Evaluation record · goose

Goose

v1.41.0 (2026-07-03)

Agentic AI Foundation (Linux Foundation); originally Block

Agentcoding-agentopen-sourcelocal-firstmcp
78
Strong
About This Agent

Open-source, local-first AI agent written in Rust, created by Block and donated to the Linux Foundation's Agentic AI Foundation (repo moved to aaif-goose/goose, April 2026). Ships as a CLI and desktop app, is MCP-native with 70+ extensions, and works with 15+ LLM providers including fully local models via Ollama. Free under Apache-2.0; vendor-neutral foundation governance is its core trust story, with the usual MCP/prompt-injection surface of local agents.

Last Evaluated: July 9, 2026
Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability
+
task completion accuracy

Capability review across coding and automation workflows; model-agnostic design means accuracy tracks the configured model

Evidence
Goose documentation — Executes coding and workflow tasks end-to-end (editing, running, testing); results depend on the user-selected LLM, from frontier APIs to local Ollama models
mediumVerified: 2026-07-09
tool use reliability

Review of MCP extension system maturity and built-in developer tooling reliability

Evidence
Goose GitHub repository (aaif-goose/goose) — MCP-native from the ground up with 70+ extensions and 15+ LLM providers across macOS, Linux, and Windows
highVerified: 2026-07-09
multi step planning

Evaluation of recipes, lead/worker mode, and long-horizon task behavior

Evidence
Goose documentation — Autonomous multi-step execution with recipes for repeatable workflows and lead/worker multi-model configurations for planning versus execution
mediumVerified: 2026-07-09
memory persistence

Review of session storage, hints files, and memory extension capabilities

Evidence
Goose documentation — Session persistence and resume, .goosehints project context files, and an optional memory extension; no managed long-term memory service
mediumVerified: 2026-07-09
error recovery

Observed recovery behavior from failing commands and the tool-inspection pipeline design

Evidence
Goose documentation — Agent loop retries failed commands and self-corrects from tool errors; a repetition inspector guards against runaway loops
mediumVerified: 2026-07-09
agent collaboration

Testing of subagent delegation and multi-model orchestration features

Evidence
Goose documentation — Subagents and recipe-driven task delegation supported; lead/worker model splits planning and execution across models
mediumVerified: 2026-07-09
🛡️Security
+
tool sandboxing

Execution isolation review; containerized deployment is possible but user-managed, mitigation is approval-based

Evidence
Goose permission modes documentation — No OS-level sandbox by default: goose runs shell and file tools with the user's privileges; permission modes (Auto, Approve, Smart Approve, Chat) gate execution rather than isolate it
mediumVerified: 2026-07-09
access control

Review of permission modes, allowlist enforcement, and the tool inspection pipeline

Evidence
Goose extension allowlist documentation — Administrators can enforce an extension allowlist (GOOSE_ALLOWLIST) restricting which MCP servers can be installed; tool calls pass a stacked inspection pipeline (security, egress, adversary, permission, repetition) before execution
highVerified: 2026-07-09
prompt injection defense

Injection surface review across MCP extensions and web-fetching tools; defenses are procedural (approvals) rather than architectural

Evidence
Goose documentation — Docs warn goose may follow commands embedded in fetched content; inspectors and approval modes reduce blast radius but the MCP/web-content injection surface matches peer local agents
mediumVerified: 2026-07-09
data isolation

Data-flow review of local-first architecture and extension access scope

Evidence
Goose GitHub repository (aaif-goose/goose) — Local-first: sessions, keys, and files stay on the user's machine; isolation between projects or extensions is not enforced beyond OS permissions
mediumVerified: 2026-07-09
open source transparency

License, source availability, and governance review

Evidence
Goose GitHub repository (aaif-goose/goose) — Apache-2.0, fully open Rust codebase (50.9K stars), now under vendor-neutral Linux Foundation AAIF governance with open security advisories
highVerified: 2026-07-09
🔒Privacy & Compliance
+
data retention

Privacy architecture review of the self-hosted, local-first model

Evidence
Goose GitHub repository (aaif-goose/goose) — Local-first design: no vendor cloud stores prompts or code; session data lives on the user's machine
highVerified: 2026-07-09
gdpr compliance

Compliance capabilities assessment across deployment configurations

Evidence
Goose documentation — As self-hosted software, GDPR posture is determined by the deployer and chosen model provider; fully local operation supports strict data-residency requirements
mediumVerified: 2026-07-09
third party data sharing

Data-flow analysis across model providers and extensions

Evidence
Goose documentation — Data flows only to the user-configured LLM provider and any installed MCP extensions; telemetry is limited and controllable
mediumVerified: 2026-07-09
local deployment option

Deployment options assessment including air-gapped local-model configurations

Evidence
Goose GitHub repository (aaif-goose/goose) — Runs fully offline with local models via Ollama and other local providers; CLI and desktop are both local applications
highVerified: 2026-07-09
👁️Trust & Transparency
+
documentation quality

