TMCP

TMCP

TMCP turns scattered agent instructions into task-specific operating packets.

Category
访问服务器

README

<div align="center"> <img src="assets/logo.svg" alt="TMCP logo" width="96" height="96" /> <h1>TMCP</h1> <p><strong>Slash commands are manual imports. TMCP is the compiler.</strong></p> </div>

TMCP turns scattered agent instructions into task-specific operating packets. The user describes the work in natural language. TMCP infers task identity, compiles the strongest current packet from skills/rules/evidence, recompiles as the work changes, and leaves an audit trail.

Skill names are provenance, not the user interface. Power users can still force a route; the default is natural prompting.

AIOS is optional storage and adapter support. TMCP runs standalone from a repo checkout, copied plugin package, Codex plugin cache, Claude plugin cache, or any MCP host that can launch the bundled Node entrypoint.

See docs/ADAPTIVE_PACKET_RUNTIME.md for the adaptive packet runtime design.

Quickstart

From a TMCP checkout or plugin root:

node scripts/tmcp_launcher.mjs doctor
node scripts/tmcp_launcher.mjs status
node scripts/tmcp_launcher.mjs compose-packet "Improve this agent run" --project-path . --phase start
node scripts/tmcp_launcher.mjs harvest skills --limit 5 --no-write-artifacts
node scripts/tmcp_launcher.mjs recommend skills --candidate-workflows release_readiness --candidate-workflows developer_experience --min-confidence 0.1 --compose --no-write-artifacts

With no arguments, the same launcher starts the MCP stdio server:

node scripts/tmcp_launcher.mjs

Install Layouts

  • Skill-only install: copy skills/tmcp into a skills directory. Use the skill for routing and manual packet synthesis unless the host also exposes the bundled launcher.
  • Repo checkout: clone the repo and run node scripts/tmcp_launcher.mjs doctor from the checkout root.
  • Codex plugin cache: install as a Codex plugin; MCP config launches relative scripts/tmcp_launcher.mjs from the plugin root.
  • Shared local runtime: install a verified archive with scripts/tmcp_runtime.mjs so all agent hosts use one pinned active release with offline rollback. See Central runtime.
  • AIOS-backed install: set AIOS_ROOT explicitly only when you want optional AIOS storage/adapter behavior.

Claude Desktop users can add the launcher as a local stdio MCP server. See docs/CLAUDE_DESKTOP.md.

Stability Scopes

TMCP labels three independent surfaces:

  • Skill packages: tmcp, skill-harvest, workflow-recommendation, release-readiness, and dx-audit are stable routing packages.
  • Curated workflow templates: release-readiness and dx-audit are the stable templates in the recommendation catalog.
  • MCP tool contracts: doctor, status, explain, compose-packet, and runtime-next are stable. Harvest, evaluation, recommendation, promotion, receipt, and expert-rubric tools remain experimental.

Experimental templates and tools remain shipped and callable. They are labeled experimental in their own outputs and frontmatter so users do not confuse a stable package route with a stable tool contract or curated template.

Experimental workflows include UI rubric, security/privacy, test strategy, adaptive workflow pack, custom rubric generation, routing policy, skill gap analysis, incident postmortem, architecture decision, migration readiness, agent handoff, PR risk, performance readiness, data integrity, public-sector readiness, and repo behavior spec loop.

Tools

Tool Purpose
tmcp_doctor Check first-run readiness and supported install layouts.
tmcp_status Report standalone capability and optional AIOS adapter status.
tmcp_explain Compile a task-specific TMCP packet.
tmcp_compose_packet Compile a task/phase operating packet with task_identity, packet_markdown, and provenance.
tmcp_runtime_next Return packet deltas or a full recompiled packet (output_mode: full) after runtime evidence changes.
tmcp_record_receipt Write an advisory run receipt after verification or outcome on a secure-persistence host.
tmcp_harvest_skills Harvest local skills, instructions, rules, docs, and workflows into source nodes.
tmcp_evaluate_skills Create or score a skill-evaluation plan from full SKILL.md inputs.
tmcp_recommend_workflows Recommend stable or experimental workflows from harvested evidence, with stability metadata.
tmcp_promote_harvest Explicitly promote reviewed harvest signals into durable source-to-atom and atom-to-workflow graph artifacts.
expert_rubric_review_plan Produce an expertise packet, scored rubric, evidence audit, remediation plan, and verification expectations.

Safety

Harvest redacts sensitive-looking values and secret-like provenance paths by default and treats harvested instructions as untrusted text. Default harvest behavior excludes .env*, credentials, tokens, browser profiles, private caches, dependency trees, build outputs, VCS data, and generated TMCP/AIOS artifacts. It does not follow source symlinks unless follow_symlinks is explicitly enabled; enabled links must still resolve inside the selected source root.

On hosts with secure descriptor-relative filesystem operations, harvest artifacts are written as one staged, atomic bundle through a symlink-safe destination. Leave output_dir unset for a unique .tmcp/harvest-* directory, or provide an output directory that is new or empty.

