press Enter to keep the default unless port 8787 is already occupied.
Typical answer:
87875.3 Public access: quick, stable, ngrok, tailscale, or local?
For this guide, enter:
ngrokThe choices mean:
quick— temporary Cloudflare quick tunnel; easiest demo, but URL changes after restart.ngrok— recommended here; reuse your account's stable ngrok development domain.stable— Cloudflare named tunnel using a domain you control.tailscale— Tailscale Funnel.local— no public tunnel; ChatGPT on the web cannot normally reach127.0.0.1directly.
5.4 Ngrok domain or URL, without /mcp
Paste the exact development domain copied from the ngrok Dashboard, for example:
example-name.ngrok-free.appDo not append /mcp; AgentDock adds the MCP route itself.
5.5 LocalWorkspaceBridge auth token for this workspace
The wizard provides a generated random default. For a normal setup, press Enter to accept it.
This saved secret is used by AgentDock's public HTTP authentication. When OAuth is active, the launcher prints it as the local OAuth approval key that you enter on the AgentDock consent page.
Do not publish this value.
5.6 Save this setup for future runs from this workspace?
Answer:
yesAgentDock saves the per-workspace profile under:
~/.local-workspace-bridgeThat is why later launches can reuse the same workspace, tunnel type, hostname, port, and local auth secret without asking everything again.
5.7 Start LocalWorkspaceBridge now?
Answer:
yesStep 6 — Read the startup output carefully
A successful public OAuth launch prints a block containing values similar to:
LocalWorkspaceBridge ready
Workspace D:\Projects\MyApp
Server URL https://YOUR_ASSIGNED_NGROK_DOMAIN/mcp
Authentication: OAuth
OAuth approval key (enter once on the LocalWorkspaceBridge consent page):
<private-generated-key>Two values matter now:
Server URL — copy the full HTTPS URL ending in
/mcpinto ChatGPT.OAuth approval key — keep the terminal open or copy the key somewhere private temporarily. You will enter it on AgentDock's own consent page during OAuth.
Do not append ?token=... to the Server URL. Query-string credentials are disabled by default and are easier to leak through browser history/logs.
Step 7 — Create the custom MCP app in ChatGPT
ChatGPT's UI is changing over time and can differ by plan/workspace. Current OpenAI documentation generally describes the path as Settings → Apps → Create (or the equivalent workspace Apps page). Some accounts/builds may still expose a Developer Mode / Plugins → + flow.
Use whichever custom-MCP creation UI your account exposes, then enter:
Name: AgentDock
Description: Local coding workspace bridge for ChatGPT
Connection: Server URL
Server URL: https://YOUR_ASSIGNED_NGROK_DOMAIN/mcp
Authentication: OAuthIf the UI has Scan Tools, start the scan. Because the server uses OAuth, ChatGPT should open/redirect to AgentDock's authorization page.
Current OpenAI help for Developer Mode/custom MCP apps:
Plan availability changes over time. If your ChatGPT account does not offer custom MCP app creation or blocks write/modify tools, check the current OpenAI plan/workspace requirements before debugging AgentDock itself.
Step 8 — Approve the OAuth request
On the AgentDock consent page:
Confirm that you initiated this connection from ChatGPT.
Check the requested client/redirect information.
Paste the AgentDock OAuth approval key printed in your local terminal.
Approve the request.
Do not paste your ngrok authtoken here. The consent page wants the AgentDock approval key, not the ngrok credential.
After approval, ChatGPT completes OAuth/PKCE and receives an access token. The MCP Server URL itself stays credential-free.
Step 9 — Verify the connection in ChatGPT
Start with a read-only request:
Use AgentDock to open the current workspace, inspect the repository, and summarize the architecture. Do not edit files.Then verify a harmless tool such as tree, search, or read. Only after the read path works should you try write/edit/bash operations.
Step 10 — Daily startup after the first setup
If the profile was saved, you do not need to repeat the ngrok account setup or ChatGPT app creation each day.
From the AgentDock checkout:
node scripts/local-workspace-bridge.mjs start --root "D:\Projects\MyApp"Or, if you installed the global CLI:
cd D:\Projects\MyApp
local-workspace-bridge startThe saved ngrok hostname remains the same, so the ChatGPT Server URL can remain unchanged.
Non-interactive ngrok launch
Once ngrok is installed/authenticated and you know the assigned domain, you can skip the wizard and launch explicitly:
node scripts/local-workspace-bridge.mjs ngrok --root "D:\Projects\MyApp" --hostname YOUR_ASSIGNED_NGROK_DOMAINFor reproducible automation, prefer a saved workspace profile after the first interactive setup rather than putting private auth values directly on command lines.
