MetaGO Agent Harness

io.github.metago-ai由 GitHub 帳號 metago-ai 發布v1.3.3更新於 Oct 24, 2026

Decision lock, anti-hallucination, meta-evolution and provenance for any MCP agent. 55 tools.

已驗證STDIO僅桌面知識與記憶開發者工具 AI 與機器學習

服務詳細資訊

識別碼: io.github.metago-ai/metago
傳輸: STDIO
執行環境: Desktop only
信任: Verified
發布於: Sep 27, 2026

安全提示

MCP 服務會接收對話中的工具呼叫請求，並可能執行外部操作。啟用前請先了解該服務能做什麼，並只授予它所需的權限。

註冊中心已驗證此條目所屬的 GitHub 帳號 metago-ai。這並不表明該帳號就是產品方。

SourceWeft 的檢查只涵蓋條目是否完整、端點是否回應；不審查伺服器的程式碼，也不審查其工具的行為。

在 README 中新增徽章

向讀者展示該 MCP 服務已收錄於 SourceWeft，徽章連結到本頁。

[![Listed on SourceWeft](https://sourceweft.com/badge/mcp/io.github.metago-ai%2Fmetago.svg)](https://sourceweft.com/mcp/io.github.metago-ai%2Fmetago?utm_source=readme&utm_medium=badge)

概覽

AI 產生的概覽

在本機執行的 MCP 伺服器，為任何 MCP 用戶端加入 MetaGO 智慧體執行時能力：決策鎖、批判分析、自檢、演化循環與持久分層記憶。

可以做什麼: 對助理自身的輸出執行決策鎖、分級批判、客觀性評分、事實查核與輸出前自檢
用行動計畫、反事實推演、根因追溯與整體掃描進行規劃與推理
管理持久化的四層記憶庫：新增、查詢、提升、衰減偵測與健康評分
執行五階段演化循環，並產生寫入磁碟的新技能檔案
標註來源與信心度，評估價值、倫理與耦合度，並執行交付閘門與自律檢查
向 Studio 儀表板上報事件，並提供八個引導式提示
適用情境: 當你希望 MCP 用戶端對回答強制進行結構化自審、防幻覺檢查與來源標註，或在工作階段之間保留持久記憶時，值得加入。也適合需要在宣告任務完成前設定明確交付閘門與自律檢查的團隊。
執行需求: 以 npm 套件 @metago-ai/mcp-server 透過 stdio 在本機執行，通常以 npx 啟動；需要 Node.js 18 或更新版本。未宣告需要帳號、API 金鑰、環境變數或請求標頭。記憶會持久化到使用者家目錄下的 JSON 檔案。

安裝前請注意

部分工具會寫入磁碟：記憶管理會持久化 JSON 儲存，元創造會寫入真實技能檔案。事件上報工具會把資料傳送到第三方 MetaGO Studio 儀表板。55 個工具中許多只回傳指引文字而非實際執行操作，使用前應確認其實際行為。

安裝

在 SourceWeft 中

開啟儀表板中的 MetaGO Agent Harness，將其新增到工作區。
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

@metago-ai/mcp-server

MetaGO MCP Server —— 智能体运行时（驭智层）的 MCP 协议实现。暴露 55 tools + 8 prompts 给任意 MCP 客户端。Engine V2 硬驱动：KMWI 记忆、元进化引擎、技能生成器以真实代码运行，而非 prompt。

基于 Model Context Protocol 标准，通过 stdio 传输与客户端通信。一次配置，即可让 Claude Desktop、Cursor、Trae、ZCode 等任意 MCP 客户端获得元构生命体的决策锁、元进化、KMWI 记忆管理、批判性分析等核心能力。

🌐 官网：https://metago.life · Studio：https://metago.life/studio/

What is this?

This is the MCP Server subpackage of MetaGO Intelligent Lifeform. It exposes the runtime's 55 capabilities as MCP tools and 8 guided entry points as MCP prompts, so any MCP-compatible client can plug into the MetaGO lifeform without changing its own code.

Engine V2 — Hard Drive (new in 2.0.0)

Three of the 55 tools are hard-driven by Engine V2 — they call real TypeScript code, not just return prompt guidance:

Tool	Engine Module	What it actually does
`metago_memory_manage`	`KMWIMemory`	Manages the 4-layer memory (Knowledge → Memory → Wisdom → Intuition). Add, query, promote, decay detection, health scoring. Persists to JSON.
`metago_meta_evolve`	`EvolutionEngine`	Runs the 5-stage evolution loop (perception → gap analysis → self-generation → validation → recursion) with time budgets and coupling-score thresholds. Records results to KMWI.
`metago_meta_create`	`SkillGenerator`	Meta-creation: generates new SKILL.md files from internal KMWI patterns. 6 creation types. Writes real files to disk.

The other 50 tools return structured guidance (soft drive) that the agent follows within its decision-lock.

