Mammoth reads a DOT file describing a pipeline graph and executes it node by node, dispatching each stage to an LLM coding agent. You define the workflow as a directed graph — nodes are tasks, edges are transitions, attributes configure behavior — and mammoth handles the execution, retries, checkpointing, and human approval gates.

Install

# From source
go install github.com/2389-research/mammoth/cmd/mammoth@latest

# Or grab a binary from GitHub releases
# macOS (arm64/amd64), Linux (arm64/amd64)

What it does

Pipeline execution from DOT files. Write a .dot file with nodes for each stage of your build — setup, code generation, testing, verification, integration. Mammoth parses the graph, validates it against 21 lint rules, and walks the DAG from start to exit.

Multi-provider LLM agents. Each codergen node spins up an agent session backed by OpenAI, Anthropic, or Gemini. Set an API key and mammoth picks the provider, or pin a specific model per node via attributes.

Checkpoint and auto-resume. Every successful node saves a checkpoint. If a run fails or gets interrupted, mammoth pipeline.dot picks up where it left off. Use --fresh to force a clean start.

Parallel branches with fan-in. Fork work across multiple branches that execute concurrently, then merge results at a fan-in node with configurable join policies (all-success, majority, first-success).

Human gates. Nodes with shape=house pause execution and prompt for human input — approve, reject, or provide guidance. Configurable timeouts with default choices for unattended runs.

Verification nodes. Octagon-shaped nodes run shell commands (like swift test or go vet) and gate progression on exit code. No LLM, no token cost, deterministic pass/fail.

Terminal UI. Run with --tui for a Bubble Tea dashboard showing the live DAG, event log, and node details. Or use the web UI via mammoth serve.

How it works

Three layers, bottom up:

llm/ — Unified client that normalizes OpenAI Responses API, Anthropic Messages API, and Gemini into a single interface with streaming, retry, and backoff.
agent/ — Agentic loop that drives an LLM through tool calls (file read/write, shell exec, patch apply), with steering, loop detection, and subagent spawning.
attractor/ — The pipeline engine. Parses DOT, validates the graph, walks nodes through a 5-phase lifecycle, manages checkpoints, and dispatches to the agent layer.

Pipeline files use standard DOT syntax with node attributes for configuration:

digraph build_app {
    setup [shape=box, prompt="Set up the project structure"]
    implement [shape=box, prompt="Implement the core logic"]
    verify [shape=octagon, command="go test ./..."]
    done [shape=doublecircle]

    setup -> implement -> verify -> done
    verify -> implement [condition="outcome=fail"]
}

Requirements

Go 1.25+ (for building from source)
At least one LLM API key: ANTHROPIC_API_KEY, OPENAI_API_KEY, or GEMINI_API_KEY
Graphviz (optional, for SVG/PNG graph rendering)