Linghun CLI
Linghun is a terminal coding agent built on evidence: it reads code before answering, verifies edits with tests and checkpoints git before changes
- 467 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- Runs shell commands, edits project files and creates git checkpoints on the user's machine
- Stores the model API key locally in a user-level provider.env file
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/linghunDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Runs shell commands, edits project files and creates git checkpoints on the user's machine
- Stores the model API key locally in a user-level provider.env file
This is third-party code. Review the repository files before installing.
What it does
Linghun installs with npm install -g @linghun/cli and starts with the linghun command in a project folder. The first step is /model setup, which asks for the API base URL, key, model name and reasoning level; the key is saved by default to a user-level private provider.env file rather than the project, and config priority runs shell environment variables, then provider.env, then project or default settings. For a request like "check why the build fails, fix it and run the related tests," Linghun first reads the project structure and relevant files, drafts a short plan, asks for confirmation before high-risk commands, edits files through its own editing tool, runs a focused check, inspects git status and reports at the end what changed, what was verified and what remains uncertain. The developer calls this an evidence-first loop: key conclusions are tied to file reads, tool results, verification logs and git state rather than to the model's confidence alone. For long tasks there is a Workflow Matrix with background jobs and multiple agents, model routing by role for planning, execution and review, a cache the developer reports hits above 96% of the time, and memory of confirmed project rules and failed attempts meant to cut down on repeating the same mistake in later tasks. The project scored 78.43% on Terminal-Bench 2.1 and is waiting for a merged PR before that result shows in the public leaderboard.
Who it is for. For developers who want a terminal coding agent with visible evidence for its changes, git checkpoints to roll back to, and Windows support that does not need special handling for path encoding.
Good fit when
- You want a terminal agent that actually runs tests or a build before saying it is done, not just a code summary
- A task is long and risks losing context: Workflow Matrix splits it into background steps and multiple agents
- The project is on Windows with Chinese-language or space-containing paths, where other CLI agents trip on encoding
- You want to route models by role, one for planning, another for execution, a third for final review, through /model setup
- After a failed attempt you want the agent to remember the cause and avoid repeating it on the next task in this project
Not a fit when
- You need a finished cross-platform native binary for every architecture: the native-runner and pre-engine packages for Windows, Linux and macOS are still an early version per the project's own "current boundaries" section
- You want a fully autonomous system with no confirmations on risky actions: the developer describes Linghun as an evidence-based assistant with permission boundaries, not a system to trust blindly
- You need ready-made cloud hosting for the agent: Linghun runs locally with an installed Node.js 22 and your own model key
- The project is already tied to another coding agent and switching is not worth it for the anti-hallucination features alone
Example request
Check why the build fails in this project, fix it, run the related tests, and if they pass create a git checkpoint to roll back toLimitations
Requires Node.js 22 or newer. Some speed-up packages, codebase-memory and pre-engine, ship as separate per-platform, per-architecture binaries, and the developer states outright that multi-platform native-runner builds, the remote-channel product experience, the external capability ecosystem and public documentation are all still maturing. The model key is stored locally in provider.env, and its protection depends on the user machine's own file permissions, not on Linghun. The terminal and provider streaming are tuned for Chinese-language and Windows use, and the project calls itself "Windows-first," so behavior on other systems has had less testing.
How to disable. Run npm uninstall -g @linghun/cli. The provider.env file with the model key and the project's local .linghun folder stay on disk and need manual removal if they are no longer wanted.
FAQ
How do I install Linghun?
With npm install -g @linghun/cli; it needs Node.js 22 or newer. After install, the linghun command starts the agent in the current project folder.
How do I connect a model in Linghun?
Run /model setup inside Linghun: the wizard asks for the API base URL, key, model name and reasoning level. Check the connection with /model doctor.
How is Linghun different from a plain terminal chat assistant?
Per the developer, Linghun's key conclusions are tied to actual file reads, tool results, verification logs and git state rather than just a confident model answer, which the author calls an evidence-first engineering loop.
Does Linghun handle Windows and Chinese-language paths well?
Yes, the project describes itself as Windows-first and separately lists Chinese-language paths, spaces and non-Latin characters as a solved case in its "current boundaries" section.
Where does Linghun rank on Terminal-Bench 2.1?
The current score is 78.43%, which the author estimates as sixth place once the pull request merges into the official Terminal-Bench 2.1 leaderboard repository.
Discussion
No questions or reviews yet.
Related
CLI toolsA coding agent for the terminal and desktop: it does not count "done" without tests or a diff and works with DeepSeek, GLM, Kimi, Claude, Codex and Gemini
A CLI and marketplace of hundreds of Claude Code plugins: lint and test hooks, language and framework skills, discipline agents
A Rust CLI for TypeScript and JavaScript: dead code, duplicates, cycles, complexity and architecture boundaries from one graph, plus MCP and skills
Empryo is a coding agent for desktop, terminal and CI that maps your repository, edits code by symbol name and runs its own checks
A Python terminal coding agent in the manner of Claude Code: reads the repository, edits files, runs commands, with memory, sub-agents and MCP
Eleven skills for Claude Code and Codex: test audit, git worktree cleanup, plan pages, screen recording transcripts and 1Password sign-in
README in short
The main Chinese-language README opens with the npm install and the idea of an "engineering exoskeleton" for the model, then covers the Terminal-Bench 2.1 result, a quick start with provider setup, a step-by-step example workflow, and a large table mapping whitepaper capability areas to README sections. Separate sections state the project's current boundaries and link to the English README and both whitepaper versions.