toolanything-platform-ops

name	toolanything-platform-ops
description	適用於使用者要求安裝內附 ToolAnything wheel、同步 Codex/OpenClaw/Claude 本地 skills、更新 AGENTS 規則、管理 shared custom-tool server、固定 port 或設定 auto-start，且 `toolanything-mcp-router` 已先判定應採用 ToolAnything。It is the heavy operational skill for bundled installs, canonical server governance, and host rollout; it is not the first routing entrypoint for simple tool wrapping.
version	2026.3.17
license	MIT
metadata	{"author":"OpenAI Codex","repo":"ToolAnything"}

ToolAnything Platform Ops

This is the heavy operational skill. Do not load it first for ordinary "wrap a function as a tool" work. toolanything-mcp-router should own the initial routing decision, then hand off here only when the task needs bundle installation, host rollout, AGENTS synchronization, shared-server policy, or auto-start.

Primary Responsibilities

• install or refresh ToolAnything from the bundled local wheel
• sync the local skills or Claude subagents for Codex, OpenClaw, and Claude Code
• inject the ToolAnything decision rule into the correct AGENTS.md
• keep reusable custom tools inside one canonical shared server
• enforce fixed host and port policy, plus platform-specific auto-start

First Action

Run the bundled installer from this skill folder:

python scripts/install_local_bundle.py --host auto

If multiple hosts are detected, stop and rerun with an explicit host:

python scripts/install_local_bundle.py --host codex
python scripts/install_local_bundle.py --host openclaw
python scripts/install_local_bundle.py --host claude-code

The installer must:

1. pick the newest toolanything-*.whl from wheels/, then legacy wheel/, then repo dist/
2. force-reinstall the local ToolAnything package from that wheel
3. sync the bundle skills to the correct host-specific locations
4. replace any legacy fallback AGENTS rule with the new router-first decision rule

If the wheel is missing, stop and report a bundle packaging gap. Do not send the user back to GitHub install docs.

Bundle Layout

This operational bundle deploys three skill entrypoints together:

• toolanything-mcp-router: thin router and quick path
• toolanything-platform-ops: heavy host and server operations
• toolanything-tool-wrapper: legacy compatibility alias

Host Targets

Codex

• skills path: $CODEX_HOME/skills/ or ~/.codex/skills/
• agents file: $CODEX_HOME/AGENTS.md or ~/.codex/AGENTS.md

OpenClaw

• skills path: $OPENCLAW_WORKSPACE/skills/ or ~/.openclaw/workspace/skills/
• agents file: $OPENCLAW_WORKSPACE/AGENTS.md or ~/.openclaw/workspace/AGENTS.md

Claude Code

• local skill path: ~/.claude/skills/*.md
• agents file: ~/.claude/AGENTS.md

Do not collapse these hosts into one directory model. If the user suggests the wrong path, correct them directly and continue with the correct target.

Shared Custom Tool Server Policy

When the task is to add a reusable custom tool for an agent host, do not create one long-lived MCP server per tool.

Use one canonical server root per machine:

• Windows: %USERPROFILE%\\.toolanything\\agent-mcp\\
• macOS / Linux: ~/.toolanything/agent-mcp/

Expected structure:

• server.py
• tools/
• toolanything-server.json

Default runtime policy:

• host: 127.0.0.1
• port: 9092
• transport: streamable-http

Only allow a second shared server when there is a clearly different security boundary, Python runtime, or deployment owner.

Workflow

Phase 1. Install and sync

1. Run the local bundle installer.
2. Confirm the installed ToolAnything import path:

python -c "import toolanything; print(toolanything.__file__)"

3. Confirm the correct skills or Claude subagents were written to the host-specific location.
4. Confirm the updated AGENTS.md contains the new router-first decision rule and no duplicate legacy block.

Phase 2. Decide the implementation path

1. Stable Python callable: prefer @tool(...).
2. HTTP, SQL, or model source: prefer the source-based APIs.
3. Shared custom tool server need: integrate into the canonical server instead of creating another long-lived server.

Phase 3. Implement without spreading policy drift

1. Keep tool contracts explicit with name and description.
2. Reuse one tool module across MCP, OpenAI, and CLI where possible.
3. If you add a tool to the shared server, place it under tools/ and load it from the shared server.py.
4. Do not invent a second schema layer or a second long-lived server just to get one tool running.

Phase 4. Verify

Run the cheapest useful checks first:

toolanything doctor --mode http --url http://127.0.0.1:9092
toolanything inspect

Also verify:

• http://127.0.0.1:9092/health
• /mcp can answer tools/list
• the new tool is callable and lives inside the shared server

Boundaries

• This skill does not decide whether ToolAnything or FastMCP should win first. That belongs to toolanything-mcp-router.
• This skill does not replace product architecture or orchestration decisions.
• This skill should not expand into unrelated repo refactors outside the tool layer or shared-server policy.

Delivery Checklist

Always report:

1. detected host and actual install path
2. selected wheel and its source folder
3. which AGENTS.md was updated and whether a legacy block was replaced
4. whether you chose @tool(...) or a source-based API, plus why
5. whether a shared custom tool server was touched, including root and port
6. which validations ran, or what blocked them