agent-runtime-mcp
Enables persistent task and goal management with AI-powered decomposition, cross-session continuity, and fault-tolerant multi-agent pipelines.
README
Agent Runtime MCP
Persistent task queues and goal decomposition for cross-session AGI autonomy with God Agent integration.
Description
Agent Runtime MCP provides persistent task management that survives across sessions, enabling true autonomous AGI workflows. Features include:
- Persistent Goals & Tasks: SQLite-backed storage that survives restarts
- AI-Powered Goal Decomposition: Break complex goals into executable tasks
- Dependency Management: Automatic task ordering based on dependencies
- Priority Queuing: Intelligent task scheduling by priority and readiness
- Relay Race Protocol (God Agent Phase 2): 48-agent pipelines with structured handoffs
- Circuit Breaker (God Agent Phase 5): Fault tolerance with automatic fallback
- Cross-Session Continuity: Resume work exactly where you left off
Installation
Using pip
git clone https://github.com/marc-shade/agent-runtime-mcp
cd agent-runtime-mcp
pip install -r requirements.txt
Using uv (recommended)
git clone https://github.com/marc-shade/agent-runtime-mcp
cd agent-runtime-mcp
uv pip install -r requirements.txt
Dependencies
pip install anthropic-mcp
Configuration
Add to ~/.claude.json:
{
"mcpServers": {
"agent-runtime": {
"command": "python3",
"args": [
"/absolute/path/to/agent-runtime-mcp/server.py"
]
}
}
}
Tools
Core Goal & Task Management (9)
| Tool | Description |
|---|---|
create_goal |
Create high-level goal with name and description |
decompose_goal |
Use AI to break goal into tasks (sequential/parallel/hierarchical) |
create_task |
Manually create task with dependencies |
get_next_task |
Get next ready task from queue (highest priority, deps met) |
update_task_status |
Update status (pending/in_progress/completed/failed/cancelled) |
list_goals |
List all goals, optionally filtered by status |
list_tasks |
List tasks by goal, status, with limit |
get_goal |
Get goal details by ID |
get_task |
Get task details by ID |
Relay Race Protocol (God Agent Phase 2) (6)
| Tool | Description |
|---|---|
create_relay_pipeline |
Create 48-agent relay race with baton passing |
get_relay_status |
Get pipeline status (progress, quality scores) |
advance_relay |
Pass baton to next agent after completing step |
retry_relay_step |
Retry failed step without restarting pipeline |
list_relay_pipelines |
List pipelines by status |
get_relay_baton |
Get current baton with context for next agent |
Circuit Breaker (God Agent Phase 5: Tiny Dancer) (7)
| Tool | Description |
|---|---|
circuit_breaker_status |
Get breaker state (CLOSED/OPEN/HALF_OPEN) |
circuit_breaker_list |
List all breakers with open/degraded circuits |
circuit_breaker_trip |
Manually trip breaker to OPEN state |
circuit_breaker_reset |
Reset breaker to CLOSED state |
circuit_breaker_configure |
Configure thresholds (failures, window, cooldown) |
circuit_breaker_record_failure |
Record failure for tracking |
circuit_breaker_record_success |
Record success (helps recovery) |
Usage Examples
Basic Goal Creation
# Create goal
goal = mcp__agent-runtime__create_goal({
"name": "Build REST API",
"description": "Create RESTful API for user authentication with JWT tokens",
"metadata": {"priority": "high", "project": "auth-service"}
})
# Returns: {"id": 1, "name": "Build REST API", "status": "active", ...}
AI Goal Decomposition
# Decompose goal into tasks (sequential strategy)
result = mcp__agent-runtime__decompose_goal({
"goal_id": 1,
"strategy": "sequential"
})
# Returns: {
# "goal_id": 1,
# "strategy": "sequential",
# "tasks_created": [101, 102, 103, 104, 105],
# "count": 5
# }
# Tasks: Research → Plan → Implement → Test → Document (with dependencies)
Parallel Decomposition
# Decompose for parallel execution
result = mcp__agent-runtime__decompose_goal({
"goal_id": 1,
"strategy": "parallel"
})
# Creates: Backend, Frontend, Testing tasks (no dependencies, run simultaneously)
Hierarchical Decomposition
# Decompose into phases
result = mcp__agent-runtime__decompose_goal({
"goal_id": 1,
"strategy": "hierarchical"
})
# Creates: Phase 1 (Foundation) → Phase 2 (Core) → Phase 3 (Integration) → Phase 4 (Optimization)
Task Queue Processing
# Get next ready task
task = mcp__agent-runtime__get_next_task()
# Returns: Highest priority task with all dependencies met
# {"id": 101, "title": "Research requirements...", "priority": 10, ...}
# Start work
mcp__agent-runtime__update_task_status({
"task_id": 101,
"status": "in_progress"
})
# Complete task
mcp__agent-runtime__update_task_status({
"task_id": 101,
"status": "completed",
"result": "Requirements documented in docs/api-spec.md"
})
# Get next (automatically handles dependencies)
next_task = mcp__agent-runtime__get_next_task()
# Returns: Task 102 (Plan approach) since Research (101) is complete
Manual Task Creation with Dependencies
# Create task with explicit dependencies
mcp__agent-runtime__create_task({
"goal_id": 1,
"title": "Deploy to production",
"description": "Deploy authentication service",
"priority": 7,
"dependencies": [103, 104] # Wait for Implementation and Testing
})
Relay Race Pipeline (48-Agent)
# Create relay pipeline for complex workflow
pipeline = mcp__agent-runtime__create_relay_pipeline({
"name": "Research Paper Analysis",
"goal": "Extract insights from 10 AGI papers",
"agent_types": [
"researcher", # Gather papers
"analyzer", # Extract key points
"synthesizer", # Find patterns
"validator", # Check quality
"formatter" # Create report
],
"token_budget": 100000
})
# Returns: {"pipeline_id": "rp_abc123", "agent_count": 5, ...}
# Check pipeline status
status = mcp__agent-runtime__get_relay_status({
"pipeline_id": "rp_abc123"
})
# Returns: {
# "current_step": 2,
# "total_steps": 5,
# "status": "in_progress",
# "quality_scores": [0.92, 0.88, ...],
# "tokens_used": 24531
# }
# Get current baton (context for next agent)
baton = mcp__agent-runtime__get_relay_baton({
"pipeline_id": "rp_abc123"
})