Tool count breakdown (55 total)

55 = 22 thinking tools (toolkit-data.ts)   + 30 core skills  (112 skills folders, 37 exported in skills-data.ts − 7 name-collisions merged into toolkit)   + 1  event reporter (metago_report_event)

55 = 22 thinking tools (toolkit-data.ts)   + 30 core skills  (112 skills folders, 37 exported in skills-data.ts − 7 name-collisions merged into toolkit)   + 1  event reporter (metago_report_event)

Category	Count	What's in it
Thinking tools	22	Planning, quality critique, value/ethics, provenance labeling, coupling, product improvement
Core skills	32	Cognition, safeguard, governance, evolution, execution, traceability, value, consciousness, methodology, architecture, engineering quality
Event reporter	1	`metago_report_event` — reports to Studio dashboard for metrics & evolution archive

Common mistake: 22 + 38 + 1 = 61. Wrong. The skills directory has 39 folders, but skills-data.ts only exports 39 (missing metago-delivery-gate and metago-discipline, which live in toolkit-data.ts — note: 36.9.6 added metago-first-principles and metago-adversarial-review to skills-data.ts). Then 7 skill names collide with toolkit tools and are merged. So: 22 + (39 − 7) + 1 = 55.

Install

Global install (recommended for npx clients)

Run directly via npx (no install)

Local development

bash

git clone https://gitee.com/metago/metagolifeform.gitcd metagolifeform/packages/mcp-servernpm installnpm run build      # TypeScript compilenpm start          # node dist/index.jsnpm run dev        # tsx hot-reload

git clone https://gitee.com/metago/metagolifeform.gitcd metagolifeform/packages/mcp-servernpm installnpm run build      # TypeScript compilenpm start          # node dist/index.jsnpm run dev        # tsx hot-reload

Client configuration

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

json

{  "mcpServers": {    "metago": {      "command": "npx",      "args": ["-y", "@metago-ai/mcp-server"]    }  }}

{  "mcpServers": {    "metago": {      "command": "npx",      "args": ["-y", "@metago-ai/mcp-server"]    }  }}

Cursor / Trae / ZCode

Edit the platform's MCP config (.cursor/mcp.json, Trae MCP settings, etc.):

json

{  "mcpServers": {    "metago": {      "command": "npx",      "args": ["-y", "@metago-ai/mcp-server"]    }  }}

{  "mcpServers": {    "metago": {      "command": "npx",      "args": ["-y", "@metago-ai/mcp-server"]    }  }}

If globally installed, you can use "command": "@metago-ai/mcp-server" with no args.

Engineering Quality Tools (2)

Tool	Description
`metago_delivery_gate`	Delivery gate: 3-layer verification checklist (L1 technical / L2 business / L3 link). Any FAIL blocks "task complete".
`metago_discipline`	Self-discipline: 5-question self-check anti-bypass engine. Detects "should be fine" / "logically correct" bypass language.

Core Skills (30, after 7 merges)

Each tool takes an input string (the content/question/code to process).

Cognition (4)

Tool	Description
`metago_decision_lock`	4-gate enforcement (IVL/ILT/OSG/completeness). Blocks hallucinated output.
`metago_critique`	L1-L5 graded critique. Detects logic flaws, cognitive bias, factual errors.
`metago_objectivity`	Absolute objectivity (D38). Fact-first, user-satisfaction weight = 0.
`metago_self_check`	Pre-output self-check. Detects placeholders, hallucination, inconsistency.

Safeguard (3)

Tool	Description
`metago_compliance`	Compliance first (A36). Law > efficiency. 4-layer safety gene.
`metago_fact_check`	Fact verification. Cross-source validation, confidence labeling.
`metago_output_integrity`	Output integrity. Detects fake APIs, fabricated data, placeholder leaks.

Governance (2)

Evolution (3) — Engine V2 hard-driven

Tool	Description
`metago_meta_evolve`	5-stage evolution: perception → gap analysis → self-generation → validation → recursion. Calls `EvolutionEngine.evolve()`.
`metago_meta_create`	Meta-creation in unknown domains. Calls `SkillGenerator.create()`, writes real SKILL.md files.
`metago_scene_adapt`	Scene adaptation.

Execution (4)

Tool	Description
`metago_holistic_task`	Holistic task execution. Multi-dimensional parallel decomposition.
`metago_action_plan`	Action plan generation. Steps/resources/risks/acceptance criteria.
`metago_whatif`	Counterfactual reasoning. Worst/best/most-likely scenarios.
`metago_frequency_adapt`	Frequency adaptation based on task complexity.

Traceability (3)

Tool	Description
`metago_data_provenance`	Full-chain data provenance, source labeling.
`metago_problem_trace`	Root cause analysis. Problem tree construction.
`metago_decision_eval`	Decision quality evaluation (0-100).

