Model or dataset
op7418/CodePilot avatar
op7418/CodePilot

CodePilot: 17 providers, session rewind and a desktop shell

A multi-model AI agent desktop client — connect any AI provider, extend with MCP & skills, control from your phone. Built with Electron + Next.js.

6,520 stars695 forksTypeScriptNOASSERTION

At a glance

What is it?
CodePilot grew out of a coding tool into a general purpose desktop agent. This is a look at the provider matrix, the four files the assistant keeps, the npm scripts behind a build from source, and where in-app updates stop working.
Who is it for?
CodePilot suits someone who wants a single desktop surface for several model providers, keeps sessions rewound to a checkpoint, and stores everything in a local SQLite file rather than in someone else's account. It suits you less if you need in-app updates on Linux, which stay manual downloads, or if you need the file editing and git features, which arrive only after you install the Claude Code CLI separately.
Can I use it commercially?
Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
Is it still maintained?
Yes. The repository last received commits 54 days ago.
What is it written in?
Mainly TypeScript, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • op7418/CodePilot6,520 ★ · +68
Star history
Repository2026-09-112026-11-15
op7418/CodePilot6.5k6.5k

68 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

One Settings screen covers 17 providers and any compatible endpoint

The provider table is the reason to look at this project first, because it decides which keys you have to hold. Two are direct APIs, Anthropic and OpenRouter. Two are cloud platforms, AWS Bedrock and Vertex AI. Then there is a block of Chinese providers, Zhipu GLM in Chinese and global variants, Kimi, Moonshot, MiniMax in two variants, Volcengine Ark for Doubao, Xiaomi MiMo and Aliyun Bailian for Qwen. Local and self-hosted work goes through Ollama or LiteLLM, image generation is handled by Gemini, and the Custom row accepts any Anthropic-compatible or OpenAI-compatible endpoint, which is the escape hatch for a model the list has not caught up with.

Every provider is configured in the same place, Settings > Providers, where you pick a preset and paste a key. The other half of that design is model switching: you can change provider and model in the middle of a conversation without losing context, so a hard question can start on one model and finish on another. Reasoning effort is offered as Low, Medium, High, Max plus a Thinking mode, and attachments use a multimodal vision path.

In-app updates stop at macOS, and Linux never gets them

The download table covers three platforms, and the update story is different for each. macOS ships an Apple Silicon dmg and an Intel dmg for arm64 and x64, and the official macOS stable builds can check for, download and install signed updates from inside the app. Windows ships a single x64 exe installer, and the rule there has a condition: after you manually install the first supported Windows version, later releases can update in the app, but the packages they come from are unsigned and hosted on the official GitHub Release. Linux offers AppImage, deb and rpm for x64 and arm64, and Linux releases remain manual downloads.

That asymmetry is the practical limit of the distribution setup. A macOS user gets signed, in-app patching. A Windows user gets in-app patching only after a manual first install, and only unsigned packages. A Linux user updates by hand every time. All three point at the same releases/latest URL, so nothing about the download step is version specific, and the repository also has a CHANGELOG.md, RELEASE_NOTES.md and an electron-builder.yml at the top level for the packaging side. If your fleet is Linux, treat the update path as a separate piece of tooling you have to build.

npm run dev is a browser, electron:dev is the app

Building from source needs Node.js 18 or newer and npm 9 or newer, which ships with Node 18. The documented path is four commands:

bash
git clone https://github.com/op7418/CodePilot.git
cd CodePilot
npm install
npm run dev              # browser mode at http://localhost:3000
# -- or --
npm run electron:dev     # full desktop app

The two dev targets are the important split. `npm run dev` is next dev, so it serves the interface in a browser on port 3000 and never opens a window. `electron:dev` is a longer script that prepares the Electron build, then runs three processes at once, Next, esbuild in watch mode, and Electron itself. The packaged entry point in package.json is dist-electron/main.js, and the workspaces field covers apps/* and packages/*, so this is a monorepo rather than a single Next project.

The rest of the scripts show what a contribution is expected to keep green. Prebuild runs a safety assertion before next build, lint is eslint, typecheck is tsc with no emit, and electron:build runs the macOS universal runtime preparation, cleans the build directory, builds Next and then the Electron bundle. Three extra checks exist beyond lint and typecheck: lint:colors, lint:hooks and lint:docs-drift each run a script under scripts/, which is where most of the project-specific rules live.

The default test target never reaches Playwright

The scripts block is where this project is unusually explicit, and one line is worth reading closely:

json
"test": "npm run typecheck && npm run test:harness-boundary && npm run test:unit",
"test:e2e": "node scripts/run-playwright.mjs"

The default `test` target chains exactly three things: the TypeScript check, a harness boundary check that runs scripts/check-harness-adapter-boundary.mjs with a canonical flag, and the Node unit suite. Playwright never enters that chain. The end to end suite, the smoke suite filtered by an @smoke grep tag, the visual suite filtered by @visual, and an SDK proof of concept all exist as separate targets, as does smoke:mcp-poc which runs a TypeScript file through tsx. A playwright.config.ts and a visual-reports directory at the top of the tree back the visual suite.

The practical reading is that a green `npm test` says nothing about whether a browser flow still works. Anyone changing the interface needs the Playwright targets run separately. That is a choice about what the fast gate covers, not a defect in it, but it changes what you may conclude from CI output alone.

