Agent Never Give Up MCP

This is a MCP server that acts as a "escape guide" for AI coding agents.

It provides structured thinking protocols to help agents unstuck themselves without human help.

The diagram below illustrates when an agent should call this server and what it receives in return:

graph TD
    %% 1. The Normal Loop
    UserLoop["<b>User Query & Agent Loop</b><br/>(Claude Code / Windsurf / Cursor)<br/>Agent executing tasks..."]

    %% 2. The Problem (Triggers)
    subgraph Triggers ["Agent Gets Stuck (Trigger State)"]
        direction TB
        S1["Circular Reasoning"]
        S2["Bug Fix Fails"]
        S3["Missing Info"]
        S4["Analysis Paralysis"]
        S5["Unclear 'Done' State"]
    end

    %% Flow: Loop -> Stuck
    UserLoop -.->|"⚠️ Detects Issue<br/>(Inner Loop Stall)"| Triggers

    %% 3. The Solution (MCP Call)
    subgraph MCP ["Agent Never Give Up MCP"]
        direction TB
        T1(logic-is-too-complex)
        T2(bug-fix-always-failed)
        T3(missing-requirements)
        T4(analysis-too-long)
        T5(unclear-acceptance-criteria)
    end

    %% Mapping Triggers to Tools
    S1 -->|Call| T1
    S2 -->|Call| T2
    S3 -->|Call| T3
    S4 -->|Call| T4
    S5 -->|Call| T5

    %% 4. The Response
    MCP -->|Returns| R1["<b>Metacognitive Protocol</b><br/>Structured markdown to:<br/>1. Re-anchor goal<br/>2. Summarize failures<br/>3. Propose NEW strategy"]

    %% 5. The Outcome
    R1 --> Result["<b>Autonomous Recovery</b><br/>Agent unblocks itself &<br/>resumes execution<br/>(No Human Hand-off)"]

    %% Cycle back to loop
    Result -.->|"Resumes Loop"| UserLoop

    %% Styling
    classDef loop fill:#e1f5fe,stroke:#01579b,stroke-width:2px;
    classDef trigger fill:#ffebee,stroke:#b71c1c,stroke-width:2px;
    classDef tool fill:#fff3e0,stroke:#e65100,stroke-width:2px;
    classDef response fill:#e8f5e9,stroke:#1b5e20,stroke-width:2px;
    
    class UserLoop loop;
    class S1,S2,S3,S4,S5 trigger;
    class T1,T2,T3,T4,T5 tool;
    class R1,Result response;

Features

• Remote MCP server at /mcp endpoint (Streamable HTTP specification compliant)

• Two-tier scenario organization:

  • Core scenarios (auto-registered as direct MCP tools):

    • logic-is-too-complex – for circular reasoning or over-complicated logic

    • bug-fix-always-failed – for repeated failed bug fix attempts

    • missing-requirements – for unclear or missing requirements

    • lost-main-objective – for when current actions feel disconnected from the original goal

    • scope-creep-during-task – for when changes expand beyond the original task scope

    • long-goal-partially-done – for multi-step tasks where remaining work is forgotten

    • strategy-not-working – for when the same approach fails repeatedly

  • Extended scenarios (discovered via list_scenarios, accessed via get_prompt):

    • analysis-too-long – for excessive analysis time

    • unclear-acceptance-criteria – for undefined acceptance criteria

    • wrong-level-of-detail – for working at wrong abstraction level

    • constraints-cant-all-be-met – for conflicting requirements or constraints

    • blocked-by-environment-limits – for environmental blockers vs logic problems

• Discovery tools:

  • list_scenarios – list all scenarios with their tier (core/extended)

  • get_prompt – access any scenario (core or extended)

• Dual mode support: Each tool supports static and sampling modes

• Community-contributed prompts via markdown files

• Public and auth-less (v0)

• Cloudflare Workers deployment

Related MCP server: multivon-mcp

Configuration

Since agent-never-give-up is a cloud-hosted MCP server, no local installation is required. Simply add the server configuration to your preferred AI tool.

