# MCP: ChatGPT and Claude

Connect once; read the current guide and action catalog from the server.

Canonical: https://docs.cyberrxn.com/en/docs/mcp

## Connection

Use `https://mcp.cyberrxn.com/mcp` as the remote Streamable HTTP MCP endpoint. Choose OAuth and authorize the intended CyberRXN workspace. Existing users can use the dashboard's [connection instructions](https://app.cyberrxn.com/en). No API key needs to be pasted into an OAuth connector.

1. In ChatGPT web, open **Plugins → + → Add custom MCP server**, enter a name and the endpoint, choose OAuth and complete the provider’s creation steps. Legacy workspaces may use **Settings → Apps → Create**.
2. In Claude, open **Customize → Connectors**, add a custom remote connector and enter the endpoint.
3. Complete CyberRXN OAuth consent and confirm the workspace.
4. In a fresh conversation, make the connector available for the message: select the ChatGPT plugin, or enable it under **+ → Connectors** in Claude. Then ask a chemistry question and inspect the real tool activity.

ChatGPT’s account plan and workspace permissions determine which custom MCP operations are available. App selection can apply to one message; a follow-up needing fresh data may need another selection. Claude’s connector availability and organization roles also vary. Once tools are available, the chemistry prompt can omit the CyberRXN name; selection by the model remains probabilistic.

Provider interface names, account eligibility and confirmation rules can change. The website button opens the provider and copies the endpoint; it cannot guarantee a prefilled connector form when the provider supplies no supported deep link. [Official ChatGPT guidance](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) · [Official Claude guidance](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp).

## A current guide, without a local skill file

Initialization instructions tell the agent to call `get_reaction_skill` before chemistry work. It returns the current Markdown guide, its digest and a live action catalog with input schemas. The same guide is available as the MCP resource `cyberrxn://skills/reaction-evidence` and as [public Markdown](https://mcp.cyberrxn.com/skills/reaction-evidence.md).

This is server-provided guidance. Connecting bare MCP does not install a native filesystem skill. Updating the hosted guide does not require the user to upload a ZIP.

## Stable action entry points

Some hosts freeze approved tool definitions. After they discover the stable tools once, agents can use a current action from the live guide through:

```json
{"name":"run_chemistry_action","arguments":{"action":"validate_reaction","arguments":{"reaction_smiles":"CCO>>CC=O","mode":"strict"}}}
```

`run_chemistry_action` accepts only supported read operations. `submit_corpus_action` accepts the supported non-destructive report/contribution writes. Unknown actions and recursive dispatch are rejected. Normal chemistry actions retain normal billing under the actual operation name; wrapping a call does not make it free.

When a direct tool is exposed, the agent can use it normally. New tool schemas may need an initial provider-side refresh. Server-side guide and supported-action updates remain available through the stable entry points. No server can force a model to select a tool.

## Expected behavior

For a recorded precedent: guide → bounded search → full observation → source-aware answer. For a supplied equation: guide → validation/balance/conversion. No result is not evidence of a corpus error. Reports require a directly observed source conflict. Contributions require public source evidence and sharing rights.

Disconnect the provider in [Applications](https://app.cyberrxn.com/en/connections) to revoke access. External agents cannot approve their own reports or contributions.