doltclaw — Minimal agent runtime

doltclaw is a minimal Rust agent runtime — LLM inference, model fallback chains, and tool calling in 1,652 lines. Part of the claw family (openclaw, zeroclaw, openfang, nullclaw, picoclaw, nanoclaw). Built for the dirmacs ecosystem, works with any OpenAI-compatible API.

Named after the claw that cuts clean — no bloat, no runtime, no Node.js. Just function.

• 1,652 lines — small enough to understand, audit, and modify
• 2.9MB binary — vs openclaw's ~500MB install
• 1.4MB library — zero dirmacs dependencies, importable by anything
• 10 dependencies — reqwest, tokio, serde, toml, thiserror, tracing, async-trait, futures, uuid, clap
• Model fallback — primary fails? automatic failover to next in chain
• TOML config — ${ENV_VAR} substitution, no JSON, no YAML
• Tool calling — register your own tools via Tool trait, agent loop handles the rest

# Clone and build
git clone https://github.com/dirmacs/doltclaw && cd doltclaw
cargo install --path . --features cli

# Set API key (NVIDIA NIM free tier)
export NVIDIA_API_KEY=nvapi-...

# Validate config
doltclaw check

# Run a prompt
doltclaw run "What is 2+2?"

# Migrate from openclaw
doltclaw migrate ~/.openclaw/openclaw.json > doltclaw.toml

use doltclaw::{Agent, Config};

#[tokio::main]
async fn main() -> doltclaw::Result<()> {
    let config = Config::load("doltclaw.toml".as_ref())?;
    let mut agent = Agent::from_config(config)?;

    // Register custom tools
    agent.register_tool(Arc::new(MyTool));

    // Execute with automatic model fallback
    let response = agent.execute("Explain this code").await?;
    println!("{}", response.content);
    println!("Model used: {}", response.model_used);
    Ok(())
}

# doltclaw.toml
[providers.nvidia-nim]
base_url = "https://integrate.api.nvidia.com/v1"
api_key = "${NVIDIA_API_KEY}"
api = "openai-completions"

[[providers.nvidia-nim.models]]
id = "qwen/qwen3.5-122b-a10b"
name = "Qwen 3.5 122B"
reasoning = false
context_window = 131072
max_tokens = 16384

[[providers.nvidia-nim.models]]
id = "stepfun-ai/step-3.5-flash"
name = "StepFun Step 3.5 Flash"
context_window = 128000
max_tokens = 16384

[[providers.nvidia-nim.models]]
id = "z-ai/glm4.7"
name = "GLM 4.7"
reasoning = true
context_window = 128000
max_tokens = 16384

[agent]
primary = "nvidia-nim/qwen/qwen3.5-122b-a10b"
fallbacks = [
  "nvidia-nim/stepfun-ai/step-3.5-flash",
  "nvidia-nim/z-ai/glm4.7",
]
max_iterations = 50

[agent.params]
temperature = 1.0
top_p = 0.95

doltclaw/
  src/
    lib.rs             # Error, Result, re-exports
    types.rs           # Message, Role, Response, ToolCallRequest, TokenUsage
    config.rs          # TOML loader, ${ENV_VAR} substitution, ModelRef parsing
    backend/
      mod.rs           # Backend trait
      openai_compat.rs # OpenAI-compat HTTP client (NVIDIA NIM, OpenAI, DeepSeek)
    tools.rs           # Tool trait + ToolRegistry (zero built-in tools)
    agent.rs           # Tool-calling loop with model fallback chain
    main.rs            # CLI (behind feature = "cli")

8 source files. The Backend trait is ported from pawan-core — proven against NVIDIA NIM in production.

MIT