PostHog AI plugin

Official PostHog plugin for AI clients. Access PostHog products directly from your AI coding tool.

Installation

Claude Code

1. Install the plugin:

   claude plugin install posthog
   

   Or install from this repo as a marketplace:

   claude plugin marketplace add PostHog/ai-plugin
   claude plugin install posthog@posthog
   

2. Authenticate via OAuth:

   # Just enter Claude Code anywhere
   claude
   # Then, use the /mcp command within Claude, select plugin:posthog:posthog, and press Enter
   /mcp
   

   Then follow the browser prompts to log into PostHog.

3. (Optional) Send Claude Code sessions to PostHog LLM Analytics.

   Add to ~/.claude/settings.json (global) or .claude/settings.local.json (per-project):

   {
     "env": {
       "POSTHOG_LLMA_CC_ENABLED": "true",
       "POSTHOG_API_KEY": "phc_...",
       "POSTHOG_HOST": "https://eu.i.posthog.com"
     }
   }
   

   Both POSTHOG_LLMA_CC_ENABLED=true and POSTHOG_API_KEY are required. Sessions are sent when Claude Code exits. Set POSTHOG_LLMA_PRIVACY_MODE=true to redact prompt/output content. Add custom properties to all events with POSTHOG_LLMA_CUSTOM_PROPERTIES (JSON string, e.g. '{"ai_product": "my-app"}'). Set POSTHOG_AI_PURPOSE to declare what the session is doing relative to the repo — documented values are authoring and review, emitted as $ai_purpose. Useful when a harness wraps an agent it can't modify (e.g. a CI job running Claude Code as a reviewer); left unset the property is omitted.

Session capture without the rest of the plugin

To send Claude Code sessions to PostHog LLM analytics without the skills, the slash commands, or the MCP server, install the capture plugin on its own:

claude plugin marketplace add PostHog/ai-plugin
claude plugin install posthog-telemetry@posthog

It registers one SessionEnd hook and nothing else, so it adds nothing to the context of a session. It needs only the Python standard library, and it reads the same environment variables as step 3 above.

Install it instead of the posthog plugin, not alongside it. Both register the same hook, so running both parses and uploads every session twice. PostHog dedupes the events themselves, because their UUIDs are derived from the session, but the two hooks race on the status file.

Cursor

Install from the Cursor Marketplace or add manually in Cursor Settings > Plugins.

The plugin installs the PostHog skills and the MCP server. Claude Code's session capture and the write-command gate don't run in Cursor.

In Cursor IDE, the MCP connection runs through Cursor's hosted service, so Cursor holds your PostHog sign-in and you may need to sign in again after updating the plugin.

Grok Bot (desktop and mobile) lists the same plugin from the Cursor Marketplace. Open Plugins, search for PostHog, install it, and log into PostHog when prompted.

Codex

1. Add the marketplace:

   codex plugin marketplace add PostHog/ai-plugin
   

2. Install the plugin from inside Codex:

   codex
   # Then run /plugins, select PostHog, and install
   /plugins
   

3. Trust the plugin's hook when Codex asks you to review it at startup, or later from /hooks. The hook approves known PostHog reads (tools, search, info and the other read-only exec commands, and calls to tools on hooks/read-tools.txt) so they run without a prompt. Every other call, including every write, gets Codex's own approval prompt. The POSTHOG_MCP_EXEC_GATE_* variables don't apply on Codex. Codex asks before every PostHog call until you trust the hook, and again after each plugin update until you re-trust it. The one exception is render-ui, which shows a PostHog chart or view in the conversation and never changes data, so it always runs without a prompt.

   A few Codex settings change this:

   • A [mcp_servers.posthog] entry in your own ~/.codex/config.toml replaces the plugin's server config entirely, including the prompt mode the gate relies on. Remove it, or add default_tools_approval_mode = "prompt" and a [mcp_servers.posthog.tools.render-ui] table with approval_mode = "approve" to it.
   • With approval policy never and full access (--dangerously-bypass-approvals-and-sandbox), Codex approves MCP calls before the hook runs, so nothing is gated.
   • With approval policy never and a restricted sandbox, calls the hook doesn't approve are denied instead of prompted.

Antigravity

1. Install the plugin:

   agy plugin install https://github.com/PostHog/ai-plugin
   

2. Authenticate via OAuth: run /mcp, select posthog_posthog, and choose Authenticate. Open the link it shows, log into PostHog, and paste the authorization code back into Antigravity.

The plugin installs the PostHog skills and the MCP server. Claude Code's session capture and the write-command gate don't run in Antigravity.