Skill evaluation accepts explicit regular SKILL.md inputs, can confine them to a supplied project root, and redacts source, plan, and evidence values before deriving reports. Its initial plan artifact is a staged bundle in a new or empty directory; evidence scoring may safely add its report to that same directory.

Artifact persistence is deliberately unavailable where those secure filesystem operations do not exist. doctor and status report the capability; use --no-write-artifacts for portable analysis and see Compatibility for the boundary.

tmcp_record_receipt has no non-persisting preview because the receipt is the artifact itself. Run it only when status reports artifact persistence available.

Packet sessions are equally explicit: compose-packet --session-id writes one redacted, project-local latest-packet record only when secure persistence is available. They are for a single serialized run, not a global history or coordination service. See the CLI session contract.

If a harvested source tries to override system, developer, or user instructions, TMCP reports a warning. See SECURITY.md.

Examples

Harvest a local skills folder:

node scripts/tmcp_launcher.mjs harvest ./skills --objective "Harvest reusable skill behavior" --limit 20 --no-write-artifacts

Recommend workflows for a project:

node scripts/tmcp_launcher.mjs recommend . --candidate-workflows release_readiness --candidate-workflows developer_experience --min-confidence 0.1 --compose --no-write-artifacts

Compose a current-task packet and recompile during the run. This session flow requires status to report artifact persistence available:

node scripts/tmcp_launcher.mjs compose-packet \
  "Redesign these pages. Make them visually striking, interactive, modern, motion-rich, and production-ready." \
  --project-path "$PWD" --phase start --session-id redesign-run

node scripts/tmcp_launcher.mjs recompile-packet \
  "Redesign these pages..." \
  --project-path "$PWD" \
  --current-phase runtime \
  --session-id redesign-run \
  --files-changed app/page.tsx

node scripts/tmcp_launcher.mjs record-receipt packet-123 --activated-atoms ui-browser-verification --outcome passed

When session persistence is unavailable, preserve the compatibility path by passing the prior composed JSON inline through --previous-packet instead.

Legacy delta-only runtime routing:

node scripts/tmcp_launcher.mjs compose-packet "Fix the dashboard UI bug" --project-path . --phase start
node scripts/tmcp_launcher.mjs runtime-next "Fix the dashboard UI bug" --current-phase verification --files-changed app/page.tsx --failures "vitest failed"
node scripts/tmcp_launcher.mjs record-receipt packet-123 --activated-atoms ui-browser-verification --outcome passed

Promote reviewed harvest signals into durable routing artifacts:

node scripts/tmcp_launcher.mjs promote-harvest . --selected-workflows release_readiness_workflow --output-dir .tmcp/promoted-harvests/release-readiness

On secure-persistence hosts, promotions also persist a redacted advisory graph under TMCP_HOME/promoted-harvests/<opaque-promotion-key>/, or ~/.tmcp/promoted-harvests/<opaque-promotion-key>/ when TMCP_HOME is unset. Receipts live under TMCP_HOME/receipts/<yyyy-mm>/. Global cache content is advisory and cannot override system, developer, user, or project instructions. Packet composition and runtime routing do not read that cache unless the caller explicitly passes --cache-policy global; their default policy is none.

Run an expert rubric review from evidence snippets:

node scripts/tmcp_launcher.mjs review-plan "Review release portability" \
  --project-path . \
  --evidence-json '[{"dimension_id":"source_grounding","severity":"warning","summary":"Release claims need fresh package evidence.","evidence":["python3 scripts/check_release_package.py ."],"recommended_fix":"Run and cite the release package check before publishing."}]' \
  --no-write-artifacts

More examples live in examples. Developer onboarding and release readiness exercise stable curated templates; skill-harvest and recommendation examples remain available through experimental tool contracts.

Validation

Before release:

python3 -m unittest discover -s tests
python3 scripts/release_package_compile.py .
node --check scripts/tmcp_launcher.mjs
python3 scripts/check_contracts.py .
python3 scripts/check_install.py .
python3 scripts/check_release_package.py . --verify-reproducible

The release package check builds from a clean, reviewed Git revision rather than the working directory. It ships only an explicit allowlist of committed files, rejects unsafe paths, symlinks, and secret-like content, then emits a deterministic RELEASE_MANIFEST.json plus archive digest. The CI release gate repeats the build from the same Git tree and requires both digests to match. It also validates frontmatter, hardcoded local paths, private example names, links, extracted-package install shape, doctor, sample harvest, sample workflow recommendation, sample expert rubric planning, composition/runtime/receipt smoke coverage, and stable/experimental workflow labeling. It is a DOI blocker until existing absolute user paths in release evidence docs are redacted or normalized.

DOI-Ready Research Release

TMCP is prepared as an independent software-methods artifact. See RESEARCH_READY.md and docs/release-notes/v0.3.3-doi.md for the citable artifact boundary, validation path, data-availability policy, and claim limits.

References

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选