Install in Cursor

1. Open Cursor Settings > MCP.

2. Click + Add new global MCP server.

3. Use the following configuration (or edit your ~/.cursor/mcp.json file directly):

{
  "mcpServers": {
    "agent-never-give-up": {
      "type": "http",
      "url": "https://agent-never-give-up-mcp.askman.dev/mcp",
      "note": "A 'Swiss Army knife' toolset to help agents recover from getting stuck"
    }
  }
}

Install in Claude Desktop

To configure the server for Claude Desktop, edit the configuration file located at:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the following entry to the mcpServers object:

{
  "mcpServers": {
    "agent-never-give-up": {
      "type": "http",
      "url": "https://agent-never-give-up-mcp.askman.dev/mcp"
    }
  }
}

Install in Cline

1. Open Cline and click the MCP Servers icon (☰).

2. Select the Remote Servers tab (if available) or click Configure MCP Servers.

3. Edit the cline_mcp_settings.json file to include:

{
  "mcpServers": {
    "agent-never-give-up": {
      "type": "http",
      "url": "https://agent-never-give-up-mcp.askman.dev/mcp",
      "note": "A comprehensive suite of tools designed to keep agents persistent and unstuck"
    }
  }
}

Install in Windsurf

1. Open Windsurf.

2. Go to File > Settings > Configure MCP Servers (or edit ~/.codeium/windsurf/mcp_config.json).

3. Add the server configuration:

{
  "mcpServers": {
    "agent-never-give-up": {
      "type": "http",
      "url": "https://agent-never-give-up-mcp.askman.dev/mcp"
    }
  }
}

Development

To contribute to this project or run a local instance:

# Install dependencies
npm install

The local server will be available at http://localhost:8787/mcp.

Contributing Prompts

Prompts are organized in two tiers within the prompts/ directory:

prompts/
├── core/                           # Core scenarios (auto-registered as tools)
│   ├── logic-is-too-complex/
│   │   └── tool.md
│   ├── bug-fix-always-failed/
│   │   └── tool.md
│   └── missing-requirements/
│       └── tool.md
└── extended/                       # Extended scenarios (via list_scenarios + get_prompt)
    ├── analysis-too-long/
    │   └── tool.md
    └── unclear-acceptance-criteria/
        └── tool.md

Prompt File Format

Each tool.md file follows a simple markdown format with YAML frontmatter and a single protocol body:

---
name: scenario_name
title: "Scenario Title"
description: "When / why the agent should call this tool, from the agent's perspective"
---

When you notice [the trigger condition], follow this exact protocol step by step.

## 1. First step title

1. Action item one.
2. Action item two.
3. Action item three.

Keep it concrete.

## 2. Second step title

...

## 3. Third step title

...

Key principles:

• The description explains when to use the tool (the trigger condition)

• The body is a single protocol with numbered sections

• Each section has 2–6 concrete steps

• Focus on how to think, not domain-specific details

• No system prompt / user prompt template sections—just one actionable protocol

See prompts/AGENTS.md for detailed guidance on writing effective prompts.

Adding a New Scenario

Scenarios are auto-discovered from the prompts/ tree and the generated files in src/prompts/generated-scenarios.ts and src/types/generated-scenarios.ts. You do not need to manually edit TypeScript files—just add the prompt and regenerate.

Core scenarios (auto-registered as tools):

1. Create a new directory: prompts/core/{scenario_name}/ (must match /^[a-z0-9]+(-[a-z0-9]+)*$/).

2. Add a tool.md file following the format above.

3. Run npm run generate:scenarios to regenerate types and discovery data.

Extended scenarios (accessible via get_prompt):

1. Create a new directory: prompts/extended/{scenario_name}/ (same naming rules as above).

2. Add a tool.md file following the format above.

3. Run npm run generate:scenarios to regenerate types and discovery data.

Deploy

# Deploy to Cloudflare Workers
npm run deploy

After deployment, your MCP endpoint will be: https://agent-never-give-up-mcp.<account>.workers.dev/mcp

License

MIT