A desktop AI client for model-hopping and messaging bridges

What it does

CodePilot is an Electron + Next.js desktop application that acts as a control center for AI agents. It connects to over 17 providers—from Anthropic and OpenRouter to Chinese platforms like Zhipu GLM and Volcengine Ark, plus local options like Ollama—and lets you switch models mid-conversation without losing context. Beyond chat, it manages persistent memory through persona files (soul.md, memory.md), runs scheduled tasks, generates images, and accepts commands from messaging apps like Telegram, Discord, QQ, and WeChat via a remote bridge.

The interesting bit

The project started as a coding assistant but morphed into a general-purpose agent desktop. Its “Assistant Workspace” uses markdown persona files and daily check-ins to build persistent memory, while a generative UI engine lets the AI spin up interactive dashboards and widgets on the fly. That combination of long-term personality state and runtime visual output is unusual for a desktop client.

Key highlights

• Provider agnostic: 17+ built-in presets including AWS Bedrock, Google Vertex, Xiaomi MiMo, and Aliyun Bailian, plus any Anthropic- or OpenAI-compatible custom endpoint.
• Session control: Pause, resume, and rewind conversations to any checkpoint; run split-screen dual sessions.
• MCP + Skills marketplace: Attach stdio, SSE, or HTTP MCP servers with runtime monitoring, and install reusable skills from skills.sh.
• Remote Bridge: Control the desktop agent from Telegram, Feishu, Discord, QQ, or WeChat on your phone.
• Local-first data: All conversation history, memory files, and usage analytics live in a local SQLite database with WAL mode.

Caveats

• The project is currently undergoing a major Runtime / Harness / macOS UI refactor in preview; the README explicitly warns against using older 0.53/0.54 builds and lists P0/P1 issues like launch failures and settings pages stuck loading.
• macOS builds are code-signed but not notarized, so Gatekeeper will likely complain; Windows and Linux builds are completely unsigned, which means SmartScreen warnings or manual security overrides are expected.
• Some advanced capabilities, like direct file editing and terminal commands, require installing the separate Claude Code CLI rather than being native features.

Verdict

Developers who want a single desktop hub to manage multiple AI providers, local memory, and third-party tools without sending everything to a browser tab should take a look. If you need a lightweight, battle-tested chat client and unsigned installers make you nervous, wait for the refactor to stabilize.