母鹅
STDIO递归AI协作的MCP服务器管理工具
递归AI协作的MCP服务器管理工具
Mother Goose is a Model Context Protocol (MCP) server that enables AI agents to recursively spawn and interact with Block's Goose CLI, creating nested AI instances that can collaborate on complex problems.
Mother Goose allows AI agents (like Claude in Anthropic Console) to:
This creates a powerful recursive capability where an AI can delegate subtasks to other AI instances and have ongoing conversations with them, enabling more complex workflows and interactive reasoning chains.
For detailed documentation, please refer to the following resources:
The fastest way to use Mother Goose is with npx:
# Run validation to check prerequisites npx mother-goose validate # Run the MCP server npx mother-goose
Before using Mother Goose, you need:
goose run --text "Hello" from your terminalMother Goose provides the following tools to MCP clients:
run_gooseSpawns a new Goose process with the specified prompt.
Example:
Use the run_goose tool to research quantum computing.
list_goslingsLists all running and completed Goose processes.
Example:
Use the list_goslings tool to see all active processes.
get_gosling_outputGets the current output from a specific gosling process with pagination support.
Example:
# Basic usage
Use the get_gosling_output tool to check the results from process [ID].
# With pagination (Unix more/less style)
Use the get_gosling_output tool with offset=100 and limit=50 to view lines 100-150.
# Get full output
Use the get_gosling_output tool with full_output=true to get the complete output.
get_gosling_statusGets a compact activity report of goslings with their current working status (working/idle), helping prevent context overwhelm when managing multiple processes.
Example:
# Get status of all goslings
Use the get_gosling_status tool to see which goslings are ready for interaction.
# Get status of a specific gosling
Use the get_gosling_status tool with process_id="[ID]" to check if it's idle yet.
send_prompt_to_goslingSends a follow-up prompt to a running gosling process in interactive mode, enabling multi-turn conversations. Goslings are automatically started in interactive mode, so they stay alive and can receive multiple prompts.
Example:
Use the send_prompt_to_gosling tool to ask the gosling at process [ID] to elaborate on its findings.
release_goslingReleases a specific gosling process when you're done with it.
Example:
Use the release_gosling tool to release process [ID].
After configuring Mother Goose with your MCP client, you can create recursive AI workflows:
I need to solve a complex machine learning problem. Use the run_goose tool to create three specialist goslings:
1. One to research the latest papers on this topic
2. One to design an experimental approach
3. One to draft code snippets for implementation
Then, I'll coordinate their efforts to produce a comprehensive solution.
You can have multi-turn conversations with running goslings, providing feedback or additional instructions:
# Create a gosling to work on a task
Use the run_goose tool to create a gosling for drafting a research proposal.
# Check initial progress
Use the get_gosling_output tool to see the gosling's progress.
# Send follow-up instructions
Use the send_prompt_to_gosling tool to refine the focus: "Please emphasize the methodology section and expand on the data collection approach."
# Check updated output
Use the get_gosling_output tool to see how the gosling incorporated your feedback.
# Continue the conversation
Use the send_prompt_to_gosling tool again: "Now please add a budget section with estimated costs."
Use the activity sensor to manage multiple goslings efficiently without overwhelming your context:
# Start multiple goslings for different tasks
Use the run_goose tool to create three specialist goslings for different parts of a complex project.
# Check which goslings are ready for interaction
Use the get_gosling_status tool to see a compact report of all goslings.
# Only interact with idle goslings
For each gosling marked as 'idle' in the status report:
1. Review their output using get_gosling_output
2. Send follow-up prompts as needed using send_prompt_to_gosling
# Check activity status again
Use the get_gosling_status tool again to see which goslings have finished processing.
# This prevents context overflow by only retrieving complete outputs when necessary
Mother Goose includes a circuit breaker system to prevent token runaway scenarios where recursive AI instances could consume excessive resources:
# Configure the circuit breaker
Use the configure_circuit_breaker tool to set safety limits:
- enabled: true/false to enable/disable the circuit breaker
- max_active_goslings: Maximum number of concurrent active goslings
- max_total_goslings: Maximum number of total goslings including completed ones
- max_runtime_minutes: Maximum runtime in minutes for any single gosling
- max_output_size_kb: Maximum output size in KB for any single gosling
- max_prompts_per_gosling: Maximum number of prompts per gosling
- auto_terminate_idle_minutes: Auto-terminate goslings idle for this many minutes
# Emergency killswitch
Use the terminate_all_goslings tool to immediately stop all running gosling processes
The circuit breaker offers these safeguards:
Default settings provide a reasonable balance between flexibility and safety.
When combined with Memory Graph, all goslings can share knowledge through structured memory domains:
# Create specialized goslings that share a common memory structure
Use the run_goose tool to create a research gosling that stores findings in the shared memory.
Use the run_goose tool to create an implementation gosling that builds upon those findings.
Use the run_goose tool to create a testing gosling that validates the implementation.
# Each gosling can read, write, and follow connections in the shared memory
See Advanced Usage for more details.
For development, you can use the watch mode:
npm run watch
You can test the server using the MCP Inspector:
npm run inspector
# Run all tests npm test # Run specific test files npm test -- tests/basic.test.ts
MIT