Mimir Agent (mimir-mcp-agent)

Multi-agent self-improving system with shared experiential memory, web search, and parallel agent execution.

Mimir is a local-first, MCP-native platform where specialized AI agents collaborate, execute tasks in parallel, learn from every outcome, and collectively become more effective over time. Zero configuration required — uses your editor's LLM via MCP sampling protocol.

Quick Start

Zero install (recommended)

npx mimir-mcp-agent

This downloads and starts the MCP server on stdio. Configure your editor (see below) and restart.

From source

git clone https://github.com/Cairolj/mimir-agent
cd mimir
npm install
npm run build
node dist/cli.js

MCP Configuration

Choose your approach:

• npx mimir-mcp-agent — run instantly, no local setup (auto-caches after first download)
• node <path>/dist/cli.js — local development, no network dependency

Windows users: If .js files open in an editor (Notepad++) instead of running, update to [email protected] which includes a proper Node.js shebang. Clear %APPDATA%\npm-cache\_npx\ if you see the old behavior.

Windsurf

File: ~/.codeium/windsurf/mcp_config.json (Windows: %USERPROFILE%\.codeium\windsurf\mcp_config.json)

{
  "mcpServers": {
    "mimir": {
      "command": "npx",
      "args": ["mimir-mcp-agent"],
      "env": {
        "MIMIR_DB_PATH": "%USERPROFILE%\\.mimir\\memory.db"
      }
    }
  }
}

Local dev alternative:

{
  "mcpServers": {
    "mimir": {
      "command": "node",
      "args": ["C:\\Users\\you\\projects\\mimir\\dist\\cli.js"],
      "env": {
        "MIMIR_DB_PATH": "C:\\Users\\you\\.mimir\\memory.db"
      }
    }
  }
}

VS Code

File: %APPDATA%\Code\User\mcp.json or via Command Palette → MCP: Configure MCP Servers

{
  "servers": {
    "mimir": {
      "command": "npx",
      "args": ["mimir-mcp-agent"]
    }
  }
}

Claude Code

File: ~/.claude/settings.json

{
  "mcpServers": {
    "mimir": {
      "command": "npx",
      "args": ["mimir-mcp-agent"]
    }
  }
}

Cursor

File: ~/.cursor/mcp.json

{
  "mcpServers": {
    "mimir": {
      "command": "npx",
      "args": ["mimir-mcp-agent"]
    }
  }
}

Smoke Test

Paste this into your editor's chat to verify Mimir works:

Test that Mimir is working correctly. Run these steps in order:

1. Use mimir_list_agents to show available agents
2. Use mimir_get_stats to show learning statistics
3. Use mimir_run_task to run echo "hello from Mimir agents"
4. Use mimir_get_advice with description "run a shell command"
5. Use mimir_submit_task to record "test completed successfully" with context {"test": "smoke", "result": "ok"}

Show the result of each step.

Features

• Zero-config LLM — uses your editor's LLM via MCP sampling/createMessage. No API keys, no .env.
• Web search — Investigator agent searches the web via DuckDuckGo Instant Answer API. No API key required.
• Parallel execution — agents run in waves using a dependency graph. Independent steps execute concurrently.
• Experiential memory — every task outcome is stored in SQLite. Future tasks retrieve the best strategy from past experiences.
• Plugin-based — agents are registered as tool lists. Extensible by design.
• Local-first — all data stays on your machine. No cloud, no telemetry.
• MCP-native — integrates with Windsurf, VS Code, Claude Code, Cursor, and any MCP-compatible editor.

MCP Tools

| Tool | Description | |---|---| | mimir_run_task | Run a task through the full agent orchestration pipeline | | mimir_get_advice | Get the best strategy from past similar experiences | | mimir_list_agents | List all available agent types and their tools | | mimir_get_stats | View learning statistics and success rates | | mimir_submit_task | Submit a task outcome — the system records and learns | | mimir_spawn_agents | Dynamically spawn agent instances for parallel work |

mimir_run_task

Executes a complete agent pipeline:

1. Planner — decomposes the description into executable steps (uses your editor's LLM via MCP sampling)
2. Executor — runs shell commands via child_process.exec
3. Investigator — fetches HTTP URLs or web searches (prefix with search:)
4. Validator — checks output for errors and warnings
5. Critic — reviews overall quality and assigns a score
6. Learning — everything is recorded in experiential memory

Dependency graph: the orchestrator builds a dependency graph from the plan and executes independent steps in parallel waves via Promise.all. Steps that depend on others wait until their dependencies resolve.

Example prompts:

"Use mimir_run_task to execute 'echo hello from agents'" "Use mimir_run_task to execute 'install express'" "Use mimir_run_task to execute 'https://example.com'" "Use mimir_run_task to execute 'search: latest news on AI agents'"

mimir_get_advice

Retrieves the best strategy from past similar experiences stored in SQLite. The more tasks you run, the better the advice becomes.

mimir_submit_task

Manually record a task outcome so the system can learn from it even without running the full pipeline:

{
  "description": "installed express successfully",
  "context": { "command": "npm install express", "result": "success" }
}

CLI

npx mimir-mcp-agent           # Start MCP server on stdio (default)
npx mimir-mcp-agent start     # Same
npx mimir-mcp-agent query <desc>    # Get advice from past experiences
npx mimir-mcp-agent submit <desc>   # Submit a task to learn from
npx mimir-mcp-agent stats          # Show learning statistics
npx mimir-mcp-agent agents         # List available agent types
npx mimir-mcp-agent run <command>  # Run a shell command directly (no LLM needed)
npx mimir-mcp-agent --help         # Show help

When installed globally:

mimir start       # Start MCP server
mimir run <cmd>   # Run a command directly
mimir query <d>   # Query memory
mimir submit <d>  # Submit task
mimir --help      # Show help

mimir run uses ExecutorHandler directly — no planning, no LLM, no server. Runs the command in a terminal and returns output. Useful for quick one-off commands.

