swival 0.1.3

A small, powerful CLI coding agent for open AI models

Swival

A coding agent for open models.

Swival connects to LM Studio or HuggingFace Inference API, sends your task, and runs an autonomous tool loop until it produces an answer. With LM Studio it auto-discovers your loaded model, so there's nothing to configure. A few thousand lines of Python, no framework.

Quickstart

LM Studio

1. Install LM Studio and load a model with tool-calling support. Recommended first model: qwen3-coder-next (great quality/speed tradeoff on local hardware). Crank the context size as high as your hardware allows.
2. Start the LM Studio server.
3. Install Swival:

uv tool install swival

4. Run:

swival "Refactor the error handling in src/api.py"

That's it. Swival finds the model, connects, and goes to work.

HuggingFace

export HF_TOKEN=hf_...
uv tool install swival
swival "Refactor the error handling in src/api.py" \
    --provider huggingface --model meta-llama/Llama-3.3-70B-Instruct

You can also point it at a dedicated endpoint with --base-url and --api-key.

Interactive sessions

swival --repl

The REPL carries conversation history across questions, which makes it good for exploratory work and longer tasks.

Updates and uninstall

uv tool upgrade swival    # update
uv tool uninstall swival  # remove

What makes it different

Your models, your way. Swival works with LM Studio and HuggingFace Inference API. With LM Studio, it auto-discovers whatever model you have loaded. With HuggingFace, point it at any supported model or your own dedicated endpoint. You pick the model and the infrastructure.

Small enough to read and hack. The whole agent is a few thousand lines of Python across a handful of files, with no framework underneath. You can read the entire thing in an afternoon. If something doesn't work the way you want, you can change it.

Structured thinking for any model. The built-in think tool gives any model (including local ones) multi-step reasoning with revisions, branches, and persistent notes that survive context compaction. It's not locked to a specific provider.

Built for benchmarking. Pass --report report.json and Swival writes a machine-readable evaluation report with per-call LLM timing, tool success/failure counts, context compaction events, and guardrail interventions. Good for comparing models systematically on real coding tasks.

CLI-native. stdout is exclusively the final answer. All diagnostics go to stderr. You can pipe Swival's output straight into another command or a file.

Documentation

• Getting Started -- installation, first run, what happens under the hood
• Usage -- one-shot mode, REPL mode, CLI flags, piping, exit codes
• Tools -- what the agent can do: file ops, search, editing, web fetching, thinking, command execution
• Safety and Sandboxing -- path resolution, symlink protection, command whitelisting, YOLO mode
• Skills -- creating and using SKILL.md-based agent skills
• Customization -- project instructions, system prompt overrides, tuning parameters
• Providers -- LM Studio and HuggingFace configuration
• Reports -- JSON reports for benchmarking and evaluation
• Using Swival with AgentFS -- copy-on-write filesystem sandboxing for safe agent runs