MCP server
Connect Hyperstruck learning to any MCP-capable host, like Claude Code or Cursor, so your agent gets sharper at the work it repeats.
In one line
Point an MCP host at Hyperstruck and your agent pulls the lessons from its prior runs before it starts, then reports back when it finishes, so it improves at recurring tasks without any integration code.
Hyperstruck runs a hosted MCP server at https://mcp.hyperstruck.com/mcp/. It is the lowest-friction way to give a host that already speaks MCP, like Claude Code or Cursor, access to Hyperstruck's learning. There is no package to install and no SDK to wire in. You add one remote server with your API key and the host gains two tools.
The two tools
The server is deliberately small. It exposes the learning loop and nothing else.
| Tool | When to call it | What it does |
|---|---|---|
resolve | At the start of a task, with the goal | Returns a run_id and any learnings from prior runs that are relevant to this goal, for the agent to follow. |
complete_run | At the end of the task, with the run_id | Records whether it succeeded, credits the learnings that helped, and captures new ones from how the run actually went. |
That pairing is the whole idea. The agent resolves before it acts so it starts from what worked last time, and completes the run after so the next attempt at a similar task is better than this one. The host decides when to call each tool, and the server's own instructions remind it to.
Connect from Claude Code
Add it as a remote streamable-HTTP server authenticated with your Hyperstruck API key:
claude mcp add --transport http hyperstruck \
https://mcp.hyperstruck.com/mcp/ \
--header "Authorization: Bearer <YOUR_HYPERSTRUCK_API_KEY>" \
--scope user--scope user makes it available in every project; use --scope local for just the current one. Check it connected with claude mcp list, and run /mcp inside a session to see the resolve and complete_run tools.
Connect from other hosts
Any host that supports remote MCP servers with a bearer auth header can use the same endpoint. Point it at https://mcp.hyperstruck.com/mcp/, set the Authorization: Bearer <YOUR_HYPERSTRUCK_API_KEY> header, and the two tools appear. Hosts that cannot attach a custom auth header to a remote server are not supported directly today.
Authentication
The server authenticates with a standard Hyperstruck API key, the same key you use for the REST API. The key needs an active subscription and scopes that cover the tools you call.
| Requirement | Scope |
|---|---|
| Connect to the MCP endpoint | agents:read or agents:write |
Call resolve | agents:read |
Call complete_run | agents:write |
For the full learning loop (resolve before a task, complete after), use a key with agents:write — it implies agents:read. Fine-grained authorization applies the same way as REST: the key's space reach controls which agents and learnings it can access.
See Access control (FGA and RBAC) for scope details, key tiers, and space selection.
The hosted server forwards your key to Hyperstruck, which re-authenticates every call, so tenant isolation and fine-grained access checks apply as everywhere else.
How it fits with the rest of Hyperstruck
The MCP server is another front door onto the same hosted learning system, not a separate one. A complete_run writes through the same path as a reasoning run or a skill, and a resolve reads the same store. Learnings captured one way are available every other way. If you would rather invoke Hyperstruck from inside your IDE through a slash command, the Claude Code and Cursor skills cover that; if you are building your own integration, the REST API and framework integrations are the lower-level options.
Related pages
- Access control (FGA and RBAC) for MCP scope requirements and API key space reach
- Agents API for the REST alternative
- Claude Code and Cursor for IDE skills
Give it time to learn
A learning is extracted in the background after complete_run, so a brand-new lesson is not retrievable the instant a run finishes. By the time the agent picks up a similar task later, the lesson is in place and resolve surfaces it.