Linghun: a local-first, evidence-first AI programming terminal
Linghun is a local-first, evidence-first AI programming terminal that connects the model to real projects, tools, permissions, verification and context.
Verified
Core facts
Read by a script from the official API, not rewritten by a model.
- Repository
- https://github.com/linghungegeg/Linghun
- License
- Apache-2.0
- Languages
- TypeScript
- Maintainer
- linghungegeg
- Verified
What it is
Linghun is a local-first, evidence-first AI programming terminal. Linghun can be understood as an engineering exoskeleton fitted onto a large model. The model is responsible for understanding, reasoning and generation. Linghun is responsible for connecting the model to real projects, real tools, real permissions, real verification and real context.
Main features
Linghun supports OpenAI-compatible, DeepSeek and Anthropic Messages style endpoints, covering streaming output, tool calls, usage, reasoning, timeout, idle timeout, provider diagnostics and failure summaries.
Linghun has built-in tool paths such as Read, Write, Edit, MultiEdit, Grep, Glob, Bash, Todo, Diff and Git.
Linghun supports Workflow Matrix, durable job, background task, agent transcript, budget, steps, logs, reports and handoff boundaries.
Linghun distinguishes PASS, PARTIAL, FAIL, TIMEOUT, STALE and CANCELLED, and also distinguishes focused verification, mock verification, real smoke and unverified conclusions.
Before and after large changes, Linghun can check Git status, create stable points, help manage worktrees, and bind Git-related conclusions to the real repository state.
Roles such as planning, execution, review, verification, summary, vision and image can use different model routes.
Getting started
The environment requirements include Node.js 22 or newer.
After installing, start Linghun first, inside the project you want to work on; besides linghun, Windows also supports the capital-L Linghun as a compatible entry, so pick either one. /model setup runs after Linghun has started and /model doctor checks the provider configuration; both are used inside the Linghun session, not separately in the system terminal.
The model provider key is saved by default in a user-private configuration outside the project.
By default, Linghun runs code execution on your machine.
npm install -g @linghun/clilinghunLinghunlinghun --version/model setup/model doctorWho it is for
Linghun is still under active development. Linghun uses the Apache License 2.0. Linghun does not require users to first memorize a pile of English commands.
MCP, Skills, Plugins, Workflows and Hooks provide extension entries, following the principles of metadata before execution, trust before enabling, and diagnosis before use. Remote notifications and approvals can connect to channels such as WeCom, Feishu/Lark and DingTalk. The remote channel is not a remote execution platform that bypasses local permissions, and it should not send full code and full transcripts to external platforms at will.
Emotion and user state cannot replace permission confirmation, cannot serve as test evidence, cannot trigger automatic agent/job/workflow execution, and cannot justify writing files beyond permission because the user is in a hurry. Evidence Merge distinguishes evidence that can support a completion claim from states that can only serve as context. Agent summary, job completed, remote event and failure learning cannot pose directly as PASS.
Controlled memory uses a candidate-first confirmation flow, to avoid writing one-off emotions, sensitive information or unconfirmed facts as lasting rules. AntiCodeBlob prompts remind the model not to keep piling code into god files, overlong functions, deep nesting or unbounded global state. The Linghun architecture system watches module boundaries, dependency direction, responsibility backflow, duplicate runtime, permission bypass, diagnostic leakage, frontend/TUI experience constraints and delivery consistency.
The ordinary main screen does not leak provider, baseUrl, endpointProfile or plaintext key. Linghun is positioned as not bound to a single model vendor, and instead provides a unified engineering exoskeleton for different models. Linghun can use a code index, search, architecture evidence, workspace snapshots and large file protection to reduce repeated file reading.
Linghun can accumulate project rules, handoff summaries, controlled memory and failure experience. Linghun pays attention to prompt cache, tool list stability, context noise, long log summaries, cache freshness and usage tracking.
Related tools
Related tools listed on this site.
QwenPaw
Personal AI assistant and agent workstation running in your own environment. Its primary description presents local/cloud deployment, multiple channels and extension through Skills and MCP.
Cursor
A coding agent for understanding repositories, planning features, changing code, and reviewing diffs.
CodeBuddy IDE
AI coding workspace for product, design and development. Its dedicated IDE documentation describes technical conversations, code completion, Figma-to-code, diagnostics and a design-to-deployment workflow.
InfCode
Enterprise AI coding agent from TokensInfinity. Its site describes using enterprise code and knowledge through IDE, extension and CLI forms for completion, cross-file changes and engineering tasks.
阿里云百炼
Alibaba Cloud Bailian is a model-service and application-development platform for model inference, knowledge bases, agents and workflows.
千问云
The discovered Qianwen Cloud entry currently presents the Qianwen AI Platform for model discovery, online trials, API access, Skills/CLI tooling and usage/cost management.
Figma AI
Collection of AI workflows and capabilities within Figma products for design exploration, content editing and connecting prototypes with code.
Pixso AI
AI design capability integrated into the Pixso canvas. Its official page describes generating editable web and mobile interfaces from text, sketches or screenshots and continuing to refine existing designs.
Ardot
AI design and collaboration tool. Its official site describes text/image-to-editable UI, vector editing, dynamic layouts and design systems, plus collaboration, annotations and design-to-development handoff.
纳米AI
Nano AI is an AI search and agent assistant whose site says it can generate videos, reports and slides from one sentence and integrates several large models.
Exa AI
Exa is a web-retrieval service for AI applications and agents, describing search, content retrieval, deep research and structured-output APIs.
Related guides
- skill-creator: Anthropic's skill for writing, editing and evaluating skillsAnthropic's skill-creator skill: draft a skill, run evals, revise it and tune its trigger description. Not tested here; follow the official text and docs.
- webapp-testing Skill: drive local web apps with PlaywrightOfficial Anthropic webapp-testing Skill: Python Playwright scripts for local web apps, not tested
- frontend-design: avoiding templated defaults in UI designAnthropic's frontend-design skill: plan the visual direction, check it against the brief, avoid templated defaults. Not tested; follow the official text.
- gh-address-comments: work through GitHub PR review commentsOfficial OpenAI Codex skill: fetch comments on the current branch's PR with the gh CLI, number them, then fix only the ones you pick; not tested.
- security-best-practices: language- and framework-specific security reviewOfficial OpenAI Codex skill: load security references matching the project's language and framework, then write safer code or a report; not tested.
Sources
This page is written from the official sources below. Open them to check the original text.
- GitHub READMEhttps://github.com/linghungegeg/LinghunFetched
Verified ·
This page is written by AI from the official sources above and checked automatically. This site has not tested it. Found a mistake? Tell us.