MCP server wrapping the goose CLI — sessions, recipes, providers, extensions
Server exposes 9 tools with basic descriptions but significant schema and parameter documentation gaps. Tool names follow verb_noun convention (goose_version, goose_session_start, etc.), which is good. However, most tools lack visible input parameter schemas with type definitions and descriptions. The source code shows tool registration via @mcp.tool() decorators, but parameter schemas are inferred from function signatures rather than explicitly documented. Descriptions are present but brief (avg ~80 chars), lacking actionable context about when to use each tool or what happens on error. No error handling guidance visible. Output schemas are not documented. Tools like goose_session_start and goose_recipe_run perform state-changing operations but lack confirmation/dry-run patterns.
List active goose extensions from the config file.
List configured goose providers from the config file.
Run a goose recipe YAML file. Rationale: Recipes are portable YAML workflows — run one by local path.
List recent goose sessions with status summary.
Start a goose session with a text prompt. Rationale: Launch goose non-interactively with a task. The session runs in the background; poll goose_session_status with the returned id.
Get the current status and output of a goose session.
Return goose CLI version and availability status. Rationale: Quick sanity-check that goose is installed and on PATH. Returns: dict with raw version string and available bool.
Parameter schemas lack explicit type definitions and descriptions. Function signatures show parameters (e.g., prompt: str, provider: str, model: str in goose_session_start) but no JSON Schema with type/description pairs visible in tool registration. LLMs cannot infer parameter constraints from Python type hints alone.
Output schemas not documented. Tools return dicts (e.g., goose_version returns {available, error, bin_path}, goose_session_start returns {id, status, pid, started_at}) but no schema documentation visible. LLMs cannot plan downstream tool calls without knowing response structure.
No error handling guidance. Tools like goose_session_start and goose_recipe_run perform state-changing operations but provide no recovery hints. Error responses (e.g., RuntimeError in goose_version) are not categorized as retryable/user-fixable/fatal, leaving LLMs unable to decide next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Show goose installation status as a Prefab card.
Show recent goose sessions as a rich Prefab card.
UI-focused tools (show_sessions_card, show_goose_status_card) have vague names and descriptions. 'show_*' is not a standard action verb; unclear whether these are display-only or trigger side effects. Descriptions mention 'Prefab card' and 'app=True' but lack context on when/why to call them.
Destructive operations lack confirmation pattern. goose_recipe_run executes arbitrary YAML workflows by path; goose_session_start launches background processes. No dry-run, confirmation, or idempotency hints visible. Agents could accidentally trigger unintended automation.