goose
Coding · tested 2026-08-03 · re-test due 2026-11-03 · by the Hlido desk, not the vendor
In short: One of the most agent-native open-source agents on the market — Rust-built, MCP-deep, provider-agnostic, and now vendor-neutral under the Linux Foundation.
5 PASS · 0 FAIL of 5 public-surface claims
Quick answer
goose scores 84/100 (STEADY) on Hlido’s independent, hands-on test (reviewed 2026-08-03). STEADY (84) because the public surface shows an unusually agent-native architecture (CLI + API + deep MCP + ACP), transparent open-source access (Apache-2.0, no pricing lock-in), a named security model, and strong third- Pricing: Open source · Usage-based · Subscription (free entry point documented).
goose (originally from Block, donated to the Linux Foundation's Agentic AI Foundation in 2026) is the rare local agent that reads as infrastructure rather than a product wrapper. It ships three surfaces — a native desktop app, a full CLI, and an API — built in Rust for portability, and it treats the Model Context Protocol as a first-class integration layer rather than a bolt-on, with 70+ documented extensions. For an agentic ecosystem picking durable components, three things stand out: it is genuinely provider-agnostic (15+ LLM providers, and it can ride your existing Claude/ChatGPT/Gemini subscription via ACP), it is Apache-2.0 open source with no pricing lock-in, and its governance now sits with a vendor-neutral foundation instead of a single company's roadmap. The security posture it advertises — prompt-injection detection, tool-permission controls, sandbox mode, an adversary reviewer — is more than most agents bother to name, though we assessed it from the public surface only and did not exercise those controls hands-on. That is the honest boundary on this review: the maturity signals (45k+ stars, 500+ contributors, an active extension ecosystem) are strong and externally corroborated, but Hlido reviewed the public surface, not a full hands-on run, so execution-grit and day-to-day reliability are inferred from the ecosystem rather than measured here. On what is visible, goose is a serious, well-architected, agent-first tool.
Why STEADY
STEADY (84) because the public surface shows an unusually agent-native architecture (CLI + API + deep MCP + ACP), transparent open-source access (Apache-2.0, no pricing lock-in), a named security model, and strong third-party maturity signals (45k+ stars, 500+ contributors, Linux Foundation governance). Held below VITAL because this is a public-surface Tier-1 review — Hlido did not run goose hands-on, so its execution reliability, the effectiveness of its advertised security controls, and real-world task success are inferred from ecosystem signals rather than independently measured.
Public-surface checklist
- PASS Homepage loads (required)
- PASS Primary value prop (required) — 'Your native open source AI agent. Desktop app, CLI, and API'
- PASS Cta present (required) — 'Install goose' / 'Quickstart'
- PASS Pricing or access — Open source, Apache-2.0, free download; user brings own LLM keys/subscription
- PASS Evidence or demo — Feature grid + 'See goose in action' + extensions directory linked
What it does well
- Agent-native by design — first-class CLI and API, not just a chat UI
- Deep, standards-based extensibility via the Model Context Protocol (70+ extensions)
- Provider-agnostic: 15+ LLM providers, and can reuse existing Claude/ChatGPT/Gemini subscriptions via ACP
- Open source (Apache-2.0), built in Rust, with no pricing or vendor lock-in
- Vendor-neutral governance under the Linux Foundation's Agentic AI Foundation
- Names a concrete security model (prompt-injection detection, tool permissions, sandbox mode, adversary reviewer)
What it fails at
- Security controls are advertised but were not exercised in this public-surface review
- General-purpose framing means capability depth varies by task type and is not uniform
- Local-first, provider-your-own-key model puts setup and cost management on the user
- Real-world reliability and task-success rate are not independently measured on this surface
- No first-party managed/hosted tier — everything runs on the user's machine and keys
Best for
- Developers who want a local, open-source agent they can drive from the terminal or embed via API
- Teams standardising on MCP who need deep, documented extension support
- Users who want to reuse an existing LLM subscription instead of paying per-token separately
- Organisations wary of single-vendor lock-in who value Linux Foundation governance
Not recommended for
- Non-technical users wanting a zero-setup, fully hosted assistant
- Buyers who need a vendor SLA and managed hosting rather than self-run software
- Workflows that require a turnkey UI over programmable, local-first tooling
Pricing & access
- ModelOpen source · Usage-based · Subscription
- Free entry pointYes — a free tier or open-source edition is documented
- Pricing findable on the public surfacePASS Open source, Apache-2.0, free download; user brings own LLM keys/subscription (tested 2026-08-03)
Derived from Hlido-held evidence only (engine checklist + editorial text); quotes are verbatim from the scorecard; not vendor-supplied; re-derived daily. Verify current prices on the vendor's pricing page. Last verified 2026-08-03.
Compared to
-
Aider
breadth-and-extensibility
aider is a focused terminal coding assistant; goose is a broader local agent platform (desktop + CLI + API) with a deep MCP extension model. Choose aider for tight git-native code editing; choose goose when you want a general-purpose, extensible local agent that also codes.
Agent relevance
API CLI MCP SDK Behavioral-testable
Agentic-Commerce Readiness 85/100 · COMMERCE-READY
Independent readiness for agent delegation & transaction. How it’s scored · check live
Strongly agent-native — drivable from a CLI, embeddable via API, and MCP-first with 70+ extensions plus ACP interop (works as an ACP server and can use ACP agents like Claude Code and Codex as providers). One of the more directly agent-integrable tools in the corpus.
Agent-friendly score: 9/10
Score over time
The longitudinal record — every point is the score as published on that date. Raw series.
Evidence
- Native desktop app, full CLI, and API, built in Rust — source (2026-08-03) verified
- Connects to 70+ extensions via the Model Context Protocol — source (2026-08-03) verified
- Works with 15+ LLM providers and supports existing subscriptions via ACP — source (2026-08-03) verified
- Open source (Apache-2.0), part of the Linux Foundation's Agentic AI Foundation — source (2026-08-03) verified
- Advertises prompt-injection detection, tool permissions, sandbox mode, adversary reviewer — source (2026-08-03)