What a coding agent can and cannot automate
If you ask Codex, Claude Code, or another local coding agent to deploy AgentDock from this README, it can usually automate:
cloning/pulling the repository;
checking Node.js, Python, Git, and ngrok availability;
running
npm install, PyMuPDF installation, build, and smoke tests;starting the AgentDock setup wizard or constructing the equivalent launch command;
diagnosing local port/process/build problems.
It normally cannot safely invent or bypass the account-interaction steps. Expect to perform or provide these yourself:
Sign in/create the ngrok account.
Copy the account's ngrok authtoken.
Copy the exact assigned ngrok development domain.
Enable the relevant Developer Mode/custom-app feature in ChatGPT if your plan/workspace requires it.
Create/authorize the ChatGPT MCP app.
Enter the locally printed AgentDock OAuth approval key on the consent page.
A good prompt for a coding agent is:
Clone https://github.com/Nippori709/AgentDock and deploy it by following README.md exactly.
Use the recommended Windows + ngrok + OAuth path.
Do not invent a domain, token, approval key, or ChatGPT account setting.
Perform all local terminal steps yourself. When an ngrok/ChatGPT account action is required,
tell me exactly what page/value you need, then continue from the value I provide.
Run the documented build and smoke checks before declaring the deployment complete.Tunnel choices
Recommended: ngrok assigned development domain
Use this when you want the simplest persistent URL. The free plan currently includes an account-assigned development domain. Your actual hostname may use a suffix different from examples in older AgentDock releases, so always copy it from the ngrok Dashboard.
Temporary demo: Cloudflare quick tunnel
For a quick test with no stable URL requirement:
node scripts/local-workspace-bridge.mjs startThe quick-tunnel URL can change after restart. If ChatGPT was configured with that URL, you must update the app/connector when the URL changes.
Advanced: your own domain + Cloudflare named tunnel
If you already own a domain (for example one purchased from Namecheap), you can use a Cloudflare named tunnel. The domain must be configured appropriately in Cloudflare DNS. This is an advanced alternative; buying a Namecheap domain is not required for AgentDock.
Local-only
For MCP clients running on the same computer:
node scripts/local-workspace-bridge.mjs start --tunnel noneThis does not provide a public HTTPS URL for ChatGPT web.
OAuth vs credentials in the URL
Prefer:
https://YOUR_ASSIGNED_NGROK_DOMAIN/mcp
+ OAuth / PKCEinstead of:
https://YOUR_ASSIGNED_NGROK_DOMAIN/mcp?token=...Query-string credentials are disabled by default, can leak through browser/history/log surfaces, and are unnecessary when the stable public URL uses AgentDock's OAuth flow.
Tool modes
LOCALWORKSPACEBRIDGE_TOOL_MODE controls how much of the MCP surface is advertised:
minimal— tight direct coding loop: workspace open, read/image, write/edit/patch, bash, change review.standard— default; adds repository analysis, tree/search, skills, and workspace management.full— adds advanced diagnostics, Git details, snapshot/inventory tools, and optional Codex-session tools.
Write behavior is separate:
LOCALWORKSPACEBRIDGE_WRITE_MODE=workspace # direct workspace edits
LOCALWORKSPACEBRIDGE_WRITE_MODE=off # read-onlyBash behavior is also separate:
LOCALWORKSPACEBRIDGE_BASH_MODE=safe
LOCALWORKSPACEBRIDGE_BASH_MODE=off
LOCALWORKSPACEBRIDGE_BASH_MODE=fullUse full only for repositories and commands you trust.
Repository analysis
inspect_workspace builds a bounded local map from manifests, source declarations, imports, tests, and Git state. Structured search can also use symbol/reference/impact intents.
The analysis layer provides evidence to the model; it does not decide or execute an implementation plan on the model's behalf.
Image access
Large images should not be pushed into model context at full resolution by default. AgentDock exposes a progressive path:
image_info → read_image preview → read_image_crop / read_image_tile when detail is neededImage paths are subject to the same workspace guards as text files.
PDF and DOCX access
AgentDock can read PDF text with read_pdf, render a PDF page for visual inspection with read_pdf_page, and extract paragraphs/table rows from Word .docx files with read_docx.
PDF support uses PyMuPDF on the machine running AgentDock:
python -m pip install pymupdfDOCX extraction reads OOXML directly and does not require Microsoft Word or an extra Python package.
