Technology RadarTechnology Radar
Trial

Full deep dive: Goose Architecture Breakdown

Block's open-source AI agent framework built in Rust — MCP-native extension system, 25+ LLM providers, client/server architecture (goosed) shared across CLI, desktop, mobile, and IDE surfaces. 35K+ GitHub stars, Apache 2.0 licensed, governed by AAIF (Linux Foundation). Origin of Stripe's Minions fork.

Architecture Overview

Goose uses a client/server design:

  1. goosed Server — Rust binary wrapping the core Agent behind REST+SSE (103 endpoints) and ACP (Agent Client Protocol, JSON-RPC)
  2. Client Surfaces — CLI, Electron desktop app, mobile clients, Slack bots, and IDE extensions (VS Code, Zed, JetBrains)
  3. Extensions — All extensions are MCP servers (built-in: Developer, Computer Controller, Memory, Tutorial, Visualiser)

Key Design Decisions

Decision Detail
MCP-native First open-source agent to adopt MCP as its extension standard — 3,000+ community extensions
Provider-agnostic 25+ LLM providers including local models (Ollama); no vendor lock-in
Lead/worker model Multi-model config splits planning (stronger model) from execution (cheaper model)
Recipes Reusable task configurations bundling instructions + extensions + parameters
Memory extension Dynamic cross-session knowledge persistence (not static config files)
ACP consolidation Moving from bespoke REST API to open Agent Client Protocol standard

Why It's in Trial

Goose offers the most flexible extension model (MCP-native) and broadest provider support of any open-source coding agent. The Stripe Minions fork validates the architecture at enterprise scale. However, benchmark results depend entirely on the connected model (no standalone SWE-bench entry), and the lack of native LSP integration means error feedback is slower than OpenCode. The AAIF governance move strengthens long-term viability but introduces governance overhead. Worth evaluating for teams that need provider flexibility, MCP extensibility, or on-machine privacy.

Key Characteristics

Property Value
Company Block → AAIF (Linux Foundation)
System Goose (CLI + desktop + mobile + IDE)
Models Any — 25+ providers (Anthropic, OpenAI, Google, local)
Key innovations MCP-native extensions, ACP protocol, recipes, lead/worker model, memory extension
Open source Fully open (Apache 2.0)
Sources GitHub, Docs, Architecture Breakdown