ai-manager

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

AI Manager

A calm desktop manager for AI coding tools

AI Manager installs, updates, configures, and repairs AI coding command-line tools on your own computer, so you never have to touch shell profiles, PATH, or hand-edited configuration files. It is a local-first desktop application built with Tauri 2 and React 18, released under the GNU AGPL-3.0-or-later.

[!NOTE] Every installer on the Releases page passes the signing and verification gates in the release runbook before it is published. A local build is not a release and must not be redistributed as one.

Who it is for

People who use Claude Code, Codex, or a similar tool every day and would rather not manage npm ownership, configuration files, and API endpoints by hand. One place to see which tools are installed, which AI service each one is really using, and whether that setup still works.

What it does

Seven destinations in a flat sidebar, with Settings in the footer.

Home summarises what is installed, what has an update, and what needs attention. Quick Check is built from real local signals: it rechecks saved service addresses only when you ask, and never claims a process is healthy when no such signal exists.

Software detects, installs, updates, and repairs ten command-line tools without crossing away from the npm, pnpm, bun, or Volta owner that already manages them. It keeps a verified version history and shows bounded, redacted process logs instead of raw shell output. Downloads are official-first, with one process-scoped retry on a recoverable network failure and a local proxy option in Settings. Desktop apps are detected separately; installing and updating those stays with each vendor’s signed channel.

API Endpoints connects a tool to an AI service from a reviewed preset catalogue or a custom HTTPS endpoint, then shows the effective connection: the endpoint the tool will really use, and why, including environment variable overrides. Testing a service reads the models it actually serves, sends one real request to the one you pick, and shows the reply or the generated image, so a rejected key no longer reads as a healthy address. Your shell environment is never edited. Local Routing and a 30 day local usage overview are secondary tabs.

Skills scans every supported local scope in one overview, installs from trusted GitHub catalogues or a local ZIP, copies a Skill to another tool, and keeps recovery copies before removal.

MCP adds stdio, HTTP, and SSE connections without pasting JSON, and enables or disables them per tool. Adoption of a tool’s existing servers never rewrites live configuration.

Global Prompts creates, edits, imports, and switches prompts without opening a configuration file, and reveals where each one lives.

Sessions browses and searches local sessions. Content loads only after selection, and supported sessions resume in a terminal through a command line built in the backend.

Settings holds language and appearance, desktop behaviour, the local download proxy, database backups, configuration import and export, and the fail-closed signed updater.

Supported tools

Command-line tools	Desktop apps
Claude Code	Codex App
Codex	Claude Desktop
OpenCode	Cursor
Gemini CLI	ZCode
Grok Build	Cherry Studio
OpenClaw
Hermes
Pi
Kimi Code
DeepSeek Harness (DSH)

Per-tool behaviour is driven by a capability registry rather than scattered conditionals, so a tool that lacks a concept (for example a “current service”) simply does not show that control.

Install

Download the installer for your platform from the Releases page.

Platform	Installer	Automatic updates
macOS Apple Silicon	Signed, notarized DMG	Signed app archive
macOS Intel	Signed, notarized DMG	Signed app archive
Windows x64	MSI, verify the SHA-256	Minisign-signed MSI
Linux x64	AppImage and deb with SHA-256	Minisign-signed AppImage

Windows installers are not Authenticode-signed, so SmartScreen shows an unknown publisher warning on first run; check the published SHA-256 before installing. Linux has no cross-distribution equivalent of platform signing either. On both, the minisign signature on updates and the published checksum are the integrity evidence.

Every release asset ships with a checksum, and the in-app updater verifies a minisign signature against the public key embedded in the application before anything is installed. Linux ARM, Windows ARM, rpm, universal macOS builds, the Mac App Store, and the Microsoft Store are outside the current release scope.

Build from source

Prerequisites:

• Node.js 20 or newer (.node-version records the version used locally)
• pnpm 10.12.3, the version pinned by packageManager (corepack enable)
• the Rust toolchain pinned by rust-toolchain.toml
• the platform packages from the Tauri 2 prerequisites guide

pnpm install --frozen-lockfile
pnpm dev            # desktop application with hot reload (tauri dev)
pnpm dev:renderer   # renderer only, in the browser (vite)
pnpm build          # production bundle for the current platform (tauri build)

The design-system gallery is development-only at http://localhost:3000/?gallery and is excluded from production bundles.

Before opening a pull request, run the checks listed in CONTRIBUTING.md.

Project structure

src/             React renderer: app shell, pages, features, entities, shared UI,
                 the native IPC client (src/native), and locales (src/i18n)
src-tauri/src/   Rust backend: commands, application, domain, adapters,
                 repositories, platform, infrastructure, and the
                 compat/ccswitch compatibility layer
docs/            ARCHITECTURE.md, ADRs, the product design specification, and
                 the release runbook
scripts/         release, brand, and boundary tooling
tests/           Vitest suites for the renderer

Start with docs/ARCHITECTURE.md for the layer model and directory boundaries, docs/adr/ for the decisions behind them, and docs/product/design-spec.md for the product specification. AI_RULES.md lists the non-negotiable coding rules that apply to every contributor.

Privacy and data safety

• AI Manager contains no analytics or telemetry SDK.
• All product data lives in the application’s local data directory. An existing upstream database is only ever read, never modified.
• Logs and crash records are size-bounded and redact known secret forms.
• Provider credentials are still stored in plaintext inside the local database, local backups, and SQL configuration exports. Treat those files as sensitive; an OS secret store and an encrypted portable archive are planned but not implemented yet.

Upstream

AI Manager began as a fork of CC Switch, copyright 2025 Jason Young, MIT License. Inherited code keeps its original paths so that upstream fixes can still be cherry-picked, and everything this project added sits beside it in the product layers. See ADR-0001, ADR-0003, and THIRD_PARTY_NOTICES.md.

License

GNU AGPL-3.0-or-later. Copyright 2026 AI Manager contributors.

Portions derived from the upstream project remain under their original MIT License, copyright 2025 Jason Young. MIT permits sublicensing, so the combined work is distributed under the AGPL with the upstream notices preserved in full.