Security model
Important defaults:
roots must be explicitly allowed;
.env, private keys, credential stores, dependency/build directories, and other blocked paths are denied;secret-looking content is redacted on reads and blocked on writes;
public/non-loopback HTTP use requires authentication unless explicitly overridden;
query-string credentials are disabled by default;
safe Bash blocks high-risk shell patterns and environment expansion;
logs are designed not to contain source contents, prompts, credentials, or full command output by default.
AgentDock is a developer bridge, not an OS sandbox. See SECURITY.md.
Verification
npm run build
npm run smoke
npm run benchmark:quickThe benchmark is local and reproducible. It reports tool-schema size, a shared direct-agent workflow, security boundary checks, and generated-repository inventory/search timing. Machine-specific benchmark output is printed rather than committed.
Linux
See LINUX_SERVER.md for a headless systemd --user setup after you have created a stable tunnel profile.
Configuration
See config.example.env. Prefer per-workspace saved profiles over committing machine-specific settings.
Attribution
AgentDock is a derivative work based on the open-source CodexPro project. The original CodexPro copyright notice and MIT license are preserved in LICENSE.
This derivative focuses on a single direct ChatGPT-to-workspace agent loop, repository/image evidence, authentication hardening, and reproducible local evaluation. It should not be represented as a from-scratch implementation of the upstream project.
License
MIT. See LICENSE.
Available Tools
22 toolsapply_patchApply PatchADestructive
Apply one unified diff patch inside the workspace. Paths are validated before applying. Prefer edit for tiny replacements and apply_patch for multi-file diffs.
Inputs: patch (required), workspace_id.
bashBashADestructive
Run one allowlisted verification command in the workspace, such as tests, build, lint, typecheck, or a project script. Do not use for git status/diff or file inspection; use show_changes, tree, search, and read instead. Do not chain commands with &&, pipes, redirects, or shell file readers.
Inputs: command (required), cwd, session_id, timeout_ms, +1 more.
close_workspaceClose WorkspaceARead-only
Forget one opened workspace from this server session. This does not delete files.
Inputs: workspace_id (required).
editEdit FileADestructive
Apply a targeted exact text replacement inside a workspace text file. Returns a unified diff.
Inputs: path (required), new_text (required), old_text (required), replace_all, +2 more.
image_infoImage InfoARead-only
Inspect a workspace-local image without returning its pixels. Reports original dimensions/bytes and a recommended 1600 px-safe tile grid for high-detail follow-up reads.
Inputs: path (required), tile_overlap, workspace_id.
inspect_workspaceInspect WorkspaceBRead-only
Build a bounded repository map with languages, project types, entrypoints, areas, symbols, relationships, and coverage warnings.
Inputs: mode, path, cursor, max_files, +5 more.
list_workspacesList WorkspacesARead-only
List currently opened LocalWorkspaceBridge workspaces for this server/config.
Inputs: no named parameters.
load_skillLoad SkillARead-only
Load the bounded SKILL.md body for a discovered workspace, user, or plugin skill by name. Does not accept arbitrary paths; use after open_current_workspace/open_workspace shows skill_inventory.
Inputs: name (required), path, source, max_bytes, +3 more.
local-workspace-bridgeLocalWorkspaceBridge SupertoolADestructive
Stable wrapper for advanced ChatGPT connector setups. Pass action plus args to call an already-registered LocalWorkspaceBridge tool without changing the visible schema; it cannot call tools disabled by the current mode.
Inputs: args, action.
local_workspace_bridge_self_testLocalWorkspaceBridge Self TestARead-only
Run a controlled local diagnostic for workspace access, tool registration, skills, git, and the configured bash policy without modifying workspace files.
Inputs: bash_probe, max_skills, workspace_id, include_global_skills.
open_current_workspaceOpen Current WorkspaceARead-only
Use this once at the start to open the configured default workspace without accepting a path. Do not call open_workspace after this unless switching roots.
Inputs: max_depth, include_tree, include_skills, include_global_skills.
open_workspaceOpen WorkspaceARead-only
Open a local project directory as a LocalWorkspaceBridge workspace. Returns a workspace_id plus git status, AGENTS.md, and a compact file tree.
Inputs: path, root, max_depth, max_files, +3 more.
readRead FileARead-only
Read a specific text file with line numbers. Avoid rereading files after write/edit/apply_patch unless exact final content is needed.
