Get started with Insights
Create an insight in the dashboard or from MCP, refine the objective, run it, and work the feed.
In short
Pick agents you can read, write what is worth surfacing in a sentence, click Refine, then Run now. The feed is Insights, then Feed. The same steps exist as MCP tools for the agents you already run.
Insights need to be switched on for your account, and the agents you name need documents, claims and learnings already on file. Empty agents produce an empty feed.
In the dashboard
- Open Insights, then Rules, then New insight.
- Give it a Name people will recognize in the feed, such as
Proactive outage outreach. - Under Agents to consult, search by name and tick the specialists that hold the facts. Only agents you can read are listed. You can pick up to eight.
- In What is worth surfacing?, write the standing question in plain language. Say who to look at, what signal matters, and what you want to do about it. A first draft can be rough:
customers annoyed by outages, reach out. - Set About each to the kind of subject (
customer,account,contract,caseordeal). - Set Run to When I click Run, or From the API if your own systems will start it.
- Click Refine. Hyperstruck rewrites the draft into a precise objective (the signal, the window, the action) and checks whether the agents you picked likely hold the data. Undo puts your wording back. A warning here is cheaper than a wasted run.
- Click Create insight, then Run now.
The builder is that form: Name, What is worth surfacing?, Refine, About each, and Run. Refine stays quiet until you pick at least one agent. Agents to consult lists only the agents you can read. Search by name and tick up to eight.
Paste a vague sentence, pick the specialists, and Refine. Paste a question about product usage and logins against support agents and Refine again: the notice should say they probably cannot answer it.
A typical Refine pair:
| You type | Agents | What Refine does |
|---|---|---|
customers annoyed by outages, reach out | Support desk and network operations | Sharpens it into a churn-risk outreach with a time window |
which accounts have falling product usage and fewer logins so we can save them | The same support agents | Warns: those agents do not hold usage or login data |
new companies setting up in australia | Market signals and the GTM desk | Sharpens it into a new-entity watch with roles and a time window |
Watch the run on the insight's page. When it finishes, open Insights, then Feed. Work from the top: read Why now, open the evidence, send or adapt the draft from your own tools, then Mark actioned or Dismiss.
Insights, then Rules is every standing question. On this account Candidate rediscovery has succeeded with three open findings; GTM Intelligence has never run.
Each card is the objective, the agents it asks, and the latest run. Run now starts a fresh look; Edit changes the objective or the agents.
From MCP and your own agents
Connect the hosted MCP server as in MCP server. When Insights are on for the account, the host also sees the insight tools. Your assistant in Claude Code, or any other MCP host, can then list, create, run and update them with the same access rules as the dashboard.
| Tool | When to use it |
|---|---|
list_insights | See the insights whose agents you can all read, with latest run and open finding count |
get_insight | One insight, same summary |
refine_insight | Sharpen a draft objective and judge whether the chosen agents likely hold the data. Saves nothing |
create_insight | Save a new insight. Saving does not start a run |
update_insight | Change name, objective, agents, or enablement. Needs to be the creator, or have write access to every agent it names |
run_insight | Start a run now. Needs agents:execute. Estimated cost is held against the account spend limit before any agent is asked |
list_insight_runs | Recent runs of one insight |
get_insight_run_events | Poll a run until it succeeded or failed |
list_insight_findings | The feed: highest priority first, with evidence. Filter by insight, run or status |
update_insight_finding | Mark a finding actioned, dismissed, or new to reopen it |
delete_insight | Remove the insight and its findings. Refused while a run is in progress |
A host that should start a run every morning calls run_insight, waits on get_insight_run_events, then reads list_insight_findings. That is the same cadence as a dashboard Run now, without anyone opening the page.
Creating an insight from MCP:
Name: GTM Intelligence
About each: company
Agents: the market-signals agent and the GTM desk (ids from list_agents)
Objective: Find overseas and US vendors that are building or expanding
go-to-market teams in Australia and New Zealand, and tell the GTM desk
which sales roles are coming, which candidates in our network fit them,
and the warmest way in.Call refine_insight with that draft and those agent ids first. Then create_insight, then run_insight.
These tools need the same scopes as the dashboard: agents:read and claims:read to list and inspect, agents:write to create or change an insight or a finding's status, agents:execute to run. Access control is the reference.