UNCUT 0.1.1 — a headless coding agent: 7 MB, four tools, no context truncation

⚡ The gist in 5 seconds

• UNCUT 0.1.1 is a headless coding agent: a single 7 MB static binary and just four tools.
• Linux, macOS, and Windows on amd64 and arm64; install via curl or go install with Go 1.26+.
• No TUI, no telemetry, no plugins: the model and key are set via UNCUT_MODEL and UNCUT_API_KEY.

🔍 What happened

UNCUT has landed on GitHub — a "lossless, headless coding agent harness" by an author going by the nickname Balanov. Version 0.1.1, MIT license, written in Go. A single statically linked binary weighing 7 MB, with no third-party Go dependencies. Platforms: Linux, macOS, and Windows, two architectures each — amd64 and arm64.

Inside are 837 lines of Go, a single provider protocol (OpenAI-compatible Chat Completions, including local servers), and a single agent loop with four tools. You launch it like uncut "Inspect this repository and fix the failing tests". The final text goes to stdout, structured events go to NDJSON via --json. Sessions are optionally written to append-only JSONL so the conversation can be resumed from the same point.

The main trick is in the name. The author promises that UNCUT never truncates context: no summarization, no pruning, no sliding windows. The entire conversation is sent to the model on every call, and file contents and concatenated command output are not cut off. If the provider hits the output limit and returns a truncated response, the harness fails with an error rather than passing off a partial answer as success.

🧠 Our take

Our assessment: this is not a product, it's a thesis. Three commits in the repo and 837 lines of code — a bid for minimalism against heavyweight IDE plugins and cloud wrappers with telemetry. The bet on CI, containers, and shell pipelines is obvious at a glance: stdin/stdout, NDJSON, a static binary with no runtime.

Lossless is a deliberate engineering choice, and it comes at a price. Context is never truncated, so the agent's behavior is reproducible and predictable — a plus for CI. But the entire conversation goes to the model at every step, and the token bill grows linearly with session length. For long interactive tasks that's expensive, and the author seems to have consciously chosen determinism over cost savings.

Who wins: those who assemble their own pipelines and want to fork the harness to fit their needs — MIT and a single agent loop invite exactly that. Who loses: proprietary harnesses that silently truncate output and call it optimization. But four tools is very few, and I wouldn't expect UNCUT to handle a large repository without some serious tweaking.

⚠️ What we don't know

The available README excerpt doesn't name the four tools themselves — only their count. The binaries are unsigned, so SmartScreen and Gatekeeper will complain, and the author himself advises verifying the SHA-256 from checksums.txt before running. The BENCHMARKS.md file exists in the repo, but there are no figures in the post: any claims about speed and quality remain unverified for now. One author, three commits, version 0.1.1 — the reader should run bench themselves rather than trust the description.

Prediction: within a quarter, at least one third-party fork of UNCUT with a different tool set will appear — 837 lines and MIT invite it. You can check via the forks list on GitHub.

Disclaimer: the UNCUT 0.1.1 release itself is official (MIT, GitHub Releases), but the claims about "lossless" context and speed are taken only from the author's README and BENCHMARKS.md and have not been independently verified.

Source: GitHub / Balanov/uncut