LeanToken

io.github.morluto由 GitHub 帳號 morluto 發布v0.1.28更新於 Oct 24, 2026

Token-bounded repository context for coding agents over MCP.

已驗證STDIO僅桌面知識與記憶開發者工具

概覽

AI 產生的概覽

LeanToken 為本機程式碼倉庫建立索引,透過 MCP 向編碼助理提供受 token 預算限制的程式碼搜尋、結構大綱、精確讀取、歷史追溯與 JSON 查詢。

可以做什麼
  • 尋找並排序相關程式碼路徑,在倉庫內進行文字或正規表達式搜尋
  • 檢視定義、簽章、匯入與行範圍,不必讀取整個檔案
  • 讀取指定符號或連續行區間,並在 Git 版本間追蹤或比對符號
  • 查詢、彙總或比較有界的即時 JSON 檔案
  • 在明確的 token 預算內組裝面向任務的上下文,並以回執避免重複送出未變更的證據
  • 回報觀測到的 token 用量與雜湊抑制情況
適用情境
適合編碼助理在大型或不熟悉的倉庫中工作時使用,以減少大範圍探索與整檔讀取所消耗的 token。適用於初步排查、故障調查與變更審查,編輯、建置與測試仍由助理本身的工具完成。
執行需求
以本機 stdio 程序執行,需指定倉庫根目錄,可透過 npx 或 Cargo 安裝的執行檔啟動。npx 方式需要 Node.js;musl Linux 等其他目標須以 Rust 1.95 以上版本及 C/C++ 工具鏈從原始碼建置。未宣告帳號、API 金鑰或環境變數。
安裝前請注意
安裝程式會將 MCP 項目與一個探索技能寫入用戶端設定目錄,且除非強制否則不會取代同名的既有手動項目,建議先檢視 dry-run 結果。索引存放在本機,伺服器被描述為唯讀檢索,但會讀取倉庫原始碼與 Git 歷史,應避免指向過廣的根目錄或敏感程式碼樹。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 LeanToken,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

LeanToken

Code intelligence for agents: find the code that matters and keep your context window and tokens lean.

Language: English · 简体中文 · 日本語 · 한국어

  • MCP Registry name: mcp-name: io.github.morluto/leantoken
[LeanToken narrowing a large codebase to the files and code an AI agent needs]

[npm] [npm downloads] [Rust 1.95+] [License: MIT OR Apache-2.0]

Install · Why LeanToken · Tools · CLI · How it works · Docs


Measured token savings: In a controlled 60-run study, LeanToken used 20.1% fewer model input tokens than the agent's built-in tools with limited repository exploration, and 37.6% fewer than those tools with broad exploration. See exactly how it was measured in the measurement methodology.

Quick start

Add LeanToken to Claude Code, Cursor, OpenCode, Codex, Gemini CLI, or Antigravity:

bash
npx leantoken setup
Setup behavior and safety

Current releases stop setup before writing when npx resolves a stale project-local or ancestor install, and point to npx leantoken@latest setup. Older releases that predate this check can be bootstrapped directly with that versioned command.

The interactive setup wizard preselects supported clients it detects; you can change that selection before continuing. It then shows the exact configuration paths and MCP launcher and asks for a separate final confirmation. Automation never treats detection as consent. An npx-based setup pins the exact LeanToken version that ran setup, so restarting a client cannot silently move to a newer release.

Global setup never stores the repository where setup happened. OpenCode gets a workspace-relative working directory; other supported clients launch LeanToken from the workspace cwd selected by the host. If a host instead starts it from the home directory or a filesystem root, LeanToken refuses to index that broad root by default.

Restart or reload the configured clients, then verify the connection and first retrieval from a repository:

bash
npx leantoken doctor

Try a broad task such as: Find the code related to request cancellation before editing. LeanToken helps the agent start with leantoken.context, while its normal tools remain available for edits, builds, and tests.

Inspect LeanToken's observed repository-local token accounting:

bash
npx leantoken savings
Local by default

Source is indexed on your machine in a local database. LeanToken is a read-only discovery and retrieval layer.
Explicit token budgets

Every response has an explicit token limit, so large files cannot take over the request.
Built for agent workflows

Find files, search code, inspect structure, read exact ranges, trace history, query JSON, and track token usage through focused tools.
Advanced setup and version management

To skip the wizard, select clients explicitly or configure all supported clients:

bash
npx leantoken setup --claude --codex --yesnpx leantoken setup --all --yes

For regular use, --private-runtime is the recommended launcher: it copies the exact package-native executable into an application-data directory named by version and full BLAKE3 digest, so clients launch one verified process directly, without persistent npm/Node wrappers. Different source builds of one version can coexist. setup --refresh --private-runtime pins the current exact artifact; other referenced artifacts stay usable. It remains opt-in so the zero-install path does not add an application-data write. Preview its path and digest with --dry-run.