The assistant keeps soul.md, user.md, claude.md and memory.md

Assistant Workspace is the feature that goes furthest from a coding tool, and its whole state is four markdown files. In Settings > Assistant you choose a workspace directory and enable Onboarding, and CodePilot then creates soul.md, user.md, claude.md and memory.md at the workspace root. Around those files sit onboarding flows and daily check-ins, so the assistant is meant to accumulate preferences over time rather than start blank in each conversation. Custom skills are managed on a separate Skills page from MCP servers, and they can be project scoped or global, or installed from the skills.sh marketplace.

The persona files sit in a directory you choose, which means they are as durable and as inspectable as that directory, and as exposed to anything else that reads it. Keep a copy of them.

Rewind is offered per session, with checkpoints left undefined

Session control is the other half of the daily use pitch: pause, resume, rewind to a checkpoint, and archive. What a checkpoint captures, such as message history or files the agent edited, is worth confirming before you rely on rewind.

The rest of the interaction table is worth knowing before you script anything: three modes named Code, Plan and Ask, per-action approval against a Default or Full Access permission setting, split screen for two sessions side by side, attachments for files and images through a multimodal vision path, and slash commands including /help, /clear, /cost, /compact, /doctor and /review. Cost and token counts are charted daily, so you can see what a session cost after the fact rather than before it.

One dependency sits outside the app. Installing the Claude Code CLI with npm install -g @anthropic-ai/claude-code unlocks direct file editing, terminal commands and git operations. It is described as recommended but not required for basic chat, which means the feature list you read at the top of the project depends on a separate global npm install. Session history from that CLI can be imported as .jsonl files afterwards.

MCP servers, skills and the phone bridge are three config surfaces

Extensions are split across three places in the interface. MCP servers are added and managed on the MCP page in the sidebar, over stdio, sse or http transports, with runtime status monitoring. Skills have their own page, and can be custom, project or global, or pulled from the skills.sh marketplace. The Remote Bridge is the third surface: Telegram, Feishu, Discord, QQ and WeChat are listed, and the promise is that you send a message from your phone and get the response on your desktop.

The repository has a .mcp.json at the top level.

The bridge deserves the same caution. Sending a conversation to a chat platform means the conversation leaves the local SQLite storage that the rest of the app keeps on your machine. Local storage is SQLite in WAL mode, and the claim that all data stays on your machine holds for the app's own storage, not necessarily for a bridge conversation that has been relayed through a messaging service.

package.json says BUSL-1.1 and the repository reports no licence

The licensing is the one place where the metadata disagrees with itself. The package manifest at the top level declares version 0.67.17, marks the package private, and states the licence as BUSL-1.1. The repository metadata reports no recognised licence identifier, and the tree does contain a LICENSE file. A manifest that says BUSL-1.1 and a repository that reports nothing recognised is a discrepancy worth resolving in your own favour by reading the LICENSE file rather than trusting either field.

The version numbers are consistent, at least. The manifest reads 0.67.17, which matches the newest release tag v0.67.17, and the two tags before it are v0.67.16 and v0.67.15, so three tags landed between 5 and 22 September 2026. The last push to main was on 22 September 2026, the same day as the newest tag.

What the manifest does not describe is the shipped artefact. main is dist-electron/main.js, so the packaged application is an Electron bundle rather than the Next app, and electron-builder.yml plus the build/, resources/ and public/ directories are what turn that into an installer. Anyone auditing a downloaded build is therefore auditing an Electron application, and the interesting files are in the packaged output rather than in src/.

Editorial conclusion

CodePilot suits someone who wants a single desktop surface for several model providers, keeps sessions rewound to a checkpoint, and stores everything in a local SQLite file rather than in someone else's account. It suits you less if you need in-app updates on Linux, which stay manual downloads, or if you need the file editing and git features, which arrive only after you install the Claude Code CLI separately. Before relying on it, check three things: what license the build you downloaded actually carries, since package.json declares BUSL-1.1 while the repository reports no recognised identifier; whether the MCP servers you need are reachable from your machine at all, since the app only reports their status; and which npm scripts your contribution has to keep green, because the default test target skips the Playwright suites entirely.

Frequently asked questions

What is codepilot?

CodePilot is a multi-model AI agent desktop client built with Electron and Next.js. It connects 17+ AI providers behind one interface, supports MCP servers and skills, keeps data in local SQLite, and offers session rewind to any checkpoint. Downloadable builds exist for macOS, Windows and Linux.

Does CodePilot work as a VS Code extension?

CodePilot ships as a standalone application: installers for macOS, Windows and Linux, a browser mode served by next dev at localhost:3000, and a full desktop mode started with npm run electron:dev. A VS Code extension does not appear in the top level tree.

Which platforms can update CodePilot from inside the app?

macOS stable builds can check for, download and install signed updates in the app. On Windows, after the first supported version is installed manually, later releases update in the app using unsigned packages from the official GitHub Release. Linux releases remain manual downloads.

What licence does CodePilot ship under?

The package.json manifest declares BUSL-1.1 and marks the package private, while the repository metadata reports no recognised licence identifier. A LICENSE file exists at the top level, so read that file to decide rather than relying on either metadata field.

Official sources

  1. Issues
  2. op7418/CodePilot on GitHub
  3. Project website
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/op7418-codepilot.svg)](https://hysenlabs.com/projects/op7418-codepilot)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/op7418-codepilot.svg)](https://hysenlabs.com/projects/op7418-codepilot)

Was this useful?