An AI coding agent that runs over the Agent Client Protocol (ACP). It orchestrates LLM calls, manages chat sessions, discovers project-level skills, and executes a registry of built-in and MCP-backed tools.

clawcode ships two binaries:

• acp — ACP stdio agent for any ACP-compatible client.
• claw-tui — Interactive terminal client with session resume and streaming responses.

┌──────────┐  ACP (stdio)  ┌───────┐
│  ACP     │◄────────────►│  acp  │
│  Client  │               └───┬───┘
└──────────┘                   │
                    ┌──────────▼──────────┐
                    │       kernel        │
                    │  session / turn     │
                    │  orchestration      │
                    └──┬───────┬───────┬──┘
                       │       │       │
              ┌────────▼┐ ┌────▼──┐ ┌──▼──────┐
              │ provider │ │ tools │ │ skills  │
              │ (LLMs)   │ │       │ │         │
              └──────────┘ └───┬───┘ └─────────┘
                          ┌────▼─────┐
                          │   mcp    │
                          │ servers  │
                          └──────────┘

• Rust stable (see rust-toolchain.toml)
• An LLM API key (OpenAI, DeepSeek, or any OpenAI-compatible provider)

Create ~/.config/clawcode/config.toml. For repo-local experiments, create ./claw.toml as the fallback config:

active_model = "deepseek/deepseek-v4-pro"
approval = "yolo"  # or "request_approval"

[tui]
theme = "dark"  # or "light"

[[providers]]
id = "deepseek"
display_name = "DeepSeek"
provider_type = "openai-completions"
base_url = "https://api.deepseek.com"

api_key = { env = "DEEPSEEK_API_KEY" }

[[providers.models]]
id = "deepseek-v4-pro"
display_name = "DeepSeek V4 Pro"
context_tokens = 1000000
max_output_tokens = 384000

# Optional: connect MCP servers
[[mcp_servers]]
enabled = false
name = "filesystem"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "."]

Run the TUI with:

# Interactive terminal UI
cargo run -p tui
# Or run by binary name
cargo run --bin claw-tui
cargo run --bin acp

The TUI also supports listing persisted sessions and resuming one:

cargo run -p tui -- --list-sessions
cargo run -p tui -- --resume <SESSION_ID>

In the TUI, use /agent to open the agent picker. It lists the main session and live sub-agents, then switches the active transcript to the selected agent session.

Use /model from the main session to open the model picker, then use the arrow keys and Enter to switch the active model. /model <provider/model> also switches directly.

Models can use the built-in agent tools to spawn and coordinate sub-agents. Sub-agents run as separate sessions with their own transcript streams, and the TUI can switch between the main session and sub-agent sessions with /agent.

Skills live in .agents/skills/<name>/SKILL.md (project) or $HOME/.agents/skills/<name>/SKILL.md (user). Each skill has YAML frontmatter with a name and description, and a markdown body that is injected into the system prompt when the user mentions $skill-name.

Project skills take priority over user skills with the same name.

Local agent state directories are intentionally ignored by Git:

• .agents/
• .claude/
• .codex/

These directories may contain local skills, tool caches, transcripts, or agent-specific runtime files. Keep durable project documentation in docs/ instead.

Licensed under either of MIT or Apache 2.0, at your option.

See LICENSES and THIRD_PARTY_NOTICES.md.