everyharness

Register a local model. Run it through a harness.

Alpha offline-first CLI + TUI for wrapping local models (sklearn, embeddings, Ollama/GGUF/HF LLMs, and more) behind one plugin interface. Not a replacement for Ollama, Gradio, or BentoML — a thin registry + harness layer.

macOS and Linux only in v1. Windows is not supported.

What works today

Pickles need --trust-pickle (loads arbitrary code). Prefer joblib from trusted sources.

Features

• Model registry — local files, Hugging Face, Ollama, or Python callables
• Harness selection — by kind / URI; override with --type
• One CLI — everyharness add → run → serve → train (train depth varies by harness)
• Offline-first — EVERYHARNESS_OFFLINE=1 hard-blocks outbound calls
• Plugin system — publish everyharness-* packages; scaffold with everyharness plugin init
• Agent prompt pack — everyharness ui writes prompts for Cursor, Claude Code, Copilot, Pi, or Codex
• Textual TUI — launch with bare everyharness

Install

pip install everyharness

# optional extras
pip install 'everyharness[tabular]'
pip install 'everyharness[llm]'          # HF + OpenAI-compatible serve deps
pip install 'everyharness[llm,llm-gguf]' # + llama-cpp-python for local GGUF
pip install 'everyharness[vision]'
pip install 'everyharness[all]'

Quick start

# register models (kind auto-detected when possible)
everyharness add ./model.pkl --trust-pickle
everyharness add embeddings:demo --type embeddings
everyharness add python:callable_demo:echo

everyharness list

# same interface, different harnesses
everyharness run --trust-pickle <id> predict --input '[[1.5, 0.5]]'
everyharness run <id> similarity --input '{"a":"cat","b":"kitten"}'
everyharness run <id> call --input '{"hello":"world"}'

# scaffold a publishable harness plugin
everyharness plugin init weather --kind harness

Launch the TUI:

everyharness

Model types

Coding-agent prompt pack

Write metadata + AGENT_PROMPT.md so your coding agent can scaffold a local web/GUI wrapper (everyharness does not generate the UI itself):

everyharness ui <model-id> --agent cursor
# → ./harness-ui/<id>/{AGENT_PROMPT.md,model-card.json,manifest.json,README.md}

Plugins

Built-in harnesses/loaders ship in the core package. Third-party packages can publish as everyharness-* on PyPI (none curated yet beyond the docs sample):

everyharness plugin search tabular
everyharness plugin install everyharness-foo
everyharness plugin init mymodel --kind harness   # harness | loader | detector | templates

• Plugin catalog
• Authoring guide
• Plugin API

Architecture

CLI/TUI → plugin host → harnesses / loaders / detectors / templates. See docs/architecture.md.

Updates & offline

everyharness update --check
everyharness update --yes

Set EVERYHARNESS_OFFLINE=1 to block update checks, HF downloads, and everyharness plugin install.

Development

uv sync --all-extras --dev
uv run pytest -q
uv run ruff check src tests

License

Apache-2.0 — see LICENSE.

Community

• Contributing
• Code of Conduct
• Security
• Changelog