Automation never treats detection as consent: --yes requires explicit client flags, --all, or --refresh for entries already managed by LeanToken. Preview the same resolved plan without changing files:

bash
npx leantoken setup --codex --cursor --dry-run

Setup adds the leantoken MCP entry plus a small owned discovery skill only in the directories used by the selected hosts: Claude Code uses ~/.claude, while Codex and the other supported hosts use ~/.agents. The skill advertises routing metadata; it does not duplicate tool schemas, add rules, or install shell hooks. Setup marks new MCP launchers as managed and refuses to replace a same-name manual entry unless you review the dry-run and pass --force-unmanaged. Remove the owned integration with:

bash
npx leantoken remove

After private-runtime upgrades, inspect retained artifacts and preview a reference-safe cleanup before applying it:

bash
npx leantoken runtime listnpx leantoken runtime prune --dry-runnpx leantoken runtime prune --yes

Refresh only existing LeanToken MCP entries after explicitly choosing a new version, or use an older version to roll back:

bash
npx --yes leantoken@latest setup --refresh --yesnpx --yes [email protected] setup --refresh --yes --allow-outdated

Common agent workflows

LeanToken works best as a small evidence loop rather than a one-shot repository dump:

  1. Orient autonomous triage in one call. Start an uncertain broad task with context and plan_only: false, then use the materialized evidence directly. Make at most one focused follow-up only when coverage identifies a concrete missing implementation or regression-test owner.
  2. Continue without resending source. Pass the prior receipt_id on the next context call, or pass returned fragment hashes as known_hashes. The response reports exact and overlapping omissions instead of silently charging the same evidence again.
  3. Investigate an observed failure. Use the investigation workflow and provide only directly observed failure_traces, paths, symbols, or test intent in workflow_evidence. Follow with exact search, outline, or read calls for the owners the evidence identifies.
  4. Review a change. Use the review workflow with base_revision set to BASE..HEAD and strict_changed_paths: true. Request a handoff when another agent needs a compact manifest of selected hashes, changed paths, assumptions, and completed validations without copied source bodies.

This one-call contract is for autonomous repository triage, not a limit on implementation agents. Human review and control-plane flows can still preview expensive or high-risk retrieval with plan_only: true before materializing. The repeated multi-agent context suite found that an iterative LeanToken profile used 50.9% more total input than thin native, while the frozen one-context-plus-optional-one-search profile saved 20.1% and had 15/20 path-set successes. Those results cover four pinned triage tasks; they do not prove a universal implementation workflow.

Explicit focus constraints are contracts. When a request supplies focus_paths, exact focus_symbols, and minimum_fragments_per_focus_path, LeanToken generates candidates within the documented per-file bounds and reports a coverage failure when distinct ranges cannot satisfy the minimum. Explain-profile plans and materialized responses also identify the bounded allocation boundary that generated, reserved, selected, or suppressed each focus candidate without changing ranking.

Why LeanToken

Most agents start by searching widely and reading whole files. LeanToken narrows that work in stages:

Typical repository explorationWith LeanToken
Scan broad directory listingsFind relevant paths in a compact tree
Read whole files to find structureSee definitions and imports without loading the entire file
Send the same code again after each turnAvoid repeating unchanged evidence
Let large files fill the requestKeep returned source within an exact source-token budget and report response overhead separately
Guess which files matterRank likely relevant code for the task

Your coding agent still handles editing, commands, tests, and conversation. LeanToken finds and returns the code those tasks need.

LeanToken does not create one giant prompt file. It answers focused searches and reads as the agent needs them.

Example

For a task like fix request cancellation during shutdown, an illustrative bounded result might look like this:

text
Budget: 1,200 source tokens
Selected evidence:  src/services/executor.rs        lines 137-147, 251-259  src/services/reconciliation.rs lines 148-175, 257-272

The agent receives these ranges instead of both full files. If the budget is too small, the response also says what was left out. Paths, scores, receipts, JSON, and MCP transport wrappers are not part of this source-token budget; see token accounting for the measured boundaries.

Available tools

ToolPurpose
leantoken.contextDefault materialized first call for autonomous broad triage; optional preview for human or control-plane review.
leantoken.searchPrefer over grep/rg for ranked search; exhaustive text/regex calls can explicitly record or reuse complete query coverage.
leantoken.filesPrefer over find/ls/glob for compact, ignore-aware path discovery.
leantoken.outlineInspect definitions, signatures, imports, and ranges without whole-file reads.
leantoken.readPrefer over cat/head/sed for one exact symbol or inclusive line range.
leantoken.historyRead, batch-diff, or trace parsed symbols across immutable Git revisions.
leantoken.jsonQuery, summarize, or compare bounded live JSON with paged keys and typed diagnostics.
leantoken.receipt_rebaseExplicitly carry only same-path, same-coordinate, same-hash evidence into a newer completed generation.
leantoken.savingsReport observed response accounting, hash suppression, failures, and explicit observation limits.
Advanced retrieval controls

Every index-backed retrieval tool, including receipt_rebase, accepts consistency: "reconcile_working_tree" when completed edits must be reconciled before the query. The default, "indexed_generation", returns the latest completed index generation without scanning or waiting for filesystem changes; it is not a Git revision boundary. leantoken.history reads immutable Git objects and leantoken.json reads exact live files, so neither accepts an index consistency mode. To constrain context to immutable history, pass BASE..HEAD as leantoken.context.base_revision with strict_changed_paths: true.

