Cybara FAQ
Answers about Cybara: what it is, the platforms it runs on, supported model providers and messaging channels, skills, plugins, MCP, ACP, LSP, hooks, multi-agent orchestration, how it handles your API keys and data, pricing, and operator controls.
What is Cybara?
Cybara is a self-hosted AI agent platform. It pairs a Bun-based agent runtime with a web UI, desktop and mobile apps, a CLI, a broad tool layer, and messaging-channel adapters — extended by skills, plugins, MCP, ACP, and LSP — plus encrypted wallet controls, so agents can code, automate browsers, run messaging workflows, and execute on-chain operations under operator control.
Is Cybara free and open source?
Yes. Cybara is free and open source under the MIT license, created by Carsen Klock, and runs entirely on infrastructure you control. Use it commercially, fork it, and self-host it with no required account, telemetry, or cloud service.
Which platforms does Cybara run on?
Cybara ships desktop apps for macOS, Windows, and Linux, a native SwiftUI macOS app, mobile apps for iOS and Android, and command-line binaries for macOS, Windows, and Linux — plus a VS Code extension and an ACP server for editors — all built from the same Bun runtime and published on GitHub Releases.
Which messaging channels are supported?
Cybara includes adapters for 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, SMS, Email, Web, and Webhook — each gated by pairing, allowlists, and per-channel access policy.
How do I migrate from another agent to Cybara?
Run cybara migrate --from openclaw or cybara migrate --from hermes. Cybara detects local installs, previews the import as a dry run, brings over skills and memory, resolves conflicts by skip, overwrite, or rename, and imports API keys only if you opt in.
Do the apps update automatically?
Yes. Desktop apps use a signed updater channel backed by GitHub Releases and install updates in place, and the CLI verifies a published SHA256 checksum on every cybara update.
How does Cybara keep operators in control?
Sensitive tool calls can require interactive approval with per-session or persistent allowlists, the filesystem supports checkpoint and rollback, wallet operations enforce policy caps and allowlists, and the gateway uses a localhost auth policy with rotatable API keys.
Which model providers can I use?
Cybara connects to 50+ providers — OpenAI, Anthropic, Google Gemini, xAI, Meta, DeepSeek, Qwen, Moonshot/Kimi, Z.ai/GLM, MiniMax, Groq, OpenRouter, Azure OpenAI, AWS Bedrock, Vertex AI, and local runtimes like Ollama, vLLM, LM Studio, and llama.cpp. You bring your own keys; Cybara pools multiple keys per provider, rotates on rate limits, routes by weight or cost, and enforces spend caps.
How does Cybara handle my API keys and data?
Everything stays on infrastructure you control. Model keys, wallet keys, memory, and conversations live locally, the wallet is encrypted at rest, and there's no required account, telemetry, or cloud relay. Keys are never printed to logs and never leave your machine unless a tool you approve sends them.
Can I drive Cybara from my code editor?
Yes. Run cybara acp to expose the agent over the Agent Client Protocol (ACP) and connect Zed or any ACP-compatible editor, so you can code with your own self-hosted runtime. The ACP server can be toggled on or off in settings.
What are skills and where do they come from?
Skills are reusable SKILL.md procedures agents follow for recurring work. Browse and install them from the ClawHub, Skills.sh, and GitHub registries, author your own, or let agents codify a verified procedure automatically — the loader picks it up in every future session.
Does Cybara support MCP servers?
In both directions. Install Model Context Protocol servers as agent tools from the official MCP registry, Smithery, or npm — or expose Cybara's own tools to Claude Desktop and other MCP clients. MCP tools run through the same approval and sandbox policy as built-in tools.
How do plugins work?
Plugins are the unified hub under Settings → Plugins. A plugin bundles skills, MCP services, and connected account apps in one manifest you can enable or disable live without restarting the gateway. Marketplace plugins ship curated skill bundles, and account connectors add OAuth-connected Google Workspace, Microsoft 365, Dropbox, and Notion — read-only by default, with any write still subject to your tool-approval policy.
What are slash commands?
Slash commands are prompt shortcuts that expand into structured templates. Built-ins like /learn, /plan, /review, /test, and /summarize kick off common workflows — /learn even sources material and writes a new SKILL.md via skill_save. Plugins can contribute their own commands, and you type them inline in any chat.
Can I hook into agent behavior with my own code?
Yes. In-process hooks let code react to events like llm_request, tool_before, and message:received, and transform hooks can rewrite tool results, LLM output, or terminal output before they reach the model. You can also register executable shell scripts that receive the event as JSON on stdin and return a decision (for example, to block a call) on stdout — with event filtering and timeouts. A tool_before hook can deny a call outright, so custom guardrails fail closed.
Does Cybara understand my codebase with LSP?
Yes. A built-in Language Server Protocol client gives agents diagnostics, go-to-definition, references, hover, and symbol lookup directly in your workspace. TypeScript and JavaScript are bundled by default, other languages are managed and installed on demand, and the IDE surfaces diagnostics, git blame, and project-wide search alongside locally indexed embeddings.
How does multi-agent orchestration work?
Cybara can fan a single turn out to several proposer agents and synthesize one answer with Mixture of Agents, or route work across providers with weighted, round-robin, lowest-cost, and priority strategies. Beyond routing, agents can spawn scoped subagents and coordinate longer runs on a durable multi-agent board — each step still flowing through your approval and sandbox policy.
Can agents use a real web browser?
Yes. Each session gets an embedded browser the agent drives — opening pages, clicking, scrolling, extracting data, and taking screenshots — and you watch it work live in the chat panel. It works cross-platform, including on Windows. Beyond the browser, a computer-use tool can drive the desktop in the background without taking over your cursor.
Can I talk to my agents with voice?
Yes. Cybara supports hands-free voice conversations with local Kokoro TTS, OS speech, or cloud voices, plus speech-to-text transcription — including a realtime full-duplex mode where you and the agent can speak naturally.
Can agents run on a schedule?
Yes. A built-in cron scheduler runs agents and tools on a timetable you set — recurring reports, monitors, and maintenance jobs — with results delivered to the UI or any connected messaging channel.
What does Cybara cost?
Cybara itself is free and MIT-licensed. You only pay for the model provider usage on your own accounts — Cybara adds no markup, subscription, or per-seat fee, and there's no hosted tier to buy.