mcpsmith

mcpsmith turns installed Model Context Protocol servers into source-grounded, agent-native skill packs.

It resolves the exact artifact behind a configured MCP, snapshots the real source, extracts per-tool evidence, synthesizes grounded skills, runs a second review pass, verifies the result, and can install the skills while removing the MCP config entry.

If an MCP mostly wraps local code or repeatable workflows, mcpsmith gives you something inspectable, versionable, and easier for agents to reuse than a live server dependency.

This repo is the standalone product. distill is historical context only.

Installation

# Homebrew
brew tap nclandrei/tap
brew install nclandrei/tap/mcpsmith

# crates.io
cargo install mcpsmith

The Homebrew formula builds from the tagged GitHub source archive now that the repository is public. cargo install mcpsmith remains the direct crates.io path.

Quick start

Use the installed mcpsmith binary below. Point --config at an MCP config file you already use. If you are running from a checkout, replace mcpsmith with cargo run --quiet --.

tmpdir="$(mktemp -d)"
config="/path/to/mcp.json"

HOME="$tmpdir/home" mcpsmith discover --config "$config"

HOME="$tmpdir/home" mcpsmith playwright \
  --dry-run \
  --config "$config" \
  --skills-dir "$tmpdir/skills" \
  --backend codex

Start with --dry-run. It runs the full pipeline, writes the staged artifacts, and leaves your live MCP config and installed skills untouched.

What mcpsmith does

• Converts the MCP you actually have installed, not just a registry listing.
• Resolves exact local, npm, PyPI, or repository-backed source before conversion.
• Uses runtime tools/list output plus source inspection to ground each tool.
• Prefers deterministic evidence extraction, with a narrow mapper fallback only for low-confidence tools.
• Produces inspectable staged artifacts that other agents and CI jobs can chain without prompts.
• Installs generated skills under ~/.agents/skills/ by default.

How it works

mcpsmith treats MCP replacement as a grounding problem, not a prompt-writing problem. It first proves what server is installed and what source it comes from, then it builds skills from evidence it can cite.

flowchart LR
  A["Discover installed MCP"] --> B["Resolve exact artifact"]
  B --> C["Snapshot source"]
  C --> D["Extract tool evidence"]
  D --> E["Synthesize skills"]
  E --> F["Review with second pass"]
  F --> G["Verify grounding and references"]
  G --> H["Install skills and update config"]

Loading

1. Resolve the MCP to an exact local path, npm package, PyPI package, or repository revision.
2. Snapshot the source for that exact artifact.
3. Inspect runtime tools and map them to handlers, tests, and docs.
4. Synthesize skill drafts from that evidence bundle.
5. Review the drafts with a second agent pass.
6. Verify format, grounding, and references.
7. Optionally install the skills and remove the MCP config entry atomically.

Remote-only or source-unavailable servers are blocked instead of being converted from metadata alone.

One-shot flow

Use one-shot when you want the full pipeline in a single command:

tmpdir="$(mktemp -d)"
config="/path/to/mcp.json"

HOME="$tmpdir/home" mcpsmith run playwright \
  --dry-run \
  --config "$config" \
  --skills-dir "$tmpdir/skills" \
  --backend codex

The positional shorthand is equivalent:

mcpsmith playwright --dry-run --config /tmp/mcp.json --skills-dir /tmp/skills --backend codex

Useful one-shot flags:

• --json for machine-readable output.
• --backend codex|claude to force a synthesis and review backend.
• --backend-auto to allow fallback when a preferred backend is unavailable.
• --skills-dir <PATH> to write preview or installed skills somewhere isolated.
• --config <PATH> to inspect one or more explicit MCP config files.

Staged flow

Use the staged flow when you want inspectable artifacts between steps:

mcpsmith resolve playwright --json --config /tmp/mcp.json
mcpsmith snapshot --json --from-resolve .codex-runtime/stages/resolve-playwright.json
mcpsmith evidence --json --from-snapshot .codex-runtime/stages/snapshot-playwright.json
mcpsmith synthesize --json --from-evidence .codex-runtime/stages/evidence-playwright.json --backend codex
mcpsmith review --json --from-bundle .codex-runtime/stages/synthesize-playwright.json --backend codex
mcpsmith verify --json --from-bundle .codex-runtime/stages/review-playwright.json

Each stage also accepts a server name directly. Staged artifacts are written under .codex-runtime/stages/, and they can be passed between agents with --from-resolve, --from-snapshot, --from-evidence, and --from-bundle.

