Agents
What an AI agent can do in your Datahop project, how to connect one, and how its work is tracked.
An agent connected to Datahop gets the same controls you have. It can learn your events, run any analysis, save insights and dashboards, build cohorts, create and change feature flags, run experiments from idea to decision, triage feedback, and leave notes for the team.
The fastest start: paste this into Claude, ChatGPT or your coding agent. It connects over MCP or the CLI and gets oriented in your project without changing anything.
Connect
- Create an API key in Settings → API keys & MCP. Start with Read only if you want to watch first.
- Connect your agent:
| Option | Best for |
|---|---|
| MCP | Claude, ChatGPT, Cursor and other MCP clients |
| CLI | Coding agents with a shell, scripts and CI |
| REST API | Anything that speaks HTTP |
All three can do the same things, with the same permissions. To brief an agent quickly, point it at llms.txt.
What agents can do
| Area | Examples |
|---|---|
| Orient | Get an overview: products that are on, top events with descriptions, cohorts, running experiments, recent changes. |
| Learn your data | List events, see their properties and common values, write event descriptions. |
| Analyse | Trends, funnels, retention, raw events, web analytics and read-only SQL. |
| Share | Save insights, build and arrange dashboards, read a whole dashboard at once, share a public link (after asking you). |
| People | Find people and companies. Create cohorts. |
| Release | Create, change, switch on and off feature flags. Check what a user would get. |
| Experiment | Draft from a description, create, launch, read results, stop and conclude. |
| Feedback | Read, summarise and triage feedback. |
| Leave context | Add annotations to every chart. Read what people and other agents changed. |
The tool reference lists all 62 tools. Agents can't change project settings, create API keys, rotate the project token or delete the project.
Results agents can read
Every chart and experiment result comes with a short plain-text summary of the numbers that matter:
signed_up 1,759 → project_created 625 (35.5%) → invite_sent 194 (31.0%)
Overall conversion 11.0%. Biggest drop-off: project_created → invite_sent (69.0% lost).The app shows the same text under each chart as Summary · what agents see, so you can always check what your agent was looking at.
Every change is tracked
Everything an agent changes appears under Changes on the Home page, with its key's name, such as 🔑 claude-growth-agent launched experiment "Annual plan discount banner". Insights, dashboards, cohorts, flags and experiments it creates show the key as their creator.
Give each agent or job its own key with a clear name, so you can tell who did what. Agents can read this log too, so the next agent sees what was already tried.
Prompts that work
- "Look at signup → upgrade over the last 30 days. Find the segment with the biggest gap, propose one experiment with a hypothesis and a primary metric, and create it as a draft. Don't launch."
- "Check all running experiments. For each, tell me the verdict, whether the split looks healthy, and what you'd decide, citing the segment results."
- "Experiment 4 is done. Conclude it with the winner, write a reason that mentions the segments, and add an annotation."
- "Describe every event that has no description yet, based on its properties."
- "Summarise this week's feedback from enterprise accounts into themes."
Stay in control
- The key's scope is the hard limit. A read-only key can analyse and recommend, but can't change anything. It doesn't even see the tools that write.
- Drafts are safe. Creating an experiment or a flag changes nothing for users. Launching an experiment and switching on or widening a flag do. Set your MCP client to ask before those.
- Everything can be undone. Stop the experiment, switch off the flag, and check Changes to see what happened.