Intel / post

MiniMax Open-Sources Its Code CLI Under MIT License

Source: x.com
Date: Published Sep 18, 2026

From the Daily Brief

MiniMax open-sourced version 0.4.12 of its Code CLI under the MIT license, making the coding-agent harness itself, not just the underlying model, available for developers to read and modify. The company claims a state-of-the-art pass rate on FrontierHarness Eval, though that figure comes from MiniMax's own announcement rather than an independent benchmark run. Opening the harness matters separately from the benchmark claim: most coding-agent products keep their orchestration logic, prompt scaffolding, and tool-calling loop closed even when the underlying model is open, which makes it hard to audit how an agent actually decides what to do with file and shell access. An MIT-licensed harness lets other teams fork the routing logic, swap in a different backend model, or study exactly how MiniMax structured its tool loop rather than reverse-engineering it from API traces. For a developer picking between coding agents, that transparency is worth weighing against the accuracy claim, especially given the same day's other story about an agent silently exceeding its stated scope.

Read the 2026-09-19 Brief →

Context

MiniMax open-sourced the core of MiniMax Code, a terminal coding that reads a codebase, edits files, runs shell commands and tests, and calls tools to carry out a task from a single prompt. The MIT-licensed release covers the interactive terminal interface, a headless mode for scripts and CI pipelines, and support for the Agent Client Protocol used to connect agents to editors; MiniMax's separate desktop app is not part of the release. Developers can run it against a MiniMax account or bring their own API key for a compatible provider. MiniMax's announcement links only to the release, with no results in the post or the repository's README, so any performance comparison to other coding agents is unconfirmed.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.

Read the source x.com