CodingFleet Desktop MCP server
Commit Nov 15, 2026An AI coding agent that reads, edits, and runs code in local directories while requesting user approval for all changes.
Also known as x4nth055/codingfleet-desktop
- Stars
- 1
- Tools
- 9
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 15, 2026
- License
- MIT
What is CodingFleet Desktop MCP?
CodingFleet Desktop is an AI coding agent that reads, edits, and runs code in local directories while requesting user approval for all changes. It exposes 9 tools, such as fs_read, fs_glob, fs_write.
Summary generated from the project's README
About CodingFleet Desktop MCP server
CodingFleet Desktop provides an AI coding agent that operates directly on your local files to read, edit, and execute code. Designed for developers who want to automate complex tasks like migrating codebases to TypeScript or fixing UI layouts from screenshots, it runs on Windows, macOS, and Linux. You can choose to run tools on your local machine to access your specific folders and local MCP servers, or use a cloud sandbox for a safe, isolated environment.
Control is maintained through a strict approval system where the agent must ask for permission before changing files, running most commands, or accessing sensitive data like environment variables. The interface includes a built-in browser for testing dev servers, a terminal, and an undo feature that can revert file changes from the last 20 runs. Users can steer the agent during active tasks and manage costs with real-time session estimates.
What you can do with it
- 1Migrate codebases to TypeScriptYou can task the agent with moving an entire project to TypeScript using multiple sub-agents working in parallel.
- 2Fix UI from visual feedbackA developer can provide a screenshot to have the agent identify and fix layout issues in the code.
- 3Optimize slow API endpointsThe agent can hunt down performance bottlenecks and implement code changes to speed up response times.
- 4Test web applications automaticallyThe agent can use an integrated browser to open a local dev server, click elements, and verify builds.
Written from the project's README. Every use case is checked against a quote from it.
CodingFleet Desktop MCP tools 9
Detected from the READMEfs_readfs_globfs_writefs_editrun_commandexecute_codeview_imagetake_screenshotbrowser
| fs_read | — | |
| fs_glob | — | |
| fs_write | — | |
| fs_edit | — | |
| run_command | — | |
| execute_code | — | |
| view_image | — | |
| take_screenshot | — | |
| browser | — |
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablex4nth055/codingfleet-desktop is public and reachablegithub.com/x4nth055/codingfleet-desktop ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/x4nth055/codingfleet-desktop ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model9 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubCodingFleet Desktop
An AI coding agent that works on your real code, on your computer, and asks before it changes anything.
Download · Features · Tools · Build it yourself · codingfleet.com
Three jobs on low-cost models: fix a layout from a screenshot, hunt down a slow endpoint (p99 3,152 ms → 36 ms), and move a codebase to TypeScript with four sub-agents in parallel. Watch the full video.
Download
| System | File | |
|---|---|---|
| 🪟 | Windows 10 / 11 | CodingFleet-Setup-<version>.exe |
| 🍎 | macOS (Apple silicon / Intel) | CodingFleet-<version>-arm64.dmg / -x64.dmg |
| 🐧 | Linux | .AppImage or .deb |
Open the app, click Sign in with your browser, and approve. That's all.
Windows: until code signing is in place, Windows may say "Windows protected your PC". Click More info → Run anyway.
macOS: the app is not signed by Apple. The first time, open it from System Settings → Privacy & Security → Open Anyway.
Features
| 🧑💻 Works on your real code | Reads, edits and runs things in the folder you pick. |
| 🛡️ You stay in control | Every command and every file change waits for your OK. |
| ↩️ Undo | Puts back every file a run changed, in one click. |
| 🧭 Steer a run | Send a correction while the agent works. |
| ☁️ Cloud sandbox | Or let the agent work on CodingFleet's servers instead of your PC. |
| 🔌 MCP | 20 built-in connectors (GitHub, Context7, Stripe, Sentry, Notion, Linear…) and your own local MCP servers. |
| 🖼️ Images and voice | Paste images, take screenshots, speak your prompt. |
| 💳 Cost before you send | An estimate next to the Send button, and what the session has cost so far. |
| @ Mention files | Type @ to pick a file or folder of the project. |
| 🌐 Browser | A browser beside the conversation. The agent opens your dev server there to check what it built: it reads the page, clicks, types, and looks at it. |
| 🖥️ Terminal | Your own shells beside the conversation, at the bottom or on the right. Ctrl+` shows and hides it. |
| ⌨️ Keyboard first | Answer approvals with Alt+Y / Alt+N, search sessions, find in a conversation. Ctrl+/ lists every shortcut. |
| 📄 Export | Copy an answer as Markdown, or save a whole session as a .md file. |
| 📂 Files in answers | Click a file the agent links to: code opens in the side panel, pictures in the viewer, pages and PDFs in the browser. Right-click for Show in Explorer, Open with default app, Copy path. |
| 🪶 Light on memory | Only the conversation on screen is drawn, and only its newest turns until you scroll up. Sessions you left are kept compressed, and reopen from memory without asking the server. |
| 🎨 Three themes | Dark, Light and Hacker, and a choice of fonts (Inter and JetBrains Mono unless you pick others). |
Tools
What the agent can use on your computer:
| Tool | What it does |
|---|---|
fs_read |
Read a file |
fs_glob |
Find files by name |
fs_write |
Write a file |
fs_edit |
Change part of a file |
run_command |
Run a shell command (also in the background, for dev servers) |
execute_code |
Run a piece of code |
read_background / stop_background |
Read or stop a background command |
view_image |
Look at an image in your folder |
take_screenshot |
Take a picture of your screen (always asks first) |
browser |
Use the app's browser: open a page, read it, click, type, take a picture, read the console |
It also has CodingFleet's server tools, such as web search and image generation, if you turn them on.
