← All agent and harness briefs
Agent runtimeagenticcodingopsdocumentedunattended capable

Operational brief

Goose

AAIF's local, general-purpose agent with desktop, CLI, API and MCP tools.

Official goose banner from goose-docs.ai

Choose it when

You want a local general-purpose agent with MCP tools, reusable recipes and an optional schedule whose history you can inspect.

Skip it when

You can't review extension authority, choose a permission mode or own failures from scheduled work.

01 / Classification

Autonomy evidence

This is our classification, not a vendor badge. The record says what the product normally does. The gaps stay visible.

How we classify it
Tool use + multi-step plans, still human-gated per session.
Runs unattended
Yes — that capability is documented in this brief.
Trigger
Human desktop/CLI/API sessions normally; a configured recipe schedule can also start background work without a chat turn.
State
Saved sessions use SQLite. Recipes retain workflow setup, not full conversation history; schedule details expose the sessions each recipe run created.
Tools
Enabled MCP extensions provide tool use. Depending on permission mode, Goose can modify or delete files; a recipe doesn't independently certify its tools as safe.
Human control
Choose a permission mode for each session: changing the settings default affects new sessions, not existing ones. Manual and Smart modes support per-tool overrides, but write-tool classification is best effort. Inspect scheduled-run sessions and pause a Desktop schedule or remove a CLI schedule; a human must own outputs and failures.

02 / Operations

Control reality

Useful agents eventually touch something real. This is the operational picture we have, not a promise hidden in a sales page.

Where it runs
Native app and CLI run on your machine; API and remote extension choices can add external services.
Data boundary
Usage telemetry is opt-in and separately configurable. Conversations, prompts and information goose accesses may still go to your chosen model provider, subject to its privacy and retention policies; remote extensions may also receive task data. Local execution isn't a no-egress guarantee. Recipe sharing excludes credentials and memory, not necessarily sensitive instructions you put in the recipe.
Identity & permissions
Use your selected provider's API credentials or supported ACP account path; connected extensions have their own authority. No account/provider compatibility was tested here.
What it can write
local tools via MCP
Surfaces
Desktop · CLI · API
Export / lock-in
CLI documents session export as Markdown, JSON or YAML; Desktop exports recipe YAML. Review exported history and recipe contents before sharing.
Operational limits
Documented, not app-tested. Scheduled recipes can run unattended, but availability, results and write authority depend on the host, provider, extensions and permissions. Optional retry can run a cleanup command, reset agent state and rerun the recipe; that's not verified checkpoint recovery or idempotence. Review cleanup and reconcile external writes before reruns. OS isolation and schedule reliability aren't established. Model and service costs and entitlements are separate from the Apache-2.0 code license.
Watch for
Docs say Autonomous Mode is the default; don't assume every write asks. The Manual Approval description is broader than its operational note: prompts depend on best-effort write-tool classification interpreted by the model provider. Check the mode of each session, including scheduled work; changing the default doesn't alter existing sessions. Optional recipe retry resets agent state and starts over, so reconcile external writes before a rerun. Upstream v1.59.2 is a captured release, not proof of your installed version. Controls and recovery aren't runtime-tested here.

03 / Receipts

Evidence ledger

Here is what this brief is built on. “Documented” means we found a source. “Tested” means this desk actually ran it. Those are different things.

Evidence status
documented
Source set
14 primary sources
Evidence review
2026-11-17

Corrections

  • 2026-10-13 · Updated the canonical repository from block/goose to aaif-goose/goose. Current README confirms AAIF/Linux Foundation lineage; license is Apache-2.0, not a foundation-membership label.
  • 2026-10-13 · Replaced the undated stub release metadata with v1.54.0's API publication timestamp. Corrected categorical unattended:false to scheduled-recipe capability; normal session classification remains agentic, with autonomous secondary and documented-only evidence.
  • 2026-10-17 · Refreshed release metadata from v1.54.0 to v1.55.0 using API publication time. Release notes document parent-bound subagent hooks/mode/approvals, pending-approval and Stop handling, malformed GOOSE_MODE rejection across ACP, and removal of background subagents/orchestrator. Scheduled recipes remain documented separately; no autonomy reclassification or runtime validation is implied. Documented, not runtime-tested.
  • 2026-11-17 · Refreshed upstream release metadata to v1.59.2, published 2026-11-16T05:26:30Z according to the retained release API. This is not an installed-version observation or a runtime test; earlier v1.55.0 release details remain historical.

04 / Change log

Related coverage

The brief tells you what the product is. Payloads track the changes that may alter the decision.