Gemini CLI

gemini extensions install https://github.com/PostHog/ai-plugin

Grok

1. Install the plugin:

   grok plugin install PostHog/ai-plugin --trust
   

2. Authenticate via OAuth:

   On first use of a PostHog tool, Grok prompts you to authorize in your browser. Log into PostHog to connect.

Claude Cowork

1. Add the marketplace: open Settings > Plugins, click Add > Add marketplace, enter PostHog/ai-plugin, and click Sync.

2. Install the plugin: click Browse, select PostHog, and install. The plugin stays in sync with this repo automatically.

3. Authenticate via OAuth: on first use of a PostHog tool, follow the browser prompts to log into PostHog.

How to develop

1. Clone and install the plugin:

   git clone https://github.com/PostHog/ai-plugin
   claude --plugin-dir ./ai-plugin
   

2. Authenticate via OAuth:

   /mcp
   

   Then follow the browser prompts to log into PostHog.

Features

This plugin provides access to 27+ PostHog tools across these categories:

• Feature flags - Create, update, and manage feature flags
• Experiments - Run and analyze A/B tests
• Insights - Query analytics and create visualizations
• Dashboards - Manage dashboards and add insights
• Error tracking - View and debug errors
• LLM analytics - Track AI/LLM costs and usage
• Documentation - Search PostHog docs
• And more

Bundled skills

The plugin also ships 30+ task-specific skills that your AI client loads on demand to follow PostHog best practices — covering HogQL query patterns, experiment creation and lifecycle, feature flags, data warehouse setup and troubleshooting, LLM analytics exploration, session replay diagnostics, and SDK instrumentation. Skills activate automatically when their description matches your request (e.g. "create an experiment", "why isn't my Stripe sync working?", "audit my feature flags"), so you generally don't need to invoke them by name.

Example usage

> What feature flags do I have?
> Create a feature flag called new-onboarding for 50% of users

> Show me errors from the last 24 hours
> Which errors are affecting the most users?

> How many users signed up this week?
> What's the conversion rate for the checkout funnel?

> Show me all my experiments
> What are the results of the checkout-flow experiment?

> Create a new dashboard called Product Metrics
> Add the signup funnel insight to the Growth dashboard

> What are the responses to the NPS survey?
> Create a feedback survey for the checkout page

> What's my most triggered event?
> Show me the top 10 pages by pageviews

Plugin usage telemetry

In Claude Code, the plugin sends one anonymous plugin skill invoked event to PostHog Cloud each time one of this plugin's skills runs, whether Claude invokes it or you type /posthog:<skill>. Skills from other plugins and your own skills aren't counted. Each event contains only:

• skill: the skill's name, such as querying-posthog-data
• invocation: model when Claude invoked the skill, user when you typed it as a slash command
• args_present: whether the skill was invoked with arguments, never the arguments themselves
• plugin_version: the installed plugin version
• permission_mode: Claude Code's permission mode, such as default or plan
• os: the operating system name from uname -s, such as Darwin or Linux
• distinct_id: a random UUID generated once per install and stored in the plugin's data directory
• $lib: always posthog-ai-plugin
• $process_person_profile, $geoip_disable, and $ip: fixed values (false, true, and 127.0.0.1) so PostHog creates no person profile and runs no GeoIP lookup, and the stored event carries no client IP address. The request still reaches PostHog's capture endpoint from your IP, as any HTTPS request does

No event is sent when POSTHOG_MCP_URL or POSTHOG_HOST points anywhere other than posthog.com, so self-hosted users don't report to PostHog Cloud. Like Claude Code's own telemetry, it's also off when Claude Code runs on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, or Claude Platform on AWS (CLAUDE_CODE_USE_BEDROCK, CLAUDE_CODE_USE_VERTEX, CLAUDE_CODE_USE_FOUNDRY, or CLAUDE_CODE_USE_ANTHROPIC_AWS set to 1, true, yes, or on). To turn it off, set any of these in your shell or the env block of ~/.claude/settings.json:

export DO_NOT_TRACK=1
export POSTHOG_PLUGIN_TELEMETRY_DISABLED=1
export DISABLE_TELEMETRY=1
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

DO_NOT_TRACK and DISABLE_TELEMETRY turn off Claude Code's own telemetry too, and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC turns off all of Claude Code's nonessential traffic. POSTHOG_PLUGIN_TELEMETRY_DISABLED turns off only this plugin's telemetry.

Self-hosted

For self-hosted PostHog instances, set the POSTHOG_MCP_URL environment variable to point to your instance:

export POSTHOG_MCP_URL="https://mcp.your-posthog-instance.com/mcp"

Documentation

• PostHog MCP documentation
• PostHog API documentation

License

MIT