Where the tools run
Each session picks one place when you start it:
| Your folder | Cloud sandbox | |
|---|---|---|
| Tools run on | your computer | CodingFleet's servers |
| Sees your code | yes, in that folder | only what you upload |
| Asks before changes | yes | no (nothing local to protect) |
| Local MCP servers | yes | no |
Pick a folder to work on real code. Pick no folder for a safe, empty sandbox.
Approvals
| What the agent does | What happens |
|---|---|
| Reads files inside the folder | Runs at once |
Runs a command that only reads inside the folder (ls, grep, git diff…) |
Runs at once |
| Reads outside the folder | Asks |
| Runs a command or changes files | Asks |
Reads .env, keys or passwords |
Asks, even in Auto-approve |
Deletes files inside the folder (rm -f tmp/a.py, Remove-Item -Force a.log) |
Same as changing files |
Deletes a folder that is made again (rm -rf node_modules, .next, __pycache__, coverage…) inside the folder |
Same as changing files |
| Deletes any other folder or anything outside the folder, rewrites git history, wipes data | Asks every time |
| Runs a script downloaded from the internet | Asks every time |
In a script, or on a line with variables or a loop (for f in *.log; do rm -f "$f"; done),
the app cannot tell where each file is. Deleting files there is treated like changing files,
unless a path is plainly somewhere else (/, ~, .., $HOME, another drive); deleting
folders with -r still asks every time.
A variable counts as a secret when it is written like an environment variable
($GITHUB_TOKEN); a script's own $key or $token does not.
These checks catch mistakes. They are not a sandbox.
Commands that only read
These run at once, as long as every path they name is inside the folder:
ls,pwd,cd,cat,head,tail,wc,echo,whichgrep,rg,find,sort,uniqgit status,diff,log,show,blame,ls-files,rev-parse, andgit branchorgit remote -vto listnode,npm,python,python3andgitwith--version
Several can be joined with |, &&, || or ;, and output can go to /dev/null. Anything
more asks as before: a variable, $(…), a redirection to a file, or a flag that writes or
runs something (find -exec, sort -o, rg --pre, git diff --output). Naming a secret
such as .env still asks, and so does reading files through a wildcard that could match one
(cat .env*, cat *.pem).
On Windows without Git for Windows, commands run in PowerShell, and these run at once:
Get-ChildItem(ls,dir),Get-Content(cat,type),Select-String,Get-Item,Test-Path,Resolve-Path,Get-LocationandSet-Location(cd)Sort-Object,Select-Object,Where-Object(without{ }),Measure-Object,Group-Object,Format-Table,Format-List,Out-String,Out-Null,Write-Output- the same
gitcommands,rg, and the same--versionchecks
They can be joined with | or ;, and output can go to $null. A variable, $(…), { },
&, a redirection to a file, any other drive than a disk (env:, HKLM:) and anything not
listed (ForEach-Object, Remove-Item, Set-Content, Out-File) asks.
Background commands
Dev servers and watchers keep running in the background. A bar above the prompt shows them, with a Stop button each. Click one to watch its output as it comes. Once a dev server prints its local address, an Open button opens it in your browser.
The agent can give a background command a time limit (stop_after, in seconds): it is
stopped when the time is up.
A command the agent waits for can move there too. After it has run 10 seconds, its step offers Run in background: the agent gets what the command printed so far and its id, and continues while the command keeps running.
Ask before changes, or Auto-approve
Switch with the shield button.
- Ask before changes (default): you answer Allow, Deny or Allow all this session. "Allow all" allows that kind of action, for example every command.
- Auto-approve: only the protected actions above still ask. Use it only in a folder you can afford to lose.
Undoing a run
Undo on the "Edited files" card puts back every file the run changed. Files you changed yourself afterwards are not touched. Undo keeps the last 20 runs until you close the app. Git is still your real safety net.