Agent Types

| Agent | Tools | Behavior | |---|---|---| | Planner | decompose_task, resolve_deps, estimate_effort | Uses editor's LLM (MCP sampling) to decompose descriptions into executable steps. Falls back to keyword-based decomposition if sampling is unavailable. Can swap in Ollama/OpenAI/Anthropic via MIMIR_LLM_PROVIDER. | | Executor | git_clone, write_file, exec_cmd, read_file | Spawns child processes via child_process.exec. Runs shell commands, captures stdout/stderr, exit codes. | | Investigator | web_search, fetch_url, scrape | Fetches HTTP URLs via fetch. Detects search: prefix and delegates to DuckDuckGo Instant Answer API. Returns structured results (abstract, topics, infobox). No API key required. | | Validator | run_tests, lint_code, check_types | Scans output for error/warning patterns. Validates exit codes and output content. | | Critic | review_output, check_consistency, suggest_improvements | Reviews overall quality, assigns a score, and suggests improvements based on output analysis. |

Web Search

The Investigator agent can search the web using the DuckDuckGo Instant Answer API — zero configuration, no API key needed.

Prefix a task description with search: to trigger a web search:

mimir_run_task: "search: latest developments in AI agents 2026"

The search returns:

• Abstract — a concise summary when available
• Related topics — relevant subtopics with URLs
• Infobox — structured data (definitions, metrics, etc.)
• Images — associated images when present

Without search:, the Investigator fetches URLs directly via HTTP.

LLM Integration

Mimir uses a sampling architecture that requires no API keys by default:

Default: Editor's LLM (MCP Sampling)

The SamplingProvider calls your editor via MCP's sampling/createMessage. Whatever model your editor uses (Claude, GPT-4, Gemini, etc.) is what Mimir uses for task decomposition. Zero config, zero cost.

Optional: External Providers

Set MIMIR_LLM_PROVIDER to switch to an external LLM:

| Provider | Env Variable | |---|---| | ollama | OLLAMA_BASE_URL (default: http://localhost:11434) | | openai | OPENAI_API_KEY | | anthropic | ANTHROPIC_API_KEY |

set MIMIR_LLM_PROVIDER=ollama
npx mimir-mcp-agent

set MIMIR_LLM_PROVIDER=openai
set OPENAI_API_KEY=sk-...
npx mimir-mcp-agent

Architecture

User/Editor
    │
    ▼ (MCP protocol / stdio)
┌──────────────────────────────────┐
│        Mimir MCP Server           │
│                                   │
│  ┌─────────────────────────────┐  │
│  │      AgentOrchestrator       │  │
│  │  ┌───────┐ ┌───────────┐   │  │
│  │  │Planner│ │ Executor  │   │  │
│  │  ├───────┤ ├───────────┤   │  │
│  │  │Invest.│ │ Validator │   │  │
│  │  ├───────┤ ├───────────┤   │  │
│  │  │Critic │ │           │   │  │
│  │  └───────┘ └───────────┘   │  │
│  │  Dependency graph + waves   │  │
│  └─────────────────────────────┘  │
│                                   │
│  ┌─────────────────────────────┐  │
│  │      Learning Engine         │  │
│  │  ┌─────────┐ ┌───────────┐  │  │
│  │  │ SQLite  │ │ Experience │  │  │
│  │  │  Store  │ │   Graph    │  │  │
│  │  └─────────┘ └───────────┘  │  │
│  └─────────────────────────────┘  │
│                                   │
│  ┌─────────────────────────────┐  │
│  │      SamplingProvider        │  │
│  │  (MCP createMessage → LLM)   │  │
│  └─────────────────────────────┘  │
│                                   │
│  ┌─────────────────────────────┐  │
│  │     WebSearchService         │  │
│  │  (DuckDuckGo Instant Answer) │  │
│  └─────────────────────────────┘  │
│                                   │
│  ┌─────────────────────────────┐  │
│  │       Agent Registry         │  │
│  └─────────────────────────────┘  │
└──────────────────────────────────┘

Memory

All data is stored locally in SQLite. The memory layer has three components:

• Experience Store — raw task records with description, context, result, and outcome
• Experience Graph — relationships between experiences (similarity, dependency, sequence)
• Learning Engine — queries past experiences to find the best strategy for new tasks

Set MIMIR_DB_PATH to customize the database location:

set MIMIR_DB_PATH=C:\Users\youruser\.mimir\memory.db
npx mimir-mcp-agent

Default: ~/.mimir/experience.db

Tests

npm test                           # All tests (66 passing across 21 files)
npx vitest run tests/demo.test.ts  # Multi-agent demo (parallel execution)
npx vitest run --reporter=verbose  # Verbose output

The demo test runs 3 tasks concurrently and measures wall time vs sequential time, plus a multi-step pipeline with real command execution.

Development

npm install          # Install dependencies
npm run build        # Compile TypeScript to dist/
npm test             # Run all tests
npm run test:watch   # Watch mode

Requires Node.js 20+.

Windows Notes

• File associations: If npx opens cli.js in Notepad++ instead of running it, ensure you have [email protected] or later (includes Node.js shebang so npm generates proper .cmd shims). Clear %APPDATA%\npm-cache\_npx\ if cached from an older version.
• Paths: Use %USERPROFILE% in editor configs for portability.
• Environment variables: Use set VAR=value in cmd or $env:VAR = "value" in PowerShell.

Links

• npm
• GitHub
• Issues