Askit build mcp

Skill product-on-purpose/agent-skills-toolkit/skills/askit-build-mcp

Toolkit and Standard for building, grading, and scaling cross-agent skill libraries (Claude Code + Codex) to a tiered Bronze/Silver/Gold quality bar.

Install

npx -y skills add product-on-purpose/agent-skills-toolkit --skill askit-build-mcp

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Creates and improves MCP server definitions (a portable .mcp.json) for a plugin to the Advanced Skill Library Standard. Use when you need to add an MCP server to a plugin, author or extend .mcp.json, or wire the per-target mcpServers manifest pointer.

SKILL.md

2.8 KB, as published. Nobody here has run it

askit-build-mcp

Purpose

Author the plugin's MCP servers as one portable .mcp.json at the plugin root (the standard { "mcpServers": { ... } } format) and wire the per-target mcpServers manifest pointer. Two modes: create adds a server; improve fixes findings. MCP is Universal-tier: the server definition is portable and both Claude and Codex ingest a bundled .mcp.json referenced by the native manifest's mcpServers pointer (Standard sec 3.9). Authoring depth is in references/authoring-mcp.md. Follows the shared builder contract (../../docs/reference/builder-pattern.md).

When to use

When the user asks to add, scaffold, author, or improve an MCP server (tools, resources, or prompts via the Model Context Protocol) for a plugin.

create mode

Brief interview: server name (kebab-case), transport (stdio or http), the command plus args (stdio) or the url (http), and any secrets as an env-var NAME (never an inline value). Skip the interview if these are already in context.
Author or extend the root .mcp.json: add the server under mcpServers. Copy templates/mcp.json if the file does not exist yet.
Register the server in library.json components.mcpServers as { name, version, tier, status }.
Wire the per-target pointer: run node scripts/generators/gen-manifest.mjs <plugin-root> --write --target=all. Each native manifest gains "mcpServers": "./.mcp.json".
Assess with node scripts/evaluate.mjs . --json and iterate to 0 errors (mcp-valid, the S3 components index, and the S6 per-target pointer).

improve mode

Run node scripts/evaluate.mjs . --json and read the report.
For each finding: a malformed server (mcp-valid) -> add the missing command or url; a committed secret (mcp-valid) -> replace the inline value with an env-var reference or bearer_token_env_var; a missing pointer (S6) -> regenerate the manifests; an undeclared server (S3) -> add it to components.mcpServers. For any other finding, read its message and apply the fix it states.
Re-run evaluate to confirm.

Scope

The root .mcp.json is the single authored source; the manifest pointers are generated, so never hand-edit a generated manifest. Secrets MUST be referenced from the environment, never committed (Standard sec 9). MCP is the one plugin-distributable component both agents ingest, and emission is just the manifest-pointer wiring (no per-type render engine, per the builder pattern).