capdisc

Discovers Claude Code capabilities — skills, agents, plugins, MCP servers, hooks — into typed Pydantic catalogs and an environment report.

Install

uv add claude-code-capabilities
# or
pip install claude-code-capabilities

CLI

capdisc            # scan this machine, write discovery-report.json + .html
capdisc --oauth     # also allow the interactive OAuth flow for HTTP MCP servers
                    # with a pre-registered client (forces a fresh MCP harvest)

Both files are written to ~/.claude/capdisc/ by default. Configure paths and MCP auth via env vars (CAPDISC_ prefix), a .env file, or ~/.claude/capdisc/config.json/config.yaml — see DiscoverySettings in settings.py for every field.

Library

Scan the machine into a typed catalog:

from pathlib import Path
from capdisc.discovery import scan_environment
from capdisc.scope import ScopeRoots

roots = ScopeRoots.discover(start=Path.cwd(), home_dir=Path.home())
catalog = scan_environment(roots)
for entry in catalog.entries:
    ...  # CatalogSkill | CatalogTool | CatalogMcpServer | CatalogPlugin

Build and persist the full environment report (what the CLI does):

from capdisc.report import build_report, write_report

report = build_report(oauth=False)
write_report(report)  # -> ~/.claude/capdisc/discovery-report.{json,html}

Inspect the on-disk scope inventory directly — every skill/agent/command/hook found, its precedence, and which ones actually win a collision:

from capdisc.scope import ScopeInventory, ScopeRoots

inventory = ScopeInventory.scan(ScopeRoots.discover(start=Path.cwd(), home_dir=Path.home()))
print(len(inventory.artifacts), "captured,", len(inventory.effective), "in effect")
for hooks in inventory.hook_configs:
    ...  # HookConfig, unified from settings.json and component frontmatter

MCP servers

Read the last harvested tool inventory (fast, no network — served from a 12h cache):

from capdisc.mcp_harvest import read_mcp_cache, cache_is_stale

servers = read_mcp_cache()  # [] if there's no cache yet
if cache_is_stale():
    ...  # trigger a refresh (below) before trusting this

Force a fresh harvest — connects to every configured server concurrently (bounded) and lists their real tool schemas:

from capdisc.mcp_harvest import refresh_mcp_cache

servers = refresh_mcp_cache(oauth=False)  # also (re)writes the cache
for server in servers:
    print(server.ref, [t.name for t in server.tools])

Bearer/OAuth auth for HTTP servers is opt-in via settings, bound to an exact hostname so a same-named server elsewhere can never receive a credential meant for another:

export CAPDISC_MCP_BEARER_ENV='{"github": {"env": "GH_TOKEN", "host": "api.githubcopilot.com"}}'

report.EnvironmentReport captures the full discovery harvest (scan roots, on-disk inventory, skills, builtin tools, plugins with per-component token cost, MCP servers). mcp_harvest and mcp_catalog enumerate connected MCP servers; plugin_catalog reads installed plugins; scope resolves which roots to scan.

The package ships py.typed.

Tests

uv run pytest
uv run mypy src
uv run ruff check .