Workflows with your agents

Enterprise teams are already willing to work in assistants they chose. Hyperstruck's job is to make those assistants apply the same expertise the rest of the company invested in, instead of each person prompting from memory.

A workflow is the named, judged version of that expertise: first-line support, triage, candidate matching. The agent in the editor is still the one talking to the human. Hyperstruck is how it knows which workflow this is, what the company already knows, and how the answer should be judged.

What this does not replace

• The agent the person is talking to. Claude Code (or any other MCP host) still plans the turn, reads the file they dropped, and writes the reply they see.
• Their other tools. Search, tickets, the file system, the browser: still theirs.
• A new place they have to live. The dashboard is where you design the workflow, watch the run log, and correct a verdict. The work itself happens in the tool they already opened.

What Hyperstruck adds is organization-wide: the catalog of workflows, the claims behind a brief, the learnings from earlier runs, and evals that judge the answer the same way every time.

Connect the host

Add Hyperstruck as a remote MCP server, the same way you would for learnings and claims. The address is https://mcp.hyperstruck.com/mcp/. An API key is one way in. It is not the only way.

• API key. A host that can send an Authorization header uses a Hyperstruck API key. MCP server has the Claude Code command.
• OAuth 2 with your company's SSO. Hosts that support remote OAuth sign the person in through the identity provider the organization already uses. There is no key to copy.

Either way, the connection inherits whatever those credentials are already allowed to do. It does not get a separate, broader grant. Your IT administrators and security team control that access: who can sign in, which agents they can read, and which scopes they hold. A person who cannot read an agent in the dashboard cannot read it through the workflow tools.

Using an editor such as Cursor as an MCP client is that same connection. It is not the Claude Code hook, and it does not require the hook. The agent in that editor gets the workflow tools, then finds and runs workflows like any other MCP host.

Accounts with Workflows turned on see six extra tools on that server.

The six workflow tools

The agent's own instructions already tell it to resolve at the start of a task and complete_run at the end. Workflows sit beside that loop, they do not replace it. list_workflows with a goal is how it chooses which saved way of working this task falls under; resolve is still how it loads learnings and claims for the rest of the turn.

A turn in practice

Someone is in Claude Code, with Hyperstruck connected.

1. They describe the work in their own words, and attach whatever they have (a ticket, a CV, a matter).
2. The agent calls list_workflows with that goal. The catalog is written so an agent that has read no docs can still send a valid body.
3. If include_context is on, the top matches include what those workflow agents learned last time: the same kind of advice resolve would offer, scoped to this workflow.
4. The agent fills the input from the attachment and from claims it looks up (a hiring brief, an account's plan). It does not invent numbers the company never filed.
5. It calls run_workflow, waits for the answer, and reports the verdicts as Hyperstruck returned them. It does not average them, re-score them, or substitute its own judgment for an eval.
6. A person who wants the case file opens the run in the dashboard.

The examples page walks this as a recruiter matching a CV to a brief. A turn looks like this (no new window, no extra app):

You: Check this CV against Juniper Labs' Founding CTO role.

Agent: I'll find the matching workflow and fill it from the CV and the brief on file.

(calls list_workflows with that goal)
Picked candidate-company-match.

(calls run_workflow)

Must-have skills: met
Seniority: above
Overall match: 2.68 / 4
Reasons: hands-on applied AI, still writes code, salary inside the band.
Gaps: founding-level CTO ownership should be checked.

The agent reported the verdicts Hyperstruck returned. It did not re-score them.

Drafting a workflow from the agent

You do not have to start in the builder. From the same host:

Draft a workflow that decides whether a customer email is a refund request and how urgent it is, then drafts a reply.

The agent calls draft_workflow with that goal and, if it has one, a sample email. It gets back a definition, suggested evals (and why), and any warnings. It should show you the draft, then save_workflow when you agree. Starter evals the account does not have yet can be installed on save.

The dashboard's Describe a workflow panel is the same compose step, for people who prefer to review a graph on a canvas before it exists.

What the credentials can do

An API key is limited to the scopes on that key. A person signed in with SSO is limited to the role and the agents your administrators assigned them. For workflows specifically:

MCP server lists the scopes for resolve, claims, and the rest. Use one key with the union of what you actually call.

Related pages