Carapace AI Agent

carapace

Stable release available. Carapace is ready for real use on its verified stable paths; partial and in-progress areas are called out explicitly in the docs.

A security-focused, open-source personal AI assistant. Runs on your machine. Works through Matrix, Signal, Telegram, Discord, Slack, webhooks, and console. Supports Anthropic, OpenAI, Codex, Ollama, Gemini, Vertex AI, Bedrock, NEAR AI Cloud, Venice AI, and local Claude CLI. Extensible via WASM plugins and guarded filesystem tools. Written in Rust.

A hardened alternative to openclaw / clawdbot — for when your assistant needs a hard shell.

Features

• Multi-provider LLM engine — Anthropic, OpenAI API key, Codex subscription login, Ollama, Google Gemini, Vertex AI, AWS Bedrock, NEAR AI Cloud, Venice AI, and local Claude CLI with streaming, cancellation, and provider-aware tool paths
• Multi-channel messaging — Matrix, Signal, Telegram, Discord, Slack, console, and webhooks
• Channel activity framework — per-channel typing indicators and append-time read receipts, with Signal as the first activity-enabled built-in channel
• Tooling and local workspace access — built-in agent tools, guarded filesystem tools for explicit roots, and channel-specific tool schemas
• Signed plugin runtime — plugins are signature-verified and run with strict permissions and resource limits
• Secure defaults — local-first binding, locked-down auth behavior, encrypted secret storage, guarded tool execution, root-scoped filesystem access, and OS-level subprocess sandboxing for protected paths
• Infrastructure — TLS, mTLS, mDNS discovery, config hot-reload, Tailscale integration, Prometheus metrics, audit logging

Security

Carapace is designed to address the major vulnerability classes reported in the January 2026 openclaw security disclosures:

Threat	Carapace defense
Unauthenticated access	Denied by default when credentials configured; CSRF-protected control endpoints
Exposed network ports	Localhost-only binding (127.0.0.1)
Plaintext secret storage	OS credential store (Keychain / Keyutils / Credential Manager) with AES-256-GCM fallback
Skills supply chain	Ed25519 signatures + WASM capability sandbox + resource limits
Prompt injection	Prompt guard + inbound classifier + exec approval flow + tool policies
No process sandboxing	OS-level subprocess sandboxing on macOS/Linux/Windows for sandbox-required paths; unsupported paths fail closed
SSRF / DNS rebinding	Private IP blocking + post-resolution validation

See docs/security.md for the full security model. See docs/security-comparison.md for a threat-by-threat comparison with OpenClaw. See docs/feature-status.yaml and docs/feature-evidence.yaml for verified-vs-partial implementation status.

Quick Start

1. Install cara from the latest release (Linux/macOS/Windows):
   • https://getcara.io/install
   • docs/site/install.md
2. Run guided setup:

   cara setup
   

3. Start the assistant:

   cara
   

4. Verify first-run outcome:

   cara verify --outcome auto --port 18789
   

5. Start local interactive chat:

   cara chat
   

Use /help in chat for REPL commands (/new, /exit, /quit).

If you use cloud models, finish one provider onboarding path before launching: set one provider key (for example ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, NEARAI_API_KEY, or VENICE_API_KEY), use Codex sign-in through cara setup --provider codex or the Control UI, or use Gemini Google sign-in through cara setup --provider gemini --auth-mode oauth or the Control UI. Codex and Gemini Google sign-in both require CARAPACE_CONFIG_PASSWORD so the stored auth profile stays encrypted at rest. Models are routed explicitly with provider:model strings such as anthropic:claude-sonnet-4-6, openai:gpt-5.5, gemini:gemini-2.5-flash, nearai:google/gemma-4-31B-it, ollama:llama3.2, or codex:default. If you are not sure where to start, choose local-chat as your first outcome, start with one provider, and add channels only after cara verify --outcome auto passes. If you want Cara to inspect one local project directory, enable the filesystem block for a single workspace root and start with the guarded local project assistant recipe.

Roadmap

Active and planned work is tracked on GitHub Issues. The feature inventory is the source of truth for what currently ships.

Recently shipped highlights are tracked on the GitHub Releases page.

Docs

• Website — install, first run, security, ops, cookbook, troubleshooting
• Getting started — full setup and operations
• Install — release binaries, signatures, and install commands
• First run — secure local startup and smoke checks
• Help — setup help, team evaluation, and cookbook request paths
• Security model — architecture and trust boundaries
• Security comparison — threat-by-threat view
• Channel setup — Matrix, Signal, Telegram, Discord, Slack, webhooks
• Channel smoke validation — live checks and evidence capture
• Cookbook — outcome-first walkthroughs
• Release & upgrade policy — upgrade, migration, rollback, release checklist
• CLI guide — subcommands, flags, and device identity
• Documentation index — architecture/protocol/security references
• Security reporting policy — private vulnerability reporting and response expectations
• Report feedback or bugs

Contributing

If you want to build from source or contribute, start here:

• CONTRIBUTING.md
• docs/README.md

License

Apache-2.0 — see LICENSE.