Inputs: path (required), end_line, max_bytes, start_line, +1 more.
read_imageRead ImageARead-only
Read a workspace-local JPEG, PNG, or WebP as native MCP image content. Safe default previews use a 1600 px longest side; values above 1600 up to 4096 are experimental in ChatGPT. For pixel-level detail, prefer image_info plus read_image_crop/read_image_tile.
Inputs: path (required), workspace_id, max_dimension, return_original.
read_image_cropRead Image CropARead-only
Read a rectangular crop directly from the original image, rendered at a safe maximum longest side of 1600 px. Coordinates are original-image pixels from the top-left corner.
Inputs: x (required), y (required), path (required), width (required), +3 more.
read_image_tileRead Image TileARead-only
Read one tile from an automatically recommended high-detail grid. Tiles come directly from original pixels, default to 8% overlap, and are kept within the safe 1600 px longest-side path.
Inputs: row (required), path (required), column (required), overlap, +2 more.
searchSearch FilesARead-only
Use this for targeted verification or code lookup. Prefer one specific final search instead of repeated broad verification searches.
Inputs: query (required), glob, path, regex, +6 more.
server_configServer ConfigARead-only
Show LocalWorkspaceBridge server configuration, safety modes, limits, and blocked paths. Does not reveal auth tokens.
Inputs: no named parameters.
set_default_workspaceSet Default WorkspaceBRead-only
Set an already-opened workspace as the default and active workspace for omitted workspace_id calls.
Inputs: workspace_id (required).
show_changesShow ChangesARead-only
Summarize the current workspace changes in one review-oriented result with git status, diff stats, and optional diff. Use this instead of bash git status, bash git diff, git_status, or git_diff when reviewing work.
Inputs: path, since, staged, include_diff, +2 more.
treeFile TreeBRead-only
List files and directories inside the workspace, excluding blocked paths.
Inputs: path, max_depth, max_entries, workspace_id, +1 more.
writeWrite FileADestructive
Create or overwrite a meaningful text file inside the workspace. Returns a unified diff; do not create empty placeholder files.
Inputs: path (required), content (required), overwrite, create_dirs, +1 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 22 tools
Most tools target distinct operations—workspace management, file read/write/edit, image inspection, and change review—with descriptions that explicitly separate them. The main sources of ambiguity are the generic local-workspace-bridge wrapper, which can call other tools indirectly, and the write/edit/apply_patch cluster, though their intended use cases are clarified.
The vast majority of tools follow a clear verb_noun snake_case pattern like open_workspace, read_image, and show_changes. The hyphenated local-workspace-bridge and the single-word bash/tree are noticeable outliers, but they are isolated deviations rather than a systematic mixing of conventions.
22 tools is on the heavy side and exceeds the typical 3–15 tool sweet spot. The count is somewhat justified by the broad scope—workspace lifecycles, file editing, image handling, diagnostics, and verification—but it still feels like more surface area than necessary.
The toolset covers workspace open/list/close/default management, file tree/search/read/write/edit/apply_patch, image inspection and cropping, bash verification, change summarization, skills, and self-test/config. Obvious gaps like an explicit delete/rename tool or git commit are absent, but apply_patch can handle file deletion and the remaining gaps are workable.
Maintenance
Related MCP Connectors
Connect ChatGPT and Claude to your Mac, Windows or Linux with secure MCP remote access.
Connect ChatGPT, Claude, Cursor or VS Code to your Windows PC: files, terminal, browser, apps.
Use your Mac, Windows or Linux computer from ChatGPT, Claude or Codex: files, commands, documents.
Share one project context across ChatGPT, Claude, Telegram and any MCP client.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables ChatGPT to inspect and edit local projects through a secure MCP interface, offering workspace management, file operations, git integration, and safe command execution.6MIT
- AlicenseAqualityCmaintenanceEnables ChatGPT Developer Mode to operate as a local coding agent by connecting to your own repository through MCP, allowing it to read, write, search, edit files, run safe verification commands, and manage handoff context for continued work.15122 npmMIT
- AlicenseNot gradedqualityCmaintenanceEnables ChatGPT and other MCP clients to browse, search, edit and upload files, run commands and interactive terminals, inspect Git history and worktrees, dispatch coding agents, and install composable plugins within locally registered workspaces. Requests execute under the user's own machine with per-request approval or auto-approval, keeping files, tasks and results local unless explicitly returned.1MIT

Flyto2 Runtimeofficial
AlicenseNot gradedqualityBmaintenanceGives ChatGPT secure access to your local machine over MCP, letting it read and edit files, run commands, tests, and Git within approved workspaces.1MIT