uncloseai.

uncloseai-cli: Agent Harness for Local LLMs

A Harness That Levels Models Up

uncloseai-cli (command: unclose) is a ReAct agent harness for local and open LLMs. North star: bring your model as it is. A smol, quirky, or brand-new model becomes a real coding agent through a purify-first tool layer — the harness adapts to the model, never the model to the harness. Pure Python. Public domain.

Quick Start

git clone https://git.unturf.com/engineering/unturf/uncloseai-cli.git
cd uncloseai-cli
make install-unclose

# Point it at any OpenAI-compatible endpoint — an endpoint
# alone is a complete config (key optional, model id
# auto-resolved from /v1/models, label from hostname):
mkdir -p ~/.uncloseai-cli
cat > ~/.uncloseai-cli/development.env <<'EOF'
MODEL_ENDPOINT_0="https://qwen.ai.unturf.com/v1"
# MODEL_API_KEY_0="sk-friends-CHANGEME"
MODEL_ENDPOINT_1="https://hermes.ai.unturf.com/v1"
# MODEL_API_KEY_1="sk-friends-CHANGEME"
EOF
chmod 600 ~/.uncloseai-cli/development.env

unclose "what changed in the last commit?"
unclose -p "summarize README.md"        # print-only, pipe-friendly
unclose -i                              # interactive REPL; /model switches live

The free uncloseai inference endpoints above need no key. Any vLLM / Ollama / cloud OpenAI-compatible endpoint works the same way.

What You Get

Purify, don't demand perfection: malformed tool calls from quirky models are canonicalized on the way in, so the model succeeds first try instead of burning recovery turns.
Twenty in-tree tools: files, bash, web crawler, vision, RAG, sub-agent delegation — plus opt-in computer use (screenshot, clipboard, accessibility tree, input injection) on linux, WSL, macOS, FreeBSD, and OpenBSD.
Self-hardening: every run's conversation and every malformed→canonical tool arg is harvested into a regression corpus. A quirk handled once stays handled for every model.
Terminal-native: markdown rendering with tables, clickable links, and LaTeX→Unicode; resumable sessions; background jobs; ~50 env-var knobs, each A/B-testable alone.

Documentation

The full reference — CLI flags, provider catalog, all twenty tools, REPL commands, architecture, and the complete tunables table — lives in the repo README:

Source + README — git.unturf.com/engineering/unturf/uncloseai-cli
HARNESS.md — design: purify layer, reflector, self-play loop
COMPUTER-USE.md — desktop-control support matrix per platform
YAML-TOOL-CALLING.md — canonical tool-call format
PATTERN-MATCHERS.md — anti-footgun detectors

Looking for microgpt or the smol model garden? They live in the sibling training repo since the 2026-07 split: training.ai.unturf.com.

License

Public domain. Knowledge unbound by gatekeepers.