For autonomous broad triage, set plan_only: false and use the materialized evidence directly. Reserve plan_only: true for human or control-plane review before expensive or high-risk retrieval: it returns bounded ranked candidate metadata without source fragments or receipt mutation. After approval, repeat the same request with plan_only: false. Set response_profile: "compact" for the smallest fail-loud response, keep the default "balanced" shape, or use "explain" for bounded individual omissions, facets, and diff evidence. The response reports the resolved choice as effective_response_profile.

The catalog stays intentionally small because every tool description and schema also consumes model context.

CLI usage

Run LeanToken directly through npx:

bash
npx leantoken statusnpx leantoken savingsnpx leantoken doctornpx leantoken --root /path/to/repo search handle_request

Or use a globally installed binary:

bash
npm install --global leantoken@latest
leantoken --root /path/to/repo indexleantoken --root /path/to/repo search handle_request --mode identifier --max-tokens 800leantoken --root /path/to/repo context \  --task "fix request cancellation during shutdown" \  --budget 2000

Audit an existing redacted experiment or host report without opening a repository index:

bash
leantoken episode audit \  --adapter multi-agent-suite-v1 \  --input benchmarks/reports/multi-agent-context-suite-v1-codex-0.144.1.json

The default projection is Markdown; add global --json for the stable normalized JSON schema. The auditor is local, bounded, and read-only with respect to its input. It retains artifact hashes, not raw prompts, source, tool arguments, or tool outputs.

npm install leantoken installs the command in the current project's node_modules/.bin; it does not add leantoken to the shell PATH. Invoke a project-local install through npx leantoken, a package script, or ./node_modules/.bin/leantoken.

Run the MCP server manually over stdio:

bash
leantoken --root /path/to/repo mcp
Manual MCP client configuration
json
{  "mcpServers": {    "leantoken": {      "command": "leantoken",      "args": ["--root", "/path/to/repo", "mcp"]    }  }}

For the Cargo distribution, install the published crate and point your MCP client at the resulting executable:

bash
cargo install leantoken --version VERSIONleantoken --root /path/to/repo mcp

The official MCP Registry entry is io.github.morluto/leantoken. Registry clients that support Cargo packages can install the matching leantoken version and use the mcp command shown above.

Installation options

The npm package includes native binaries for:

  • macOS on ARM64 and x64
  • glibc Linux on ARM64 and x64
  • Windows on x64

Installation does not run lifecycle scripts or download an executable from a postinstall hook. Other targets, including musl Linux, must build from source. Install Rust 1.95 or later and a native C/C++ toolchain, then run:

bash
cargo install --locked --git https://github.com/morluto/leantoken leantoken

Updating

MCP entries created through npx stay pinned to the exact LeanToken version that configured them. Update existing client integrations explicitly:

bash
npx --yes leantoken@latest setup --refresh --yes

For a globally installed CLI or a CLI installed with Cargo:

bash
leantoken upgrade --checkleantoken upgrade --yes

update is an alias for upgrade. For a project-local npm installation:

bash
npm install leantoken@latest

Pinned MCP entries never silently move to @latest. If the exact package is not available locally or online, startup fails rather than selecting another version. Updating the CLI does not change existing MCP entries. See the usage guide for rollbacks, cache management, and version details.

Cache management

Inspect local repository caches or preview cleanup before applying it:

bash
leantoken cache listleantoken cache list --summaryleantoken cache list --incompatible-with-currentleantoken cache prune --incompatible-with-currentleantoken cache compact --id v15-0123456789abcdef --dry-runleantoken cache prune --older-than 30 --dry-runleantoken cache prune --max-total-bytes 1073741824 --yes

See the usage guide for cache states, pagination, and cleanup safety rules.

How it works

text
repository    │    ▼file discovery ──► code structure extraction ──► local search index                                                    │                                                    ▼agent request ──► ranked / exact retrieval ──► focused code within a token budget

LeanToken indexes source once, then serves compact paths, ranked matches, structural outlines, exact source ranges, and task-specific context. It avoids resending unchanged evidence across turns.

Dependency-heavy workspaces can opt into a separate, cache-identified first-party index without changing the default whole-repository behavior:

bash
leantoken --index-include 'src/**' --index-include 'tests/**' index

Status and every retrieval disclose whether the active index is full or scoped, so an empty scoped result is never presented as whole-repository absence. See the usage guide for bounds, cache identity, and MCP registration examples.

LeanToken's goal is to return the code an agent needs with fewer input tokens.

Documentation

GuideContents
Usage and tool referenceCommands, MCP tools, request options, and examples
Architecture and reliabilityComponents, data flow, storage, and failure behavior
RoadmapCurrent direction and planned work
Development and testingLocal setup, validation, and release workflow
Benchmark methodologyToken-economy measurements and interpretation
Measurement harnessesExperiment, wire-cost, and profiling tools

License

Licensed under either of the following, at your option:

來源:README.md,提交 5cfab83

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

1
  1. v0.1.28最新Sep 16, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分

    請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。

  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分