smasher

AI workflow orchestration in Rust. Write your pipeline as a DOT directed graph, point it at an LLM, and run it.

Based on attractor by strongDM, reimplemented from scratch in Rust. The original attractor defined the idea of DOT-graph-driven AI pipelines; smasher takes that concept and rebuilds it with a layered crate architecture, multi-provider LLM support, and a web dashboard.

Five crates, bottom to top:

Crate	What it does
smasher-llm	Talks to OpenAI, Anthropic, and Gemini through one client. Handles streaming, retries, the usual.
smasher-agent	Agent loop with tools (read, write, edit, shell, grep, glob). Steering rules, subagents, sandboxed execution.
smasher-attractor	The graph engine. Parses DOT, resolves node types from shapes, dispatches handlers, runs the pipeline.
smasher-cli	smasher complete, smasher chat, smasher run, smasher serve, smasher resume, smasher render, smasher ingest, smasher archive, smasher lint.
smasher-web	HTMX dashboard on port 21541. Live event stream, graph visualization, human Q&A.

Setup

You need Rust 1.85+ and at least one API key.

git clone https://github.com/2389-ai/smasher.git
cd smasher
cargo build --release

Set a provider key (any one works, or set all three):

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...
export GEMINI_API_KEY=...

Or drop them in a .env file at the repo root.

Usage

One-shot completion

smasher complete "explain quicksort in three sentences"
smasher complete "explain quicksort" --json  # full response object

Interactive chat

smasher chat

The chat REPL gives the agent all six tools (read, write, edit, shell, grep, glob), so it can work on files in the current directory.

Run a pipeline

smasher run examples/old-examples/hello-world.dot
smasher run examples/old-examples/conditional.dot --var route=yes
smasher run examples/old-examples/multi-step.dot --var model=claude-sonnet-4-20250514

Pipelines are standard DOT digraphs. Node shapes tell the engine what each node does:

• circle / point = start
• doublecircle = exit
• box / rectangle = codergen (runs an LLM agent)
• diamond = conditional (branches on variables)
• oval / ellipse = interviewer (asks a human, captures response)
• house = manager (human approval gate)
• parallelogram = parallel fan-out
• hexagon = tool
• component = sub-pipeline (nested DOT file)

See examples/ for working pipelines and docs/dot-reference.md for the full spec.

Web dashboard

smasher serve
# opens on http://127.0.0.1:21541

Submit pipelines from the browser, watch events arrive over SSE, answer human gate questions through the web UI.

Development

cargo check --workspace
cargo test --workspace       # ~2,700 tests
cargo clippy --workspace

There's a Makefile with shortcuts if you prefer that.

Project structure

crates/
  smasher-llm/          # LLM client, provider catalog, types
  smasher-agent/        # Agent loop, tools, steering, events
  smasher-attractor/    # DOT parser, graph engine, handlers
  smasher-cli/          # CLI binary
  smasher-web/          # Web dashboard (axum + askama + HTMX)
docs/                   # Reference docs, quickstart guide
examples/               # Sample DOT pipelines
scripts/                # CI helpers

Docs

• Quickstart
• API reference
• DOT reference
• Handler reference
• CLI reference
• Config reference

License

See LICENSE for details.