Getting started
This page takes you from nothing to your first answer out of a namespace. It assumes you have an AI host — codex, Claude Code, or Antigravity — and a namespace route. If you just want to try it, the public worked example is:
https://theorymcp.ai/theorycloud/mcp
Step 1 — Point your host at the route
The MCP server id is theorymcp in every host. Pick yours; the full matrix is on
Connect a host.
codex — in .codex/config.toml:
[mcp_servers.theorymcp]
url = "https://theorymcp.ai/theorycloud/mcp"
oauth_resource = "https://theorymcp.ai/theorycloud/mcp"
scopes = ["mcp:tools", "ai.kb.query"]
Claude Code — in the repo-root .mcp.json:
{
"mcpServers": {
"theorymcp": { "type": "http", "url": "https://theorymcp.ai/theorycloud/mcp" }
}
}
Antigravity — in .agents/mcp_config.json, through the bridge:
{
"mcpServers": {
"theorymcp": { "command": "npx", "args": ["-y", "mcp-remote", "https://theorymcp.ai/theorycloud/mcp"] }
}
}
Step 2 — Authenticate
- codex:
codex mcp login theorymcp(the configured server name) — completes browser OAuth natively. - Claude Code: open
/mcp, choosetheorymcp, authenticate in the browser. - Antigravity: the first connection opens a browser once; the mcp-remote bridge holds the token afterward.
One route, one authorization
OAuth is route-scoped. The token you mint for …/theorycloud/mcp authorizes that route only. If
you later connect an agent endpoint like …/agents/apptheory/mcp, that is a distinct authorization.
Step 3 — Ground before you act
Never assume a namespace’s tools. Paste this into your host:
Ground me in the namespace at https://theorymcp.ai/theorycloud/mcp.
Call describe_interface and summarize the available tools, the visible knowledge bases,
whether agent installs are available, and the route-derived identity. Use only the tool
names describe_interface returns.
You’ll get back the real tool surface for this route, the knowledge bases you can search, and your
resolved client_namespace / endpoint_kind.
Step 4 — Ask the namespace something
Now query its knowledge. The pattern is search first (previews), then fetch detail:
In the theorycloud namespace, list the knowledge bases, then search the most relevant one
for "<your question>". Show me the top results as previews (title + summary + relevance),
then call get_unit on the best match and answer my question from its authoritative body.
Cite the unit_id you used.
That’s the whole loop: connect → authenticate → ground → query.
What next?
- Want continuity across sessions and a named expert? → Use an agent MCP.
- Want an agent’s soul and skills as files in your project? → Integrate an agent.
- Want to publish your own agent into the namespace? → Authoring (gated).
- Want the prompts without the prose? → Copyable prompt library.