CLI quick reference

Common entrypoints:

mcpsmith --help
mcpsmith discover --help
mcpsmith resolve --help
mcpsmith run --help

Common commands:

mcpsmith discover --json --config /tmp/mcp.json
mcpsmith playwright --dry-run --config /tmp/mcp.json --skills-dir /tmp/skills
mcpsmith run playwright --json --config /tmp/mcp.json --skills-dir /tmp/skills
mcpsmith resolve playwright --json --config /tmp/mcp.json
mcpsmith snapshot --json --from-resolve .codex-runtime/stages/resolve-playwright.json
mcpsmith evidence --json --from-snapshot .codex-runtime/stages/snapshot-playwright.json
mcpsmith synthesize --json --from-evidence .codex-runtime/stages/evidence-playwright.json --backend codex
mcpsmith review --json --from-bundle .codex-runtime/stages/synthesize-playwright.json --backend codex
mcpsmith verify --json --from-bundle .codex-runtime/stages/review-playwright.json
mcpsmith catalog sync
mcpsmith catalog stats

Catalog and source resolution

mcpsmith has two source inputs:

• Local MCP config entries from discovered config files or explicit --config paths.
• Public catalog data for census work and limited resolution fallback.

Catalog sync defaults to the official and smithery providers.

Resolution order is deterministic:

1. Local path
2. npm package and version
3. PyPI package and version
4. Repository URL and revision
5. Cached catalog fallback only when direct identity is insufficient

That order is intentional. mcpsmith prefers exact source identity over registry prose, and it refuses to generate skills when the source cannot be grounded.

Backend behavior

mcpsmith currently supports codex and claude for the synthesize and review stages.

Backend selection order is:

1. Explicit --backend
2. backend.preference from ~/.mcpsmith/config.yaml
3. Auto-detect installed backends in codex, then claude order

Use --backend-auto when you want the CLI to fall back automatically.

Minimal config example:

backend:
  preference: codex

Use MCPSMITH_CODEX_COMMAND and MCPSMITH_CLAUDE_COMMAND for tests or local command overrides.

Output and artifacts

Generated skills default to ~/.agents/skills/:

~/.agents/skills/
  playwright/
    SKILL.md
    .mcpsmith/
      manifest.json
  playwright--execute/
    SKILL.md

Staged commands write JSON artifacts under .codex-runtime/stages/. One-shot runs also emit a run report under .codex-runtime/runs/ with paths for resolve, snapshot, evidence, synthesis, review, and verify.

Evidence artifacts include:

• The chosen registration and handler snippets for each tool
• Confidence and diagnostics for deterministic matching
• Test and documentation citations when they exist
• Mapper fallback output only for low-confidence tools

run installs reviewed skills and removes the MCP config entry unless --dry-run is set. In --dry-run mode it still writes the full staged artifact set plus a preview skill tree.

Examples

Sample fixtures live under examples/:

• examples/sample-mcp-config.json: minimal MCP config input
• examples/sample-review.json: reviewed skill bundle artifact
• examples/sample-run-report.json: one-shot run report
• examples/sample-skill-pack-tree.txt: installed skill tree preview

Release Automation

Pushing to main triggers the Release workflow, and the same workflow can be run manually with workflow_dispatch.

When the version in Cargo.toml has not been published yet, the workflow will:

• Build release artifacts
• Publish a GitHub release
• Publish mcpsmith-core and mcpsmith to crates.io
• Update nclandrei/homebrew-tap from the tagged GitHub source archive

The workflow creates the v<version> tag automatically. Do not push release tags manually.

Required GitHub Actions secrets:

• CARGO_REGISTRY_TOKEN
• HOMEBREW_TAP_TOKEN

Troubleshooting

• No servers resolved: pass --config /path/to/mcp.json and confirm the file contains an mcpServers object.
• Backend not found: install codex or claude, or pass --backend and --backend-auto explicitly.
• Artifact resolution blocked: inspect the resolve artifact to see whether the server is remote-only or missing exact source identity.
• Synthesis blocked: inspect the evidence artifact for missing handler, test, or documentation citations.
• Review rejected a skill: inspect the review artifact and rerun synthesis with a better backend or better source evidence.
• One-shot blocked: rerun with --dry-run first, inspect the staged artifacts, then retry without --dry-run.

Isolated verification

Never verify against your real home directory. Keep mutable state isolated with a temporary HOME, an explicit --config, and an isolated --skills-dir.

Repo-local checks:

./scripts/local-checks.sh

Smoke helpers in this repo:

• ./scripts/smoke/smoke-test-installed-mcpsmith.sh
• ./scripts/smoke/mock_fixture_flow.sh
• ./scripts/smoke/live_public_mcp.sh --server memory
• ./scripts/smoke/live_public_mcp.sh --server chrome-devtools
• ./scripts/smoke/cli_verify_smoke.sh

For manual verification, the baseline rule is simple: set an isolated HOME, point at an explicit config file, and write generated skills into an isolated directory.

License

MIT