Doc Project

agents-mercato (mct) — Project Documentation

What is agents-mercato?

agents-mercato (mct) is a decentralized package manager for Claude Code agents and skills. It solves the problem of distributing, versioning, and synchronizing Claude agent/skill definitions across teams and projects — without requiring any central registry or infrastructure.

The Problem

Claude Code agents and skills are markdown files (.md / SKILL.md) that live in a project's .claude/ directory. Teams need to:

• Share curated sets of agents and skills across multiple projects
• Keep those definitions in sync when upstream authors publish updates
• Detect when someone locally modifies an installed skill (drift detection)
• Manage dependencies between skills (skill A requires skill B)
• Avoid conflicts when multiple sources provide overlapping definitions

The Solution

mct uses Git repositories as distribution channels (called "markets"). A market is just a Git repo with a conventional directory structure containing agent and skill files. mct clones markets locally, indexes their contents, and lets users install entries into their projects. It then tracks versions, detects drift, resolves dependencies, and synchronizes updates — all through Git.

How It Works End-to-End

1. Register a market: mct market add <git-url> — mct shallow-clones the repo into ~/.cache/mct/<name>/, auto-detects its format (hierarchical or flat), and saves it to ~/.config/mct/config.yml

2. Browse & search: mct search <query> or mct tui — mct builds a BM25 full-text index from all market entries, supports fuzzy matching, filtering by type/market/category, and an interactive TUI with three-panel browsing

3. Install an entry: mct add <market>@<path> — mct reads the file from the cached clone, validates its YAML frontmatter, recursively resolves skill dependencies (with cycle detection), injects tracking metadata (mct_ref, mct_version, mct_checksum, etc.), writes files to .claude/, and records the installation in ~/.cache/mct/installed.json

4. Check status: mct check — for every installed entry, mct computes its state by comparing the local file hash (xxhash) against both the recorded version and the latest upstream version. States: clean, update_available, drift, locally_deleted, update_and_drift, deleted, orphaned

5. Sync updates: mct sync (or mct refresh + mct update separately) — fetches from all markets, prunes stale install locations and upstream-removed files, then offers to restore any locally-deleted file interactively ([r]/[k]/[a]/[n] prompt; --restore-all / --restore-none flags; CI mode skips silently), and finally applies updates. Drift conflicts are handled via policy flags (--all-keep, --all-merge)

6. Audit health: mct doctor — read-only, offline. Surfaces every issue in one shot: drift, locally-deleted files, stale locations, upstream-removed files, orphaned packages. JSON output via --json for CI tooling. Run mct refresh first if you need a fresh upstream view.

7. Export/restore: mct save writes a .mct.json manifest; mct restore re-installs from it. Combined with optional Git hooks (mct hook install), this enables fully automatic save-on-push and restore-on-pull workflows

Market Format

Hierarchical (full market):

market-repo/
  profile-a/category-x/
    README.md              ← profile metadata (tags, description)
    agents/agent1.md       ← agent definition
    skills/skill1/SKILL.md ← skill definition (+ supporting files)
  profile-b/category-y/
    README.md
    agents/agent2.md

Flat (skills-only repo):

repo/
  skills/
    skill1/SKILL.md
    skill2/SKILL.md

Format is auto-detected on market add. The first two path segments form a "profile" that groups related entries.

Entry Type Inference

Determined by path convention, not declaration:

• .md file under agents/ → Agent (installed to .claude/agents/<name>.md)
• SKILL.md under skills/<name>/ → Skill (installed to .claude/skills/<name>/)
• .md file under commands/ → Command (installed to .claude/commands/<name>.md)
• .json file under hooks/ → Hook (merged into .claude/settings.json)
• Everything else → Ignored by mct

Commands are flat single-file entries (like agents), Claude-Code-specific — other transform targets (codex, gemini, etc.) skip them. Commands support requires_skills in frontmatter, resolved via the same dependency engine as skills.

Hooks are JSON snippets, NOT files copied to disk. Install merges the snippet's hooks[] array into .claude/settings.json under hooks[<event>] and injects an mct_id field (xxhash of the ref) on every hook object so removal can target the right entries. Drift detection compares the canonical body checksum (mct_id excluded) against the recorded value, so user reformatting doesn't trigger drift but command edits do. Two markets providing hooks with the same (event, matcher) are blocked at install with ErrConflictHookEventMatcher. Note: mct hook install (the existing CLI command) installs git hooks for save/restore automation — unrelated to the new hook entry type.