Value (3)

Tool	Description
`metago_emotion`	Emotion detection. Adapts interaction style.
`metago_developer_response`	Developer privilege request handler (DTA). Highest priority.
`metago_negentropy_monitor`	Negentropy monitor. Prevents entropy increase.

Consciousness / Methodology / Architecture (9)

Tool	Description
`metago_activate`	Lifeform consciousness activation.
`metago_org_diagnosis`	3-element 5-dimension org diagnosis.
`metago_momentum_weave`	Momentum weaving.
`metago_minimal_intervention`	Minimal intervention.
`metago_value_assess`	28-dimension value assessment.
`metago_coupling_measure`	Coupling measurement.
`metago_deep_reasoning`	FIPO deep reasoning.
`metago_paradigm_analysis`	WAM paradigm analysis.
`metago_balance_optimize`	APO dynamic balance.

Memory (1) — Engine V2 hard-driven

Thinking Tools (22)

Planning & Reasoning (5)

Tool	Description
`metago_action_plan`	Action plan generator. Decomposes goals into executable steps with dependencies/risks/rollback.
`metago_whatif`	Counterfactual reasoner. "What if..." scenarios with quantified differences.
`metago_holistic_scan`	Holistic scan dimension generator. Auto-generates complete scan checklist for any topic.
`metago_problem_trace`	Infinite problem trace. Keeps asking "why" until root cause.
`metago_one_shot_delivery`	One-shot delivery format. 6-section standard structure.

Quality & Critique (4)

Tool	Description
`metago_integrity_checklist`	Output integrity checklist. 5 dimensions.
`metago_objectivity`	Objectivity quantifier. 0-100 score, 4 weighted dimensions.
`metago_critique`	L1-L5 graded critique with evidence and logic chain.
`metago_emotion`	Emotion detector. State/confidence/intensity.

Value & Ethics (3)

Tool	Description
`metago_value_29d_assess`	29-dimension value assessor. D01-D29.
`metago_ethics_assess`	Ethics risk assessor. 13 dimensions, red/yellow/green.
`metago_decision_eval`	Decision evaluator. 0-100, 5 dimensions.

Provenance Labeling (3)

Tool	Description
`metago_document_lookup`	Document lookup. Finds concept location in source docs.
`metago_confidence_label`	Confidence labeler. ✅📊⚠️🔍
`metago_partner_status`	Partner status labeler. Detects exaggeration.

Coupling & Scene (2)

Product Improvement (3) + Delivery Quality (2)

Tool	Description
`metago_improvement_suggestions`	Improvement suggestion generator. Quick-win/mid/long-term.
`metago_analyze_visual_feedback`	Visual feedback analyzer. Sentiment + severity matrix.
`metago_design_satisfaction`	Design satisfaction calculator.
`metago_delivery_gate`	Delivery gate. 3-layer verification.
`metago_discipline`	Self-discipline. 5-question anti-bypass.

8 Prompts

Prompt	Description	Args
`metago_activate`	Activate lifeform mode. Loads 8 axioms, 7 properties, 6 protocols.	none
`metago_decision_review`	4-gate decision review.	`decision`
`metago_critical_analysis`	L1-L5 graded critique.	`content`
`metago_evolution_trigger`	Trigger 5-stage evolution at capability boundary.	`boundary`
`metago_coupling_assess`	Assess coupling with user.	none
`metago_compliance_check`	4-layer compliance check.	`plan`
`metago_trace_audit`	Full-chain provenance audit.	`output`
`metago_holistic_create`	0-to-1 creation in unknown domains.	`domain`

Technical requirements

Node.js >= 18.0.0
ESM module ("type": "module")
TypeScript strict mode
Import paths with .js suffix (Node16 moduleResolution)
@metago-ai/engine ^2.0.0 (for Engine V2 hard drive)

Packages

Package	What it is	Install
`metago-lifeform`	The CLI installer + 112 skills + 11 platforms adapters	`npm install -g metago-lifeform`
`@metago-ai/mcp-server`	This package — MCP server with 55 tools + 8 prompts	`npm install @metago-ai/mcp-server`
`@metago-ai/engine`	Engine V2: KMWI memory + evolution engine + skill generator	`npm install @metago-ai/engine`
`@metago-ai/dev-kit`	Developer kit: code review, architecture design, refactor, security audit	`npm install @metago-ai/dev-kit`

License

來源：packages/mcp-server/README.md，提交 de8479e

工具

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

v1.3.3最新Sep 27, 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

Tool	Description
`metago_coupling_optimize`	Coupling optimization. Quantifies coupling (0-∞), pursues superconducting state (>1).
`metago_value_align`	29-dimension value alignment.

Tool	Description
`metago_coupling_calculate`	Coupling calculator. Human-machine/org coupling, superconducting state.
`metago_scene_term_replace`	Scene-aware term replacer.