Agent Runtime

doltclaw

Minimal agent runtime — LLM inference, model fallback chains, and tool calling in 1,652 lines of Rust. Part of the claw family.

1,652 LOC · 22 tests · 2.9MB binary · 10 deps · Zero bloat

View on GitHub Documentation

Why doltclaw

2.9MB Binary

vs 500MB+ for openclaw. Single static binary, no Node.js, no runtime dependencies. Compiles to a release binary smaller than most favicons.

Model Fallback Chains

Primary model fails with 429 or 5xx? Automatic failover to the next backend in the chain. Non-retryable errors (401, 403) abort immediately — no wasted latency.

Tool Calling Loop

Register custom tools via the Tool trait. The agent loop handles generate → tool_calls → execute → re-generate automatically, up to max_iterations.

TOML Config + ${ENV_VAR}

Environment variable substitution at load time. No JSON, no YAML. Clean TOML that's version-controllable and diffable.

Zero Built-in Tools

doltclaw is a runtime, not a toolkit. Consumers (doltdot, ares, pawan) register their own tools. The crate stays minimal and focused.

10 Dependencies

tokio, reqwest, serde, toml, thiserror, tracing, async-trait, futures, uuid, clap. No framework bloat. Every dep earns its place.

Use as a Library

Add doltclaw to your Cargo.toml and get inference + fallback + tool calling in 10 lines.

// Cargo.toml: doltclaw = "0.1"

use doltclaw::{Agent, Config};

use std::sync::Arc;

#[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(MySearchTool));

// Execute with automatic model fallback

let response = agent.execute("Analyze this codebase").await?;

println!("{}", response.content);

// response.model_used tells you which model succeeded

}

Quick Start

# Install

$ git clone https://github.com/dirmacs/doltclaw && cd doltclaw

$ cargo install --path . --features cli

$ export NVIDIA_API_KEY=nvapi-...

# Validate config and model chain

$ doltclaw check# Shows: StepFun Flash → GLM 4.7 → Qwen

# Run a prompt through the agent loop

$ doltclaw run "Explain Rust ownership"

# Migrate from openclaw

$ doltclaw migrate ~/.openclaw/openclaw.json > doltclaw.toml

Configuration

One TOML file. Three models. Automatic fallback. Environment variable substitution.

# doltclaw.toml

[providers.nvidia-nim]

base_url = "https://integrate.api.nvidia.com/v1"

api_key = "${NVIDIA_API_KEY}"

[[providers.nvidia-nim.models]]

id = "stepfun-ai/step-3.5-flash"

name = "StepFun Flash"

context_window = 128000

[agent]

primary = "nvidia-nim/stepfun-ai/step-3.5-flash"

fallbacks = ["nvidia-nim/z-ai/glm4.7"]

max_iterations = 50

The Claw Family

doltclaw is part of a family of agent runtimes, each with different tradeoffs. We built the smallest one.

Name	Language	Binary	RAM	What
openclaw	TypeScript	~500MB	~394MB	Full-featured personal AI assistant
zeroclaw	Rust	8.8MB	<5MB	Trait-driven, provider-agnostic runtime
openfang	Rust	32MB	~40MB	Agent OS with autonomous Hands
doltclaw	Rust	2.9MB	<5MB	Minimal runtime for dirmacs
nanoclaw	Python	N/A	>100MB	Container-isolated, small codebase
zeptoclaw	Zig	43MB	N/A	NVIDIA NIM native, 21 skills

Architecture

8 source files. Single crate with optional CLI behind a feature flag. Backend ported from pawan-core.

src/

├── lib.rs — Error, Result, re-exports

├── types.rs — Message, Role, Response, TokenUsage

├── config.rs — TOML loader, ${ENV_VAR} substitution

├── tools.rs — Tool trait + ToolRegistry

├── agent.rs — Tool-calling loop + model fallback

├── backend/

├── mod.rs — Backend trait

└── openai_compat.rs — NVIDIA NIM, OpenAI, DeepSeek

└── main.rs — CLI (feature = "cli")

Ecosystem Integration

doltclaw is the inference layer for the dirmacs stack. It plugs into ares, pawan, and doltares.

# In your Cargo.toml

doltclaw = { git = "https://github.com/dirmacs/doltclaw" }

# doltares uses doltclaw as its inference backend

doltclaw = { path = "/opt/doltclaw" }

# pawan shares the same OpenAI-compat backend pattern

# doltclaw was ported FROM pawan-core's proven backend