DISCOVER THE FUTURE OF AI AGENTSarrow_forward

PostHog Code

calendar_todayAdded Apr 22, 2026
categoryAgent & Tooling
codeOpen Source
TypeScriptElectron桌面应用Multi-Agent SystemModel Context ProtocolAI AgentsAgent FrameworkAgent & ToolingDeveloper Tools & CodingData Analytics, BI & Visualization

An AI coding workbench for product developers that drives autonomous development and diagnostics from product usage data, with multi-agent parallel management and automatic instrumentation.

PostHog Code is an AI coding workbench built by PostHog, currently in Pre-alpha with a planned public release in Spring 2026. Its core innovation extends AI Agent decision signals beyond the codebase to the product analytics platform — ingesting multi-dimensional production signals including in-app activity, logs, error tracking, session replays, funnel analysis, Feature Flags, experiment results, internal discussions, and support tickets, enabling a full loop from signal detection to automatic Pull Request generation.

The project uses a TypeScript Monorepo architecture (pnpm workspaces + Turborepo) with three entry points: an Electron desktop app, a React Native mobile app, and a Jujutsu-based CLI tool. The desktop app provides a Command Center supporting split-screen monitoring of multiple coding Agents (side-by-side / 2×2 / 3×3), with built-in workflow panels for Signals Inbox, Plan, Review, and Skills. The Agent layer wraps the Claude Agent SDK, supporting multi-model switching between OpenAI GPT-5.x series and Anthropic Claude 4.x series.

For product engineering practices, PostHog Code automatically adds Event Instrumentation, configures Error Tracking, creates Feature Flags with gradual rollout conditions, and sets up A/B experiment variants during new feature development — achieving a "develop-and-observe" progressive release workflow. Through built-in MCP Server management, it can extend connectivity to external services including Attio CRM, HubSpot, Stripe, Salesforce, and Sentry, with additional integrations for Linear, Slack, GitHub, Granola, and Zendesk.

Architecturally, it enforces strict Main/Renderer process separation — the Main process handles all business logic, orchestration, and system I/O, while the Renderer process holds only pure UI state and tRPC action wrappers. It uses InversifyJS for dependency injection, Biome for code quality, and prohibits barrel files to ensure tree-shaking. The project is licensed under Business Source License 1.1, permitting personal use and contributions.

Unconfirmed information: exact release date (marked Spring 2026), pricing strategy, Windows/Linux support, mobile app release status, and model versions listed on the website (GPT-5.2/5.3-codex/5.4 and Claude 4.6 series are not yet publicly released and may be future plans or placeholder copy).

Related Projects

View All arrow_forward

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.

rocket_launch