When the computer sleeps
While a run goes, the app keeps the computer awake. If the computer goes to sleep anyway, the run stops and is kept: type continue to go on. Cloud sandbox runs do not need your computer.
The browser
The panel under the conversation (or on the right) has two views: Terminal and Browser. The browser is for your own apps: type an address, or click Open on a running dev server. The agent uses the same page, so you see what it does.
| What the agent does in the browser | What happens |
|---|---|
Opens or uses a page on this computer (localhost) |
Runs at once |
| Opens or uses a site on your list (Preferences) | Runs at once, typing included |
| Opens or uses another site | Asks once per site: Allow example.com this session, or Always allow example.com, which adds it to your list |
| Types into another site | Asks every time |
In Preferences > Agent tools, Sites the agent may use without asking holds your own sites, such as a staging server. Let the agent use any site without asking skips the per-site question; it is less safe (a page can hold instructions meant for the agent), and typing into a site that is not on your list still asks every time.
The agent can use the browser while the app is behind other windows. If the app is minimized when the agent needs to click or type in a page, the window comes back without taking the keyboard: a minimized window's page cannot be clicked.
A page left behind a closed panel for two minutes is closed, to free its memory. It comes back where it was, back and forward included, when you open the panel or the agent uses the browser.
The browser has its own profile: none of your real browser's logins or cookies, no downloads, and no camera, location or notifications. Clear browsing data in Preferences empties it. Turn Browser for the agent off in Preferences > Agent tools to keep the agent out of it.
Project instructions (AGENTS.md)
Put an AGENTS.md file in your folder, and the agent follows it:
- Run `npm test` before you say a change works.
- Never edit anything under `generated/`.
What a run costs
The estimate next to Send comes from what earlier runs cost. Above 100 credits, the app asks first and suggests cheaper models.
Run it from the source code
You need Node.js 22 or newer.
git clone https://github.com/x4nth055/codingfleet-desktop.git
cd codingfleet-desktop
npm install
npm start
Sign in with your browser, or paste an API key from codingfleet.com/agent-api. The key is encrypted by your system (Windows DPAPI, macOS Keychain, Linux libsecret/kwallet).
Windows users: install Git for Windows. The agent writes better Bash than PowerShell. Without Git, it uses PowerShell.
Tests
npm test # unit tests
npm run test:layout # layout checks
npm run test:memory # the sessions kept in memory, in the real window
npm run e2e:memory # the real app against a scripted API: memory, pictures, a run, the browser
npm run e2e # a real run on codingfleet.com (needs CODINGFLEET_API_KEY)
Build it yourself
npm run dist # Windows installer
npm run dist:mac # macOS (needs a Mac)
npm run dist:linux # Linux (needs Linux)
The files go to dist/. Close the app before you build.
How it fits together
src/core/ the agent logic, no Electron (tools, approvals, runs, MCP, undo)
src/main/ the Electron main process (window, keys, updates, crashes)
src/renderer/ the window (it never sees your API key)
test/ tests
Logs and crash reports
The app writes a log in its data folder (Preferences → Diagnostics). The log never contains your messages, files or keys. After a crash, the app asks if you want to send a report. Nothing is sent without your OK.
Releasing
Change version in package.json, commit, and push a tag:
git tag v1.0.0
git push origin v1.0.0
GitHub Actions builds Windows, macOS and Linux and publishes a release. The app gets its updates from these releases.
Code signing
- Windows: signed for free by SignPath Foundation (see
.signpath/). - macOS: not signed by Apple, so the app cannot update itself. It tells you when a new version is out.
- Linux: no signing needed.
Code signing policy
Free code signing provided by SignPath.io, certificate by SignPath Foundation.
Every signed file is built from this repository by GitHub Actions, and each release is approved before it is signed.
Privacy
This program will not transfer any information to other networked systems unless specifically requested by the user or the person installing or operating it.
In practice:
- It talks to CodingFleet only after you sign in, to do what you ask. See the CodingFleet privacy policy.
- Each run also says which settings the app runs with: Ask before changes or Auto-approve, your system and shell, which features and built-in connectors are on, which coding CLIs it found, and whether each tool call ran at once or after you answered. These are words from a fixed list and counts, never a path, a file or server name, or what is in a file.
- It checks for updates only if you said yes.
- It sends a crash report only if you click Send report.
- Images from other websites load only when you click them.
- It connects to an MCP server only if you added it.
License
CodingFleet Desktop MCP questions
What tools does CodingFleet Desktop MCP provide?
It exposes 9 tools, including fs_read, fs_glob, fs_write, fs_edit, run_command.
How does the agent handle file changes?
The agent asks for permission before every file change or command execution to ensure you stay in control.
Can I undo changes made by the AI?
Yes, you can use the undo feature to revert every file changed during a run with a single click.
Does it support MCP?
Yes, it includes 20 built-in connectors and supports your own local MCP servers.