Cybara features
Explore Cybara's features: multi-agent orchestration, real tool execution, browser and computer automation, voice conversations, self-improving skills, persistent memory, MCP support, and operator controls — self-hosted and open source.
- Multi-agent orchestration. Fan a single turn out to several proposer agents and synthesize one answer with Mixture of Agents, or route by weighted, round-robin, lowest-cost, and priority strategies.
- Real tool execution. A large built-in tool library for code, files, browser and API automation, media generation, and a tool-calling code sandbox — all with interactive approval.
- Self-improving skills. Agents codify a verified multi-step procedure once, and the loader picks it up for every future session. Memory and skills compound over time.
- Encrypted wallet control. On-chain execution across ETH, BTC, and SOL behind an encrypted local wallet with policy caps, allowlists, and private-address SSRF protection.
- Broad provider catalog. Dynamic model discovery across major providers with multi-key credential pools, rate-limit rotation, spend caps, circuit breakers, and coding-plan awareness.
- Operator in the loop. Per-session and persistent tool allowlists, filesystem checkpoint and rollback, transform hooks, and localhost auth policy keep control where it belongs.
- Code-aware IDE. A built-in IDE with semantic workspace indexing — embeddings run locally via Transformers.js — plus LSP diagnostics, git blame, and project-wide search and replace.
- Usage-aware routing. Coding-plan windows are detected automatically from your subscriptions. The router skips exhausted plans, downweights near-limit providers, and honors spend caps.
- MCP in both directions. Install MCP servers as agent tools from the official registry, Smithery, or npm — or flip it around and expose Cybara's own tools to Claude Desktop and other agents over MCP.
- Editor integration over ACP. Drive Cybara agents straight from your editor over the Agent Client Protocol. Run `cybara acp` and connect Zed or any ACP client to code with your own self-hosted runtime.
- Live browser preview. A session-bound embedded browser the agent drives — open pages, click, scroll, extract data, and screenshot — while you watch it work live in the chat panel.
- 25+ messaging channels. Run agents on Telegram, Discord, Slack, WhatsApp, Signal, iMessage, Matrix, Teams, and 20+ more — each gated by pairing, allowlists, and per-channel access policy.
- Persistent memory. Agents recall prior work, decisions, people, and preferences from a searchable memory store, with semantic recall that compounds across every session.
- Skill registry & install. Browse and install reusable SKILL.md procedures from ClawHub, Skills.sh, and GitHub, or let agents author their own — self-improvement that grows your toolkit.
- Voice, in and out. Talk to your agents hands-free with local Kokoro TTS, OS and cloud voices, speech-to-text transcription, and a realtime full-duplex conversation mode.
- Computer use. Agents drive the desktop in the background — capture, click, type, scroll, and drag — without taking over your cursor, in a sandboxed flow you approve.
- Scheduled automations. A built-in cron scheduler runs agents and tools on your timetable — recurring reports, monitors, and maintenance jobs that fire even while you're away.
- Plugins & account connectors. One hub for skill bundles, MCP services, and encrypted account connectors — Google Workspace, Microsoft 365, Dropbox, and Notion — read-only by default with approval-gated writes.
- Media generation. Generate images, video, music, and speech from chat through swappable media providers, with artifacts and a canvas to collect what agents produce.
- Sandboxed execution. Shell and git run inside a configurable command sandbox — sandbox-exec on macOS, Podman or Docker on Linux — with network allow/deny and path-safety guards for secrets.