Documentation completeness and accuracy review

Evidence
Goose documentation — Comprehensive docs covering permissions, allowlists, extensions, recipes, and enterprise guidance, maintained under the AAIF
highVerified: 2026-07-09
execution traceability

Review of session logs, permission records, and observability hooks

Evidence
Goose documentation — Full session transcripts of tool calls and outputs; permission decisions persisted in permission.yaml; observability integrations available but not enterprise-grade by default
mediumVerified: 2026-07-09
decision explainability

Assessment of pre-execution visibility and reasoning transparency

Evidence
Goose permission modes documentation — Approval modes surface each intended tool call with allow/deny prompts before execution, making agent intent visible step by step
mediumVerified: 2026-07-09
open source code

Open source and governance review; repo migration to aaif-goose completed April 2026

Evidence
Goose GitHub repository (aaif-goose/goose) — Apache-2.0 Rust codebase under Linux Foundation AAIF governance, donated by Block alongside Anthropic's MCP and OpenAI's AGENTS.md at the foundation's launch
Linux Foundation AAIF announcement — AAIF formed 2025-12-09 with goose as an anchor project; platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI
highVerified: 2026-07-09
community activity

Community engagement analysis of stars, release cadence, and post-donation governance

Evidence
Goose GitHub repository (aaif-goose/goose) — 50.9K stars, 5.5K forks, active Discord, and steady releases (v1.41.0 on 2026-07-03)
Goose blog - move to AAIF — Repository and docs migrated from block/goose to the Agentic AI Foundation on 2026-04-07, broadening the contributor base beyond Block
highVerified: 2026-07-09
⚙️Operational Excellence
+
ease of integration

Setup time and integration surface assessment across CLI and desktop

Evidence
Goose documentation — Single-binary CLI install or desktop app on macOS/Linux/Windows; provider setup is a guided config flow, and MCP extensions install in one step
highVerified: 2026-07-09
scalability

Assessment of headless usage, automation hooks, and organizational deployment patterns

Evidence
Goose GitHub repository (aaif-goose/goose) — Primarily a single-user local agent; headless/API usage and recipes enable automation, but fleet orchestration is left to the deployer
mediumVerified: 2026-07-09
cost predictability

Pricing model analysis of free software plus BYO model costs

Evidence
Goose GitHub repository (aaif-goose/goose) — Free Apache-2.0 software; costs limited to the chosen LLM API, and fully local models make zero-marginal-cost operation possible
highVerified: 2026-07-09
monitoring capabilities

Monitoring features assessment for individual and enterprise use

Evidence
Goose documentation — Session logs and cost/token tracking built in; centralized monitoring, alerting, and audit dashboards require external tooling
mediumVerified: 2026-07-09
production readiness

Maturity assessment from release history, Block production usage, and governance continuity

Evidence
Goose GitHub repository releases — Mature 1.x series (v1.41.0, 2026-07-03) with frequent releases; used internally at Block scale, and foundation governance reduces single-vendor abandonment risk
mediumVerified: 2026-07-09
Strengths
  • +Vendor-neutral governance: donated by Block to the Linux Foundation's Agentic AI Foundation (Dec 2025 announcement; repo moved April 2026)
  • +Local-first Rust application: fast, no vendor cloud, works fully offline with local models
  • +MCP-native with 70+ extensions and 15+ LLM providers
  • +Layered controls: permission modes, extension allowlist, and a stacked tool-inspection pipeline
  • +Free Apache-2.0 with an active community (50K+ stars) and steady release cadence
  • +Both CLI and desktop app across macOS, Linux, and Windows
Limitations
  • !No OS-level sandbox by default; shell and file tools run with full user privileges
  • !Prompt injection via MCP extensions and fetched web content remains an open risk, mitigated only by approval modes
  • !Single-user focus: no built-in multi-tenant, fleet management, or enterprise audit dashboards
  • !Output quality varies widely with the configured model, especially small local ones
  • !Post-donation transition (repo/docs moves, org rename) creates some ecosystem link rot and tooling churn
Metadata
license: Apache-2.0
repository: https://github.com/aaif-goose/goose
supported models
0: Anthropic Claude
1: OpenAI GPT models
2: Google Gemini
3: Local models via Ollama
4: 15+ providers total
languages
0: Rust
deployment type: Local CLI and desktop app (macOS, Linux, Windows); self-hosted
architecture: Local-first Rust agent with MCP-based extension system, permission modes, and recipes
first release: January 2025 (open-sourced by Block)
governance: Agentic AI Foundation (AAIF) at the Linux Foundation since April 2026 (announced 2025-12-09); founded alongside MCP and AGENTS.md
current version: 1.41.0 (2026-07-03)
github stars: 50900+
pricing: Free; users pay only their chosen LLM provider (or nothing with local models)

Use Case Ratings

code generation

Strong local coding agent with LSP-free simplicity and MCP extensibility; quality tracks the chosen model

data analysis

Good for scripted analysis and workflow automation via shell and MCP extensions

research assistant

Capable with web/MCP extensions; injection caution needed when fetching untrusted content

content creation

Serviceable for docs and technical writing; not its design center