接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
Semantic code search for code symbols and docs.
unch indexes functions, methods, types, classes, interfaces, and attached docs with Tree-sitter, then lets you search them from the terminal. It is useful when you know what the code does, but not the symbol name or file path. Local llama.cpp embeddings are the default; OpenRouter and remote publishing are opt-in.
unchunch stores top-level API objects and attached docs for Go, Rust, TypeScript, JavaScript, and Python.npm is the recommended install path:
npm install -g @uchebnick/unch
unch codex install
It installs the unch CLI and downloads the matching native release binary for your platform.
Run unch codex install once to register unch in Codex through codex mcp add and install the unch skill in ~/.codex/skills/unch/SKILL.md, then restart Codex.
Homebrew on macOS:
brew install uchebnick/tap/unch
Release installer on macOS and Linux:
curl -fsSL https://raw.githubusercontent.com/uchebnick/unch/main/install.sh | sudo sh -s -- -b /usr/local/bin
This installs unch into a system PATH directory so you can run it immediately after the installer finishes.
On Apple Silicon macOS, use /opt/homebrew/bin instead of /usr/local/bin:
curl -fsSL https://raw.githubusercontent.com/uchebnick/unch/main/install.sh | sudo sh -s -- -b /opt/homebrew/bin
PowerShell installer on Windows:
iwr https://raw.githubusercontent.com/uchebnick/unch/main/install/install.ps1 -useb | iex
This makes unch available immediately in the current PowerShell session.
To keep it available in future PowerShell or terminal sessions, add $HOME\AppData\Local\Programs\unch\bin to your user PATH.
Source install:
go install github.com/uchebnick/unch/cmd/unch@latest
This install path requires a cgo-capable Go toolchain. It is smoke-tested in CI in the official Debian-based Go container image. To pin a specific release, use @<tag>.
If you want to build from the current checkout:
go build -o unch ./cmd/unch
Published release archives:
arm64, x86_64arm64, x86_64arm64, x86_64 (unch.exe)On those targets, the installers use published release archives by default, so Go is not required. install.sh and install/install.ps1 only fall back to go install when no matching archive is available. install.sh is smoke-tested in CI on Ubuntu, Debian, Arch, and NixOS-like environments, including the default no--b path selection flow; the PowerShell installer is smoke-tested on Windows arm64 and x86_64.
Verify the installed binary:
unch --version
unch --help
See Compatibility for the support matrix and upgrade rules, and Benchmarks for the checked-in smoke, ci, and default suites.
Model selection accepts either a known model id or a direct .gguf path:
unch index --model embeddinggemma
unch index --model qwen3
unch search --model qwen3 "create a new router"
OpenRouter is also supported as an embedding provider:
unch auth openrouter --token sk-or-...
unch index --provider openrouter --model openai/text-embedding-3-small
unch search --provider openrouter --model openai/text-embedding-3-small "create a new router"
By default unch auth openrouter writes ~/.config/unch/tokens.json, so you do not need to keep the token in your shell environment. Use --local if you want to store it in .semsearch/tokens.json for one repository.
cd path/to/repo
unch index --root .
unch search "create a new router"
unch search --details "get path variables from a request"
OpenRouter flow:
cd path/to/repo
unch auth openrouter --token sk-or-...
unch index --root . --provider openrouter --model openai/text-embedding-3-small
unch search --provider openrouter --model openai/text-embedding-3-small "create a new router"
Real output from indexing gorilla/mux:
$ unch index --root .
Loaded model dim=768
Indexed 278 symbols in 16 files
$ unch search "create a new router"
1. mux.go:32 0.7747
2. mux.go:314 0.8135
$ unch search --details "get path variables from a request"
1. mux.go:466 0.7991
kind: function
name: Vars
signature: func Vars(r *http.Request) map[string]string
docs: Vars returns the route variables for the current request, if any.
The first local llama.cpp run may download the default embedding model, fetch local yzma runtime libraries, and create ./.semsearch/.
Each provider and model pair keeps its own active index snapshot. Rebuilding openrouter/openai/text-embedding-3-small does not replace the active llama.cpp/embeddinggemma snapshot until the new run finishes successfully.
The recommended MCP setup is through the npm wrapper:
npm install -g @uchebnick/unch
unch codex install
Then restart Codex. The setup command registers the unch MCP server and installs the unch skill in ~/.codex/skills/unch/SKILL.md.
After restart, ask Codex a codebase question as usual. The skill tells Codex to use these MCP tools before broad file reads:
workspace_status to inspect the repository root, state directory, active provider/model, and index statussearch_code to search indexed code symbols before opening many filesindex_repository to build or refresh the index when neededremote_sync_index to refresh from a bound GitHub Actions index before rebuilding locallycreate_ci_workflow, bind_remote_ci, and remote_download_index when you ask Codex to set up or use remote CI-backed indexesAll MCP tools accept an optional directory argument. The installed skill passes the active repository path so unch searches the workspace Codex is currently editing, even if the MCP server was launched from somewhere else.
Codex starts the stdio server for you with unch start mcp; you normally do not need to run that command by hand.
Go, Rust, TypeScript, JavaScript, and Pythonauto, semantic, and lexical search modes--comment-prefix and --context-prefix only for index fallback on unsupported files or parser failuresindexBuild or refresh the local index for a repository.
unch index --root .
Useful flags:
--exclude to skip generated, vendor, or irrelevant paths--provider to switch between local llama.cpp embeddings and remote openrouter--model to use embeddinggemma, qwen3, or a custom .gguf path--ctx-size to override the selected model context size; 0 uses the model default--lib to use an existing yzma runtime directory--state-dir to keep index state in a custom .semsearch directorysearchQuery the current index.
unch search "sqlite schema"
unch search --mode lexical "Run"
Useful flags:
--mode for auto, semantic, or lexical--limit to control result count--max-distance to narrow semantic matches--provider to select the embedding provider that matches the indexed snapshot--model to search with embeddinggemma, qwen3, or a custom .gguf path--ctx-size to override the selected model context size; 0 uses the model default--state-dir to search against a custom .semsearch directory--details to print symbol metadata, signature, docs, and body context for each matchRemote indexing is optional. Use it when you want GitHub Actions to publish an index for a repository. The generated workflow file is .github/workflows/unch-index.yml.
Create the workflow scaffold:
unch create ci
Bind the local repository state to a GitHub repo or workflow URL:
unch bind ci https://github.com/uchebnick/unch
After the workflow is committed and runs successfully once, unch search can refresh the published index automatically when a newer remote version exists. Use unch remote sync when you want to force a refresh before searching.
The cross-platform benchmark matrix stays separate from ordinary push CI. It runs on release-tag pushes and manual workflow_dispatch, uses the checked-in ci suite with a lighter 1 cold / 1 warm / 1 search repeat profile, and publishes per-platform summaries for Linux, macOS, and Windows.
Local-first semantic code search for repository annotations via GGUF embeddings and sqlite-vec.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Unch 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 Local-first semantic code search for repository annotations via GGUF embeddings and sqlite-vec.