MCP endpoint
Every Wavetable workspace exposes its own MCP (Model Context Protocol) endpoint. Add it to Claude, to a partner’s agent, or to your own scripts, and that agent can work with your business graph within the policies you approved.
This is a deliberate strategic posture: be the home, not the competitor. Wavetable does not try to be the only agent you use. It tries to be the place every agent you use comes back to, because that is where the data and the artifacts live.
Tools
The endpoint exposes a small, typed tool surface:
| Tool | What it does |
|---|---|
query |
Read-only queries against projections, attribute-scoped by policy |
memory.read |
Read an entity’s cited narrative |
draft.create |
Create a draft (message, note) for human review |
intent.propose |
Propose a new workflow, which lands as an Inbox approval |
task.create |
Create a task for a human |
Note what is absent: no external sends, no claim acceptance, no direct writes. External agents propose, exactly like Wavetable’s own edge agent. The blast radius is structural, not a prompt asking the model to behave.
Auth and scope
- Access uses OAuth 2.1 with capability tokens minted by your approved policy artifacts.
- Tokens are scoped (which entities, which attributes, which tools) and revocable.
- Policy changes are artifacts, so widening an agent’s access is a diff someone approves, with history.
Audit
Every MCP call is logged as an event in your log, same as an email arriving or a human approving a merge. “What did that agent read last week?” is a query.