Dependency Resolution

Skills can declare dependencies in their frontmatter:

requires_skills:
  - file: skills/foo/SKILL.md              # same-market dep
  - file: skills/bar/SKILL.md
    market: https://github.com/org/other.git  # cross-market dep
    pin: abc123                               # optional version pin

mct resolves dependencies recursively with cycle detection. Cross-market deps auto-register the foreign market if not already known (with user confirmation in interactive mode).

Drift Detection

After installation, mct tracks each file's content hash (xxhash64). On mct check, it compares:

• Local file content vs. original content at the recorded git version
• If they differ → the user has locally modified the file → marked as "drift"

This lets users customize installed skills while still being warned about upstream updates that may conflict.

Conflict Detection

Two types:

1. Ref collision — Same filename exists in multiple markets (ambiguous install)
2. Version mismatch — Same skill required at different versions by different agents

Project Isolation

Files are copied, not symlinked. Each project gets its own independent copy in .claude/. The install database (installed.json) tracks which projects have which entries, enabling multi-project cleanup via CleanStaleLocations().

Tech Stack

• Go 1.26 with Cobra (CLI), Charmbracelet (TUI), go-git (Git ops)
• BM25 + Snowball for full-text search, xxhash for drift detection, YAML for config

Architecture

Hexagonal (Ports & Adapters):

cmd/mct/main.go          → entry point, wires DI
inbound/commands/         → Cobra CLI handlers
inbound/queries/tui/      → Bubble Tea interactive UI
app/                      → business logic (implements all service interfaces)
domain/                   → types, errors, frontmatter parsing, state
domain/service/           → port interfaces (queries + commands)
domain/repositories/      → adapter interfaces (git, fs, config, state, installdb, tx)
outbound/gitadapter/      → go-git implementation
outbound/fsadapter/       → OS filesystem implementation
outbound/cfgadapter/      → YAML config, JSON state, install DB with file locking
outbound/txadapter/       → disk-backed transactional staging for atomic install/update/remove

Atomicity (transactional install/update/remove)

Every mct write path (Add, Update, Remove, Prune) opens a per-package tx.Tx from the tx.Manager port. All file writes, deletes, and the install database save are buffered in a per-operation staging directory under <cacheDir>/staging/<txID>/. On Commit the manager promotes everything atomically (renames where same-fs; copy+remove cross-fs); on Rollback or process crash nothing on disk has changed. A startup recovery pass replays any leftover committing staging dirs from a previous crashed run.

Package Index

Key Concepts

• Market — A Git repo containing agent/skill definitions in a hierarchical or flat layout
• Entry — An agent (.md under agents/), skill (SKILL.md under skills/<name>/), command (.md under commands/), or hook (.json under hooks/, merged into .claude/settings.json)
• Profile — First two path segments group entries (e.g., dev/go)
• MctRef — Canonical reference in market@path format
• Drift — Local modifications detected via xxhash comparison against upstream
• Sync — Fetch upstream changes, compute state (clean/update/drift/deleted), apply updates
• Dependency resolution — Skills can require other skills, even cross-market, with cycle detection
• Conflict detection — Filename collisions across markets, version mismatches on shared deps
• Git hooks — Optional post-merge (mct restore) and pre-push (mct save) automation

Data Flow

User runs `mct add market@dev/go/skills/foo`
  → CLI parses ref → App.Add()
    → GitRepo.ReadFileAtRef() reads from cached clone
    → Frontmatter.Parse() validates YAML header
    → Dependency resolution (recursive, cycle-safe)
    → InjectMctFields() adds tracking metadata
    → Filesystem.WriteFile() writes to .claude/
    → InstallDB.Save() records in installed.json

Configuration

• Config: ~/.config/mct/config.yml — markets, policies (conflict, drift), local path, SSH
• Cache: ~/.cache/mct/ — shallow clones, installed.json, sync-state.json
• Env vars: MCT_SSH_ENABLED, MCT_LOCAL_PATH, MCT_CONFLICT_POLICY, MCT_DRIFT_POLICY