Registry
One entry per model with its capabilities, context size, VRAM and licence.
Why. Decisions about a call need facts about the model.
Boundary. Wrong entries mean wrong refusals.
Model access
Refuse the model call that can't work, and record the ones that do.
One interface for local and hosted AI models, with structured output, GPU scheduling and full call records.
Prevents: Prompts silently cut at the context limit, JSON rescued in different ways at every call site, empty replies from thinking models, and models crashing each other on one GPU.
Why it exists
The failure appears after the happy path.
Calling models reliably from ordinary software, especially on your own hardware.
Scope. A client library and CLI; not an agent framework, gateway or evaluation tool.
Five-minute orientation
See the smallest complete path.
pip install "git+https://github.com/honeworks/hone-models"Real output
What it looks like running.
Architecture
The architecture is a set of promises.
A registry of model facts drives checks before each call; providers for Ollama, OpenAI-compatible servers and LiteLLM share one result shape; a file-locked ledger grants GPU leases; spans go to local SQLite.
- A request that doesn't fit the model is refused before it's sent. Inspect test
- A reply cut at the output limit is always reported. Design note
- structured_path says how each structured answer was obtained. Inspect test
- Every call leaves a span that can be replayed. Inspect test
Core concepts
The few concepts you need before reading the code.
Operations
What happens after install.
`hone-models models` and `hone-models calls list / show / stats`.
Current boundary
What this project does not solve.
Alpha. No streaming or async clients. Checks can't make an answer correct.
Near-term roadmap. Follow the design history in design/changes/.
What it implements
- capability registry and model catalog
- pre-flight context and capability checks
- structured output with a reported path
- decision questions
- GPU leases across processes
- embeddings, speech, transcription and generation models
- call records and replay
Engineering checklist
What the repository ships.
- Quickstart
- Present
- Success tests
- Present
- Failure tests
- Present
- Architecture notes
- Present
- Security notes
- Not published
- Runbook
- Not published
- Limitations
- Present
- Changelog
- Present
Need the control, not just the component?
Fit it to the system that has to survive.
The repository exposes the mechanism. Production work is defining the permissions, data, failure costs, evidence, and owners around it.