What is the Whet MCP?
Model Context Protocol (MCP) is an open protocol that lets host applications (Claude Desktop, Cursor, custom agents) discover and execute tools exposed by external servers. Whet exposes its own MCP server so any compatible agent can operate the workbench with the same permissions as a human.
Why does it matter to us?
Whet is designed agent-native by default: every action a human takes in the UI has parity in CLI and MCP. That means an agent can:
- Create pipelines from an intent ("follow this competitor and generate daily drafts").
- Read the inbox of pending artifacts.
- Refine drafts (shorter, change tone, match voice).
- Publish with explicit confirmation.
- Diagnose a pipeline that is not producing what was expected.
- Drive a drafting session end-to-end: snapshot upstream items, attach variants, generate, export.
- Read and edit analytics dashboards as living artifacts.
Without looking at the UI, without re-implementing the workbench, without bypassing the contracts.
Architecture
The MCP server runs inside the Whet Desktop binary itself — there is no separate process to start. When Whet is running, the MCP server is running.
- Transport: WebStandard Streamable HTTP (stateless per-request).
- Endpoint:
POST http://127.0.0.1:7891/mcp(alsoGETandDELETE). - Auth: Bearer token in
Authorizationheader. Same token used by the CLI. - SDK:
@modelcontextprotocol/sdkv1.29.
Each request creates a fresh McpServer instance. There is no shared
session — state lives in the binary's local database.
What it exposes
| Surface | Count | Details |
|---|---|---|
| Core tools | 8 | start_draft, read_inbox, list_posts, diagnose_pipeline, create_pipeline_from_intent, publish_artifact, connect_account, fetch_url. See Tools. |
| Sessions tools | 18 | Lifecycle + selection + variants + exports + presets for the drafting canvas. See Tools and Sessions concept. |
| Dashboards tools | 9 | Read/write of dashboards and widgets, with dataset catalog and widget execution. See Tools. |
| Prompts | 2 | Guide templates: draft_from_handle, refine_post. See Prompts and resources. |
| Resources | 2 | Accessible URIs: pipeline://{id}, artifact://{id}. See Prompts and resources. |
Get started
Choose your client:
- Claude Desktop
- Cursor
- Any other MCP client supports Streamable HTTP — use the same config.
Permission model
Today the agent token has full workspace scope (MVP). Granular scopes (org:read, artifacts:publish, etc.) are on the roadmap. In the meantime, the operational rule is manual: if you don't want an agent to publish, don't give it a token; use it in flows where the decision to publish goes through a human confirming.
Whet does not auto-publish. Not from MCP, not from the CLI, not from the UI. Every publish_artifact represents an explicit decision. This is by design, not a toggle.