BenchFlow

Agents

Verified ACP agents the runtime supports.

Each agent runs in a sandboxed environment with hardened verifiers, mounted skills, and host subscription auth where available. Web tools are disabled by default.

  • Claude Code

    ACP

    claude-agent-acp

    Claude Code via Anthropic's Agent Client Protocol.

    Auth
    ANTHROPIC_API_KEY
    Skills
    $HOME/.claude/skills

  • Codex

    ACP

    codex-acp

    OpenAI Codex agent through the Zed ACP package.

    Auth
    OPENAI_API_KEY
    Skills
    $HOME/.agents/skills

  • Gemini CLI

    ACP

    gemini

    Google Gemini CLI running with `gemini --acp --yolo`.

    Auth
    GOOGLE_API_KEY or GEMINI_API_KEY
    Skills
    $HOME/.gemini/skills

  • OpenCode

    ACP

    opencode

    Open-source TypeScript coding agent through `opencode acp`.

    Auth
    Inferred from model/provider
    Skills
    $HOME/.opencode/skills

  • OpenHands

    ACP

    openhands

    OpenHands agent via ACP with multi-model provider support.

    Auth
    LLM_API_KEY
    Skills
    $HOME/.agents/skills

  • OpenClaw

    ACP shim

    openclaw

    OpenClaw agent wrapped by BenchFlow's ACP shim.

    Auth
    Inferred from model/provider
    Skills
    $HOME/.claude/skills

  • Pi

    ACP

    pi-acp

    Pi coding agent using the Pi ACP launcher.

    Auth
    Inferred from model/provider
    Skills
    $HOME/.pi/agent/skills