# MCP catalogs and agent hosts

> **Upcoming in 1.0 (preview). These pages describe the main branch of LLMSlim Core, which becomes the 0.13 release candidate and then 1.0. `pip install llmslim` installs 0.12.0 today, which doesn't have everything described here.** Docs for what pip installs: https://www.llmslim.app/docs.md

> Read complete tool contracts while keeping execution in your host.

Use LLMSlim's MCP catalog sources, context plans, scoped caching, and the optional OpenAI Agents SDK bridge.

Category: Guides & Strategies · Version: 1.0 · Updated: November 2026 · 4 min read

## Install an optional integration

MCP and Agents extras require Python 3.10 or newer. Base planning and text compression remain available without these SDKs. This page is about reading other MCP servers' tool catalogs; to give an MCP client memory, see the [MCP quickstart](https://www.llmslim.app/docs/1.0/quickstart-mcp).

terminal:

```bash
pip install 'llmslim[mcp]'
# For the optional Agents SDK bridge:
pip install 'llmslim[agents]'
```

## List, plan, and hydrate

Configure a trusted endpoint explicitly. HTTP is allowed only for localhost development; remote endpoints require HTTPS. Stdio sources accept an executable and literal argument list. Plans default to the full catalog.

catalog.py:

```python
import asyncio
from llmslim.mcp import MCPToolCatalogSource, plan_catalog_context, hydrate_plan

async def main():
    source = MCPToolCatalogSource.from_streamable_http(
        'http://127.0.0.1:8000/mcp', source_id='local-tools'
    )
    snapshot = await source.list_tools()
    plan = plan_catalog_context(snapshot)
    schemas = hydrate_plan(snapshot, plan)
    print(len(schemas), plan.metrics.catalog_tokens)

asyncio.run(main())
```

## Freshness and host ownership

The shortest page TTL governs the full snapshot. Private entries are isolated by the caller's cache principal; only explicitly public responses may be shared. Use source.refresh() or source.invalidate() when your host learns a catalog changed. Hydration rejects a plan from a different catalog fingerprint.

> **Selection is not permission:** Never use rankings or fingerprints as authorization. Keep provider credentials and execution policy in the host. Live-model task success is not established by the offline integration benchmark.

## FAQ

### Does this execute MCP tools?

Catalog sources only list tools. The optional Agents bridge requires your async invocation callback; your host remains responsible for authorization, argument validation, and execution.

### Does the default save schema tokens?

No. Full and measure-only plans retain every authoritative schema. Selective exposure is an explicit research-only experiment.

## Related

- [Adaptive Context Planner](https://www.llmslim.app/docs/1.0/adaptive-context-planner.md)
- [MCP quickstart](https://www.llmslim.app/docs/1.0/quickstart-mcp.md)

Canonical page: https://www.llmslim.app/docs/1.0/mcp-catalogs