# UNCUT 0.1 is a 7 MB headless coding-agent harness that never trims context or output

Canonical: https://theagenticweekly.com/news/uncut-0-1-lossless-headless-coding-agent-harness · Evidence: Reported · First seen: 2026-11-02 02:03 UTC · Source published: 2026-11-01 17:24 UTC · Brief published: 2026-11-04 20:16 UTC · Updated: 2026-11-04 20:16 UTC · Author: Agentic Weekly

> UNCUT is an MIT-licensed open-source coding-agent harness first released as v0.1.0 on GitHub on November 1, 2026 (v0.1.1 followed the same day). It is a single static Go binary with four tools that works with OpenAI-compatible endpoints, and its README says "The entire conversation is sent on every model call."

## What changed

- The repository was created on November 1, 2026; v0.1.0 was tagged that day and v0.1.1, a Windows reliability patch, about an hour later. Binaries ship for Linux, macOS and Windows on amd64 and arm64.
- It has exactly four tools (read_file, write_file, replace, run_command) and works with any OpenAI-compatible Chat Completions endpoint, including local servers such as Ollama.
- There is no TUI, telemetry, plugin system, automatic retries or context compaction: file contents and command output are never truncated, and a response cut at the provider's output limit makes UNCUT exit with an error.
- It is built for pipelines and CI: piped input is appended to the prompt, --json emits NDJSON events, --session resumes an append-only JSONL history, and --max-steps defaults to 32.
- The project's own Windows measurement reports a 9.461 ms median --version launch and a 7,109,120-byte executable, and its README counts 837 lines of implementation Go.

## In the source's words

> "The entire conversation is sent on every model call."
> (UNCUT README (GitHub), https://github.com/Balanov/uncut)

> "If the provider cuts a response at its output limit, UNCUT exits with an error instead of presenting a partial answer as success."
> (UNCUT README (GitHub), https://github.com/Balanov/uncut)

## What we have not verified

- We did not install UNCUT or run it against a model.
- The launch-time and size numbers are the project's own, measured on one Windows machine; we did not reproduce them.
- Because the full history is sent on every call, token cost on long sessions grows with every step; we did not measure it.

## Timeline

- Source published: 2026-11-01 17:24 UTC
- First seen by our watcher: 2026-11-02 02:03 UTC (via GitHub discovery sweep; primary source is the GitHub repository)
- Brief published: 2026-11-04 20:16 UTC

## Sources

- [UNCUT repository and README on GitHub](https://github.com/Balanov/uncut) (primary source) (fetched 2026-11-04)
- [UNCUT benchmarks](https://github.com/Balanov/uncut/blob/main/BENCHMARKS.md) (primary source) (fetched 2026-11-04)
- [UNCUT v0.1.1 release](https://github.com/Balanov/uncut/releases/tag/v0.1.1) (primary source) (fetched 2026-11-04)

## Related

- [Claude Code: verdict and setup guide](https://theagenticweekly.com/tools/claude-code)
- [Codex CLI: verdict and setup guide](https://theagenticweekly.com/tools/codex-cli)
- [Aider: verdict and setup guide](https://theagenticweekly.com/tools/aider)
- [Ollama: verdict and setup guide](https://theagenticweekly.com/tools/ollama)
- [All news briefs](https://theagenticweekly.com/news)
- [This week's releases](https://theagenticweekly.com/this-week)
- [New AI models this week](https://theagenticweekly.com/models/this-week)

## How to cite

Agentic Weekly, "UNCUT 0.1 is a 7 MB headless coding-agent harness that never trims context or output", published 2026-11-04 20:16 UTC, https://theagenticweekly.com/news/uncut-0-1-lossless-headless-coding-agent-harness