Cybara – Open AI Agent Platform Built with Bun

TL;DR

TL;DR: Cybara is an MIT-licensed, self-hosted AI agent platform built with Bun that combines a chat UI, CLI, desktop shells, MCP support, and channel adapters for Telegram, Discord, Slack, and more — all under operator control.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: cybara.ai ← verified via curl
• Source repository: github.com/metaspartan/cybara ← README.md read in full
• License: MIT ← verified from LICENSE.md in repo
• HN launch thread: news.ycombinator.com/item?id=42525309 ← identified via search

What Is Cybara?

Cybara describes itself as “an agent operating system for developers and operators.” It combines:

• Bun-based runtime with a web UI, CLI, Tauri desktop app, and React Native mobile companion
• Multi-agent orchestration with model provider routing (Azure OpenAI, Anthropic on Vertex, Google Gemini, and more)
• MCP host and client mode — expose Cybara tools to other MCP clients, or consume external MCP servers
• Channel adapters for Telegram, Discord, Slack, WhatsApp, Signal, iMessage, Matrix, Microsoft Teams, Feishu/Lark, DingTalk, and 20+ other platforms
• Browser automation and API automation tools built in
• Encrypted wallet controls for ETH, BTC, and SOL with policy controls
• Mixture of Agents (MoA) — fan a turn out to several proposer agents and synthesize one answer

The project was confirmed MIT-licensed in its LICENSE.md, uses TypeScript, React 19, and Tauri for the desktop shell.

Setup Workflow

Install the CLI

On macOS or Linux:

curl -fsSL https://cybara.ai/install.sh | bash

On Windows PowerShell:

powershell -c "irm https://cybara.ai/install.ps1 | iex"

The installer selects the correct release for the current platform and verifies its SHA256 checksum.

Run Without Installing

bunx cybara

Launch the Full GUI

Download a desktop build from cybara.ai/download or the GitHub Releases page for macOS, Windows, or Linux.

Connect an External Editor via ACP

Cybara supports the Agent Client Protocol (ACP) v1 for connecting compatible editors over stdio:

cybara acp

Deeper Analysis

Architecture

Cybara is unusual in how deliberately it targets the “operator in the loop” model. Rather than delegating fully to an autonomous agent, every tool call, filesystem operation, and wallet transaction can be approved or denied interactively. The README frames this as a design principle, not a workaround.

The runtime is Bun (not Node.js), which gives it faster startup and native TypeScript execution. The desktop shell is Tauri, and the mobile companion is React Native. The web UI is React 19.

Model Routing

The provider router supports weighted, round-robin, lowest-cost, priority, and mixture-of-agents strategies. It has circuit breakers, rate limits, and spend caps. Dynamic model discovery is in src/core/providers.ts, which covers Azure OpenAI, Azure AI Foundry, Anthropic on Vertex AI, and Google Gemini on Vertex AI.

MCP Support

Cybara runs in both MCP host mode (exposing its own tools to external MCP clients) and MCP client mode (consuming external MCP servers). This is a first-class feature, not a shim.

Channel Adapters

The adapter list covers an unusually broad range: Telegram, Discord, Slack, WhatsApp, Signal, iMessage, Matrix, Mattermost, Microsoft Teams, Feishu/Lark, DingTalk, WeCom, Zulip, LINE, Google Chat, IRC, ntfy, Twitch, Nextcloud, Synology, Zalo, Home Assistant, Web, Webhook, SMS, and Email. All of these are listed in the README’s capability snapshot.

Practical Evaluation Checklist

• [ ] Install script runs without error on macOS or Linux
• [ ] bunx cybara starts the CLI without installing
• [ ] Web UI loads on localhost after launch
• [ ] MCP host mode exposes tools to an external MCP client
• [ ] A channel adapter (e.g. Telegram) can be configured and receives messages
• [ ] Mixture of Agents routing produces a synthesized response from multiple providers
• [ ] Encrypted wallet controls load and show ETH/BTC/SOL policy options

Security Notes

Cybara emphasizes operator control. Interactive tool approval, per-session and persistent allowlists, filesystem checkpoint and snapshot with rollback, and transform hooks (tool_result, llm_output, terminal_output) are all listed in the README’s capability snapshot. Wallet operations require explicit approval and use encrypted local wallet controls.

When running on a server, gateway operator controls handle localhost auth policy, API key rotation, and log viewing.

FAQ

Q: Does Cybara require an API key for the AI models? A: Yes. Cybara is a runtime and UI — you supply your own API keys for the model providers (Azure OpenAI, Anthropic, Google Gemini, etc.). It does not include a hosted model service.

Q: How does it compare to tools like n8n or LangFlow? A: Cybara is more agent-centric than workflow-automation tools. It focuses on multi-agent orchestration with operator approval at every step, rather than visual workflow编排. It also includes a Bun-based runtime, Tauri desktop shell, and React Native mobile companion that most workflow tools lack.

Q: Is there a hosted or cloud version? A: No. Cybara is fully self-hosted. The README states “run agents that can code, automate browsers, manage messaging workflows, and execute wallet operations while keeping operator control in the loop” — all under operator control, meaning the operator owns the infrastructure.

Q: Does it support local models? A: The README mentions local workspace indexing with Transformers.js embeddings and ONNX Runtime, but the primary model interfaces are cloud providers (Azure, Anthropic, Google). Local model support appears to be partial, focused on embeddings rather than chat completions.

Conclusion

Cybara is a feature-dense, self-contained AI agent platform that makes unconventional choices — Bun over Node.js, Tauri over Electron, React Native mobile, ACP for editor integration — to deliver a unified runtime for agents that stay under operator control. Its channel adapter breadth and MCP host/client duality are standout features for developers who need agents to interact with real-world services and tooling. MIT licensed, source on GitHub.