AI agent for PostHog: product analytics, feature flags, experiments and surveys
PostHog connects with OAuth through PostHog Cloud, US or EU; each PostHog project becomes its own connection. The agent has 29 reads and 20 writes. Insights, dashboards, cohorts and draft experiments and surveys follow your approval setting; turning a flag on, changing a rollout, launching or stopping an experiment or survey, and deletes always wait for a person. Person data is masked everywhere.
Connecting PostHog
- 1Open Integrations → the PostHog card, choose US or EU (the cloud your PostHog organization lives in) and press Connect.
- 2Sign in to PostHog and approve the permissions it lists (see the setup guide for the full list).
- 3Agent Planners saves one connection per PostHog project the sign-in reaches, grouped by region and PostHog organization.
What the agent reads
- Analytics: read-only HogQL (SQL) queries, trends, funnels, retention, paths, lifecycle, stickiness, sessions in aggregate (duration, bounce rate, by channel, page, device or country) and click / scroll heatmaps for a page.
- Saved work: insights with their queries and cached results, dashboards and their tiles, event definitions.
- Product: feature flags with their release conditions, experiments and their results when PostHog has them, surveys with a response summary (counts per answer — open-text answers and respondents are never returned), cohorts (never their members).
- Errors and recordings: error-tracking issues, and session-recording METADATA (duration, clicks, errors, the start page) with who it was replaced by "Person N" — the replay itself is never fetched.
- Health: whether the project is reachable and still receiving events.
What changes, and who decides
- Your approval setting decides (by default each waits for your approval card): creating and editing insights, dashboards and dynamic cohorts, adding an insight to a dashboard, creating an experiment or survey as a DRAFT, creating a feature flag INACTIVE, renaming or tagging a flag, and an error issue's status or assignee.
- Always a person, on every lane: turning a flag on or off, changing a flag's rollout, release conditions or variants, launching or stopping an experiment or survey, and deleting anything (PostHog keeps a deleted object restorable for a while).
- Refused: persons, deleting recordings, members, billing, API keys, project settings, batch exports, destinations and webhooks.
Things to ask
- "Signup funnel from pageview to signup to first project, last 30 days."
- "Weekly retention for users who signed up in October."
- "Which feature flags are on, and at what rollout?"
- "Create a draft experiment on the new pricing page, 50/50 — I'll launch it."
Frequently asked questions
- Does Agent Planners support PostHog?
- Yes — PostHog Cloud in the US or EU region, connected with OAuth. The agent queries analytics and reads flags, experiments, surveys and errors, and makes changes with your approval.
- Can the agent turn on a feature flag?
- Only with a person's approval, every time. Turning a flag on or off and changing who it reaches are held for a person whatever the approval setting; a new flag is always created inactive.
- Does the agent see who my users are?
- No. Emails, names and person properties are masked in every answer, reading people is refused over MCP and on automated runs, and recordings come back as metadata with the person replaced by "Person N".
- Does it work with self-hosted PostHog?
- No. Only PostHog Cloud's US and EU regions can be connected.