PostHog

PostHog AI acts inside PostHog itself, creating dashboards, surveys, alerts and feature flags, and reaches GitHub for writing, where it starts a coding task on a connected repository and returns a pull request for review.

Scouts, the scheduled agents in the self driving suite, which is in open beta for any team that sets it up, can use the team's own shared MCP servers beside the PostHog MCP, so they act in whatever outside systems those servers reach, and they post findings to Slack channels or direct messages. Warehouse sources such as Stripe, HubSpot, Salesforce and Zendesk are synced in for reading.

Sourceposthog.com/docs/self-driving/scoutsread 2026-11-09

One PostHog AI conversation chains many tool calls, moving from SQL to the data schema to session replays and then saving the result as a notebook. In the open beta self driving suite, scheduled scouts emit signals, a pipeline groups related signals into reports, and research and implementation agents turn a report into a pull request, with separate steering instructions for each stage. Customers get no builder for their own multistep agent graphs, branching or handoffs between agents they define.

Sourceposthog.com/docs/self-driving/scoutsread 2026-11-09

Answers draw on the customer's own events, persons, sessions, cohorts, saved insights and synced warehouse tables, queried at question time through SQL and the data schema. A semantic layer stores team approved definitions of metrics such as active users and MRR, which the agent applies every time and marks as approved, and Azure runs semantic search over the project's actions and cohorts. There is no store for uploading company documents, wikis or policies for the agent to retrieve from.

Sourceposthog.com/airead 2026-11-09

Approval policies on the Scale package turn feature flag and experiment changes into change requests that named users or roles must approve, with a quorum, optional conditions such as a rollout jump above 20%, and self approval off by default; approving over MCP needs a typed confirm. An admin must approve PostHog AI's use of outside model providers before the first chat. Scouts receive write access per object type only from their author or a project admin, and a dry run blocks every write.

Sourceposthog.com/docs/settings/approvalsread 2026-11-09

Access control on the Boost package sets who can see and change data and features, Scale adds SAML and OpenID Connect single sign on, and Enterprise adds custom roles, SCIM provisioning from the identity provider, IP restrictions and property level access control. PostHog AI sees only what the asking user can see. The company holds a SOC 2 Type 2 attestation from an external audit, with the latest report open to read, runs annual penetration tests, and signs a HIPAA BAA on Boost and above.

Sourceposthog.com/platform-packagesread 2026-11-09

Activity logs record who in the organization viewed or changed entities, kept 7 days on Boost, 2 months on Scale and 60 months on Enterprise, and every write a scout makes lands there under its owner, marked as made through MCP. For the agents a team builds, each trace is a waterfall of spans and generations with inputs, outputs, tool calls, tokens, cost and latency, linked to the user's session replay, and failed calls become grouped issues in error tracking with trace summaries and clustering alongside.

Sourceposthog.com/ai-observabilityread 2026-11-09

PostHog AI keeps a project memory of up to 10,000 characters that persists between conversations. The /init command researches the product and saves what it finds, /remember adds facts, and people edit the text from the chat gear icon or the project's AI settings; past 5,000 characters the agent reads only the first and last 2,500. No retention period is set for it. Scouts, in open beta, keep their own memory across scheduled runs of what is normal and what they already reported.

Sourceposthog.com/docs/posthog-ai/edit-memoryread 2026-11-09

Customers choose PostHog Cloud US, on AWS in Virginia, or PostHog Cloud EU, on AWS in Frankfurt, where data stays inside EU jurisdiction, and the MCP server routes each account to its own region. The core product is also available as MIT licensed open source that a company can run on its own servers. Transfers from the UK and EU into the US region rely on Standard Contractual Clauses, and a signed DPA is self serve on every plan, including the free one.

Sourceposthog.com/handbook/company/securityread 2026-11-09

Creating an evaluation opens a template catalog grouped into answer quality, safety, agent behavior, user insights and cost and speed, with ready judges for response quality on a 1 to 5 scale, task outcome, user correction, relevance, hallucination and prompt injection, plus Hog code and sentiment templates. PostHog also ships built in scouts that a project switches on, suggests new scouts from the project's own data, and stores agent skills with version history that a team shares through MCP.

Sourceposthog.com/docs/ai-evalsread 2026-11-09

Online evaluations run on generations as they arrive, sampled anywhere from 0.1% to 100% and filtered by event or person properties, and a summary report fires after every 100 new results by default, delivered by email or Slack. Anomaly alerts learn the normal level of cost, latency and errors and fire without fixed thresholds. PostHog AI answers @PostHog mentions in Slack threads, and scouts run on intervals from 30 minutes to 30 days or on cron expressions in the project's time zone.

Sourceposthog.com/docs/ai-evalsread 2026-11-09

LLM judges run on whichever model the customer selects from its own provider keys for OpenAI, Azure OpenAI, Google Gemini, Anthropic, OpenRouter, Fireworks or Together AI, with a choice between explained verdicts and faster decision models and a reasoning level from minimal to high. The prompt playground compares models side by side on real inputs. PostHog AI itself runs on Anthropic, OpenAI and Google models that PostHog selects, and the customer cannot pick among them.

Sourceposthog.com/docs/ai-evalsread 2026-11-09

A hosted MCP server at mcp.posthog.com accepts OAuth or a personal API key as a bearer token, lists every tool in a public reference, and creates and edits PostHog's own objects, including flags, dashboards, experiments, surveys, prompts and evaluations, with read only mode, tool filtering and project pinning. A REST API sends traces in, queries them back and drives evaluations, reviews and clustering from the customer's code, and the PostHog CLI exposes the same tool catalog as shell commands for terminals and CI.

Sourceposthog.com/docs/model-context-protocol/faqread 2026-11-09

Evaluations score live generations, whole traces or whole sessions with LLM judges, deterministic Hog code or a local sentiment model, returning pass or fail, a numeric score or categories, with passing rules and pass rates over time. Offline evaluations run against curated datasets before a change ships, backfills regrade up to 30 days of past data, and prompt versions can be split across users in an A/B experiment that measures cost, latency and evaluation pass rate. Annotation queues route outputs to human reviewers.

Sourceposthog.com/docs/ai-evalsread 2026-11-09

Nothing in the AI layer drives a browser, a remote desktop or another application's interface. Agents work through PostHog's own query engine, MCP tools and API, and coding tasks run in a cloud sandbox against a cloned repository. Replay Vision scanners watch recorded user sessions and turn what they see into data a team can query, which reads past sessions rather than operating a live interface on anyone's behalf.

Sourceposthog.com/airead 2026-11-09