ardkit (Python)

Plug-and-play ARD (Agentic Resource Discovery) publisher & registry middleware for Python backends.

ARD lets AI clients discover your agents, MCP servers and skills before they invoke them — instead of every tool description being crammed into the model's context window. ardkit turns any FastAPI / FastMCP / AG2 backend into an ARD publisher (and, optionally, a search registry) in a few lines.

pip install "ardkit-ai[fastapi]"      # import name stays `ardkit`

The PyPI distribution is ardkit-ai (the bare name ardkit was already taken); the import package is ardkit.

Quickstart — one call

If your app already mounts its MCP server(s), publish finds them and exposes the lot — no manual catalog needed:

from fastapi import FastAPI
from ardkit.integrations.fastapi import publish

app = FastAPI()
app.mount("/mcp/billing", billing_mcp.http_app())  # your FastMCP, as usual

publish(app, host="Acme AI", publisher="acme.com", base_url="https://app.acme.com")

publish never mounts anything — it reads the route table, maps each mounted FastMCP back to its instance (identity-based, via fastmcp's app state), introspects its tools into capabilities, and builds the correct public URL from base_url + the mount path. Pass mcps=[billing_mcp, ...] to expose specific servers (unmounted instances are ignored), and introspect=False to expose only those.

That serves:

Or build the catalog by hand

from ardkit import Catalog
from ardkit.integrations.fastapi import mount_ard

catalog = Catalog(host="Acme AI", publisher="acme.com", identifier="did:web:acme.com")
catalog.add_mcp_server(
    name="Billing",
    url="https://acme.com/mcp",
    capabilities=["create_invoice", "list_invoices"],
    representative_queries=["create an invoice", "show unpaid invoices"],
)
mount_ard(app, catalog)

Validate a manifest any time from Python:

from ardkit.validation import validate_manifest

errors = validate_manifest(catalog.to_dict())   # [] when it conforms to ai-catalog v1.0

Catalog as a cacheable document origin

When discovery is published on an apex host but the catalog's source of truth is a backend on another host, expose the manifest as a cacheable JSON document on the backend and have the apex consume it (see ardkit-ts remoteCatalog):

from ardkit.integrations.fastapi import add_catalog_route

add_catalog_route(app, catalog, path="/api/ard/ai-catalog.json")

add_catalog_route (and mount_ard) emit a weak ETag + Cache-Control and answer conditional If-None-Match with 304, so the document is efficiently cacheable by the apex/edge.

Auto-derive entries from your frameworks

from ardkit.adapters import from_mcp_server, from_ag2_agent

# Introspect a FastMCP / mcp-SDK server -> mcp-server-card entry (capabilities = tool names)
catalog.add(from_mcp_server(my_mcp_server, url="https://acme.com/mcp", publisher="acme.com"))

# Map an AG2 / AutoGen agent -> a2a-agent-card entry
catalog.add(from_ag2_agent(my_agent, url="https://acme.com/a2a/agent", publisher="acme.com"))

Optional: be a registry (bring your own search)

ardkit owns the ARD wire contract (request parsing, filters, pagination, federation, error codes, response shaping). You own ranking by implementing SearchProvider — or use the bundled InMemorySearchProvider for tests/demos.

from ardkit.registry import RegistryService
from ardkit.registry.memory import InMemorySearchProvider

service = RegistryService(
    InMemorySearchProvider(catalog.entries),     # swap for your search engine
    source="https://registry.acme.com/ard/",
)
mount_ard(app, catalog, registry=service, registry_prefix="/ard")
# -> POST /ard/search, POST /ard/explore, GET /ard/agents

Implement your own backend by satisfying the protocol:

from ardkit.registry import SearchQuery
from ardkit.registry.models import SearchPage, ScoredEntry

class MySearch:                       # SearchProvider (structural typing)
    async def search(self, query: SearchQuery, *, page_size, page_token) -> SearchPage:
        hits = await my_vector_db.query(query.text, filters=query.filter, k=page_size)
        return SearchPage(items=[ScoredEntry(entry=h.card, score=h.score) for h in hits])

explore and list_agents are optional — omit them and the endpoints return HTTP 501 as the spec allows.

What's in the box

• Spec models (ardkit.models) — Pydantic v2, mirroring ai-catalog.schema.json (URN validation, url XOR data, representativeQueries 2–5, trust manifest).
• Builder (ardkit.Catalog) — mints URNs, sets media types, dynamic catalogs.
• Adapters (ardkit.adapters) — FastMCP / mcp-SDK / AG2.
• FastAPI integration (ardkit.integrations.fastapi) — publish (one call, reads your mounted MCPs), mount_ard, add_catalog_route.
• Registry (ardkit.registry) — BYO-search service + FastAPI router + in-memory provider.
• Discovery helpers (ardkit.discovery) — robots Agentmap, <link>, DNS records.
• Validation (ardkit.validation) — conformance against the vendored JSON Schema.

Install extras

The adapters duck-type framework objects, so you can use them without the matching extra installed.

Development

uv sync --group testing --group lint
uv run pytest
uv run ruff check src tests && uv run ruff format src tests --check
uv run pyright src/ardkit

The ARD schemas under src/ardkit/schemas/ are vendored from ards-project/ard-spec; refresh them with python scripts/vendor_schemas.py.

License

Apache-2.0 — matching the ARD specification. See LICENSE.