Create customizable AI assistants, automations, chat bots and agents that run 100% locally. No need for agentic Python libraries or cloud service keys, just bring your GPU (or even just CPU) and a web browser.

LocalAGI is a powerful, self-hostable AI Agent platform that allows you to design AI automations without writing code. Create Agents with a couple of clicks, connect via MCP and give it skills with skillserver. Every agent exposes a complete drop-in replacement for OpenAI's Responses APIs with advanced agentic capabilities. No clouds. No data leaks. Just pure local AI that works on consumer-grade hardware (CPU and GPU).

Are you tired of AI wrappers calling out to cloud APIs, risking your privacy? So were we.

LocalAGI ensures your data stays exactly where you want it—on your hardware. No API keys, no cloud subscriptions, no compromise.

• 🎛 No-Code Agents: Easy-to-configure multiple agents via Web UI.
• 🖥 Web-Based Interface: Simple and intuitive agent management.
• 🤖 Advanced Agent Teaming: Instantly create cooperative agent teams from a single prompt.
• 📡 Connectors: Built-in integrations with Discord, Slack, Telegram, GitHub Issues, and IRC.
• 🛠 Comprehensive REST API: Seamless integration into your workflows. Every agent created will support OpenAI Responses API out of the box.
• 📚 Short & Long-Term Memory: Powered by LocalRecall.
• 🧠 Planning & Reasoning: Agents intelligently plan, reason, and adapt.
• 🔄 Periodic Tasks: Schedule tasks with cron-like syntax.
• 💾 Memory Management: Control memory usage with options for long-