Use Caro from your agent

Caro-as-a-tool — outward integrations

Working today

• Claude Code (skill) ✅ Working Claude Code skill validated 2026-04-26

  Bundled skill at `.claude/skills/caro-shell/SKILL.md`. Triggers when the user asks for shell-command synthesis; shells out to the published `caro` binary with `--dry-run` and presents the safety-validated suggestion for approval.

  # Install the caro CLI first
  cargo install caro
  
  # Then point Claude Code at the skill bundled in this repo:
  #   .claude/skills/caro-shell/SKILL.md
  # Claude Code auto-discovers skills from .claude/skills/ in any project.

  https://www.anthropic.com/claude-code

In progress

• Claude Code (MCP server) 🚧 In progress MCP server

  Planned `caro mcp serve` subcommand exposing `generate_command`, `validate_command`, `explain_safety`, `show_decision_tree` over the Model Context Protocol. Spec drafted in `.github/first-time-issues/06-mcp-claude-code-integration.md`.

  Tracking: see GitHub issues labeled `integration` + `mcp`

• OpenAI-compat HTTP shim 🚧 In progress OpenAI-compat HTTP

  Planned `caro serve --openai` mode. Exposes an OpenAI Chat Completions endpoint backed by caro's safety-validated generation. Single integration unlocks Codex, Cursor, Continue, Aider, Tabby, and most long-tail tools that already speak OpenAI's schema.

  Tracking: see GitHub issues labeled `integration` + `openai-compat`

• crush 🚧 In progress MCP server

  Charm.sh terminal coding agent. Codex MCP server config pattern in flight.

  Tracking: PR #789 (Crush MCP config)

• Gemini CLI 🚧 In progress Native backend

  Google's agentic CLI. Caro-side Gemini backend in flight.

  Tracking: PR #782 (Gemini integration workflows)

Planned

• OpenAI Codex ⏳ Planned OpenAI-compat HTTP

  Will work via the OpenAI-compat shim once that ships — point Codex at `http://localhost:PORT/v1` instead of `api.openai.com`.

  https://openai.com/codex

• Cursor ⏳ Planned OpenAI-compat HTTP

  Cursor accepts a custom OpenAI base URL in settings. Once the OpenAI-compat shim ships, set base URL to caro and enjoy validated shell-command tool calls.

  https://cursor.com

• Continue.dev ⏳ Planned OpenAI-compat HTTP

  Continue (VS Code / JetBrains) supports OpenAI-compatible providers via `config.json`. OpenAI-compat shim covers it.

  https://continue.dev

• Aider ⏳ Planned OpenAI-compat HTTP

  Aider speaks OpenAI; OpenAI-compat shim covers it.

  https://aider.chat

• Tabby (self-hosted) ⏳ Planned OpenAI-compat HTTP

  Self-hosted coding assistant. Point its model endpoint at the caro shim.

  https://tabby.tabbyml.com

• opencode ⏳ Planned MCP server

  Charm.sh terminal coding agent. Will integrate via MCP server or OpenAI shim.

  https://opencode.ai

• droid ⏳ Planned MCP server

  Mobile-first agent. Plan: MCP server or OpenAI shim.

• Sourcegraph Amp ⏳ Planned MCP server

  Enterprise coding agent. MCP-based integration once `caro mcp serve` ships.

  https://sourcegraph.com/amp

• Letta (MemGPT) ⏳ Planned MCP server

  Persistent-memory agent runtime. Tool registration via MCP.

  https://letta.com

• Qwen Code ⏳ Planned OpenAI-compat HTTP

  Qwen-based coding agent; OpenAI-compat shim or native backend.

• Jules (Google) ⏳ Planned Native backend

  Google's async coding agent. Coordinate with the Gemini backend work.

Backends Caro can drive

Caro itself is a multi-backend client. Use any of these for inference; safety validation runs the same regardless of backend.

• Anthropic Claude API ✅ Working

  Implemented in `src/backends/remote/claude.rs`. Set `ANTHROPIC_API_KEY` and run with `--backend claude`.

  export ANTHROPIC_API_KEY=sk-ant-...
  caro --backend claude "find large files modified today"

• Ollama ✅ Working

  Local or remote Ollama server. Implemented in `src/backends/remote/ollama.rs`.

  ollama pull qwen2.5-coder:7b
  caro --backend ollama --model qwen2.5-coder:7b "tar this dir excluding .git"

• vLLM ✅ Working

  Self-hosted vLLM server (OpenAI-compatible). `src/backends/remote/vllm.rs`.

• Exo ✅ Working

  Distributed local inference across your devices. `src/backends/remote/exo.rs`.

  https://github.com/exo-explore/exo

• MLX (embedded, Apple Silicon) ✅ Working

  On-device inference via MLX + llama.cpp. Default on macOS arm64. No network.

  caro --backend embedded "kill the process on port 3000"

• Candle CPU (embedded) ✅ Working

  Cross-platform CPU inference via Candle. Slower but works everywhere.

• OpenRouter (incl. `auto`) ⏳ Planned

  Single backend → 100+ models including OpenRouter's `auto` routing. Planned: clones the vLLM backend shape.

• Claude Code session token ⏳ Planned

  Reuse the user's active Claude Code authentication (subscription or API). Default to Haiku for cheap fast inference.