Proma
把最丝滑的通用 Agent 体验带进你的工作流，为 100x 专业用户而生的未来产品，正在实现 proactive Agent 阶段。基于 Claude Agent SDK 的完整开源实践，原生支持飞书群聊调用、灵活接入任意大模型供应商 —— 让顶级 Agent 能力真正跑在你每天用的地方。
Last verified:
What is Proma?
Proma is a local-first AI desktop application that brings multi-model Chat, universal Agent capabilities, workspaces, Skills, MCP (Model Context Protocol), remote robots, and memory into one open-source client. It is designed as an Agent workspace where users can handle simple queries in Chat mode and delegate complex multi-step tasks to Agent mode, with data and configuration stored locally by default.
Key features include: Chat mode with multi-model conversation, attachment parsing, image input, Markdown/Mermaid/KaTeX/code highlighting, side-by-side conversations, system prompts, and context management; Agent mode based on Claude Agent SDK with workspace isolation, permission modes, file operations, streaming long-task output, plan confirmation, and user follow-up questions; SubAgent/Tasks for splitting complex work into child agents; per-workspace Skills and MCP Server configuration; remote robot support for Feishu/Lark, DingTalk, and WeChat bridge integration; shared memory between Chat and Agent with web search and built-in chat tools; and desktop features like auto-update, proxy settings, file preview, global hotkeys, quick task windows, system-level voice input, and light/dark/system themes.
Proma is for professionals who want to build long-term AI-powered workflows: developers, researchers, writers, analysts, and knowledge workers who need AI to take action on local files, conduct multi-step research, generate reports, and integrate with team chat platforms. It supports macOS (Apple Silicon + Intel) and Windows, with API channels from Anthropic, DeepSeek, Kimi, OpenAI, Google,智谱 AI, MiniMax, Doubao, Tongyi Qianwen, and custom endpoints.
Proma pricing
Pricing model: Freemium
Proma offers an open-source version available for free download from GitHub Releases (current version v0.9.40 as of May 2026), which includes all core features: Chat mode, Agent mode, workspaces, Skills, MCP, remote robots, and memory. The commercial version is available separately at proma.cool/download with built-in channels and subscription plans, designed for out-of-the-box use with reduced API configuration costs. Both versions can run in parallel. The commercial version supports subscription billing and provides pre-configured AI channel access.
Proma pros
- Local-first architecture with JSON/JSONL file storage in ~/.proma/
- Multi-model Chat supporting OpenAI, Anthropic, Google, DeepSeek, Kimi, and more
- Universal Agent mode based on Claude Agent SDK for complex tasks
- Workspace isolation with independent Skills and MCP Server configuration
- SubAgent/Tasks automatically split complex work into child agents
- Remote robot bridge for Feishu/Lark, DingTalk, and WeChat integration
- Shared memory system between Chat and Agent modes
- System-level global voice input with Doubao streaming support
- Side-by-side conversations for comparing multiple model outputs
- Streaming long-task output with plan confirmation and permission modes
- Built-in web search and Chat tools for enhanced responses
- Auto-update, proxy settings, and global hotkeys for desktop convenience
- Markdown, Mermaid, KaTeX, and Shiki code highlighting support
- File attachment parsing with image input capability
- No local database required - uses JSON config and JSONL logs
Proma cons
- Agent mode requires Anthropic or Anthropic-compatible API protocol
- OpenAI and Google channels do not support Agent mode yet
- Requires local Node.js/Bun and Git environment for Agent mode
- Commercial version with subscription coming soon, not yet fully available
- Custom endpoints only support Chat mode, not Agent mode
- Learning curve for setting up Skills and MCP Server configurations
- Memory feature requires separate API credential configuration
- Remote robot bridges may need additional setup for team deployment
- Currently in active development with frequent version updates
Frequently asked questions about Proma
What is Proma and what does it do?
Proma is a local-first AI desktop application that combines multi-model Chat, universal Agent capabilities, workspaces, Skills, MCP, remote robots, and memory into one open-source client. It is designed as an Agent workspace for building long-term AI-powered workflows, where simple queries use Chat mode and complex multi-step tasks are delegated to Agent mode.
What platforms does Proma support?
Proma supports macOS (both Apple Silicon/arm64 and Intel/x64) in DMG/ZIP format, and Windows (x64) in NSIS installer format. The macOS version is signed and notarized by Apple.
What is the difference between Chat mode and Agent mode?
Chat mode is for daily Q&A, explanations, translation, polishing, light code discussions, reading attachments for summaries, and using web search or memory tools. Agent mode is for modifying/creating/organizing local files, research, writing reports, multi-step tasks, using MCP/Skills/Shell/Git, and tasks requiring permission confirmation, plan mode, background execution, or remote robot follow-up.
Which AI model channels are supported?
Proma supports Anthropic, DeepSeek, Kimi API, Kimi Coding Plan, OpenAI, Google, 智谱 AI, MiniMax, Doubao, Tongyi Qianwen, and custom endpoints. Chat mode supports all channels, but Agent mode currently requires Anthropic or Anthropic-compatible protocols (Anthropic, DeepSeek, Kimi, 智谱 AI, MiniMax, Doubao, Tongyi Qianwen).
How is data stored in Proma?
Proma uses local file storage in ~/.proma/ directory with JSON/JSONL files for conversations, agent sessions, workspaces, attachments, settings, and channels. API Keys are encrypted via Electron safeStorage. No local database is required.
What are Skills and MCP in Proma?
Skills are reusable AI agent behaviors that can be configured per workspace for repeatable capabilities. MCP (Model Context Protocol) Servers provide external context to Agents. Each workspace can independently configure Skills, MCP Servers, and workspace files, allowing reusable capabilities to accumulate over time.
How do remote robots work in Proma?
Proma supports Feishu/Lark, DingTalk, and WeChat bridge integration. You can trigger Agent workflows from phone or group chat. Feishu bots support @mention triggers in group chats and private messages, with automatic thread replies and automatic injection of the last 20 group chat messages as context.
What is the memory feature in Proma?
The memory system allows Chat and Agent modes to share memories, enabling AI to truly understand and remember user preferences and habits. It uses HTTP direct connection to MemOS Cloud API with three providers (Anthropic/OpenAI/Google) supporting Tool Use calls for memory tools (recall_memory/add_memory).
How do I set up Proma for the first time?
First, complete environment check (Agent mode requires Git, Node.js/Bun, and available Shell). Then go to Settings > Channels to add at least one AI provider channel with Base URL, API Key, and model list. For Agent mode, select default Agent channel, model, and workspace in Settings > Agent. Configure memory, web search, and remote robot bridges in their respective setting tabs.
Is Proma open source?
Yes, Proma is fully open source under Apache-2.0 license and available on GitHub. It is built as a Bun workspace monorepo with TypeScript, React 18, Electron 39, and uses Claude Agent SDK. The open-source version includes all core features, while the commercial version adds built-in channels and subscription options.