# Returns: {
# "baton": {...},
# "prompt": "You are the Synthesizer. Previous output: ..."
# }
# Advance to next step
mcp__agent-runtime__advance_relay({
"pipeline_id": "rp_abc123",
"quality_score": 0.88,
"l_score": 0.85,
"output_entity_id": 456,
"tokens_used": 8234,
"output_summary": "Found 3 key patterns across papers"
})
# Retry failed step
mcp__agent-runtime__retry_relay_step({
"pipeline_id": "rp_abc123",
"step_index": 2
})
Circuit Breaker (Fault Tolerance)
# Check agent circuit breaker status
status = mcp__agent-runtime__circuit_breaker_status({
"agent_id": "researcher_agent"
})
# Returns: {
# "agent_id": "researcher_agent",
# "state": "CLOSED",
# "failure_count": 0,
# "success_count": 42
# }
# Record failure
mcp__agent-runtime__circuit_breaker_record_failure({
"agent_id": "researcher_agent",
"failure_type": "timeout",
"error_message": "API request timed out after 30s"
})
# List all circuit breakers
breakers = mcp__agent-runtime__circuit_breaker_list()
# Returns: {
# "total_breakers": 10,
# "open_circuits": ["failing_agent_1", "failing_agent_2"],
# "half_open_circuits": ["recovering_agent"],
# "breakers": [...]
# }
# Configure thresholds
mcp__agent-runtime__circuit_breaker_configure({
"agent_id": "researcher_agent",
"failure_threshold": 5,
"window_seconds": 60,
"cooldown_seconds": 300,
"fallback_agent": "generalist"
})
# Manually trip (emergency stop)
mcp__agent-runtime__circuit_breaker_trip({
"agent_id": "researcher_agent",
"reason": "Manual intervention - debugging required"
})
# Reset after fix
mcp__agent-runtime__circuit_breaker_reset({
"agent_id": "researcher_agent"
})
Cross-Session Resume
# Session 1: Create goal and start work
goal = mcp__agent-runtime__create_goal({"name": "Big Project", ...})
mcp__agent-runtime__decompose_goal({"goal_id": goal["id"]})
task1 = mcp__agent-runtime__get_next_task()
mcp__agent-runtime__update_task_status({"task_id": task1["id"], "status": "in_progress"})
# [Close Claude Code, restart later]
# Session 2: Resume exactly where left off
pending = mcp__agent-runtime__list_tasks({"status": "in_progress"})
# Returns: [task1] - still marked as in_progress
task1_updated = mcp__agent-runtime__update_task_status({
"task_id": task1["id"],
"status": "completed"
})
next_task = mcp__agent-runtime__get_next_task()
# Automatically gets task2 (next in dependency chain)
Requirements
- Python: 3.10+
- Dependencies:
anthropic-mcp(MCP SDK) - Storage:
~/.claude/agent_runtime.db(SQLite)
Database Schema
Tables in ~/.claude/agent_runtime.db:
goals- High-level goals with status and metadatatasks- Individual tasks with dependencies, priority, resultstask_queue- Queue position and scheduling inforelay_pipelines- Relay race pipeline definitions (God Agent Phase 2)relay_batons- Baton state for pipeline stepscircuit_breakers- Circuit breaker state and history (God Agent Phase 5)
Decomposition Strategies
Sequential
Task 1 → Task 2 → Task 3 → Task 4 → Task 5
Each task depends on previous. Linear execution.
Parallel
Task 1 (Backend) ─┐
Task 2 (Frontend) ─┼─→ All run simultaneously
Task 3 (Testing) ─┘
No dependencies. Maximum parallelism.
Hierarchical
Phase 1 (Foundation)
↓
Phase 2 (Core Implementation)
↓
Phase 3 (Integration)
↓
Phase 4 (Optimization)
Large phases that can be further decomposed.
Testing
# Run test suite
python3 test_agent_runtime.py
# Test relay protocol
python3 test_relay_protocol.py
# Test circuit breaker
python3 test_circuit_breaker.py
God Agent Integration
Phase 2: Relay Race Protocol
- 48-agent sequential pipelines
- Structured baton passing with context
- Quality gates at each step
- L-Score tracking for output quality
- Single-step retry (no full restart)
Phase 5: Circuit Breaker (Tiny Dancer)
- Automatic failure detection
- State machine: CLOSED → OPEN → HALF_OPEN → CLOSED
- Configurable thresholds and cooldowns
- Fallback agent routing
- Recovery monitoring
Links
- GitHub: https://github.com/marc-shade/agent-runtime-mcp
- Issues: https://github.com/marc-shade/agent-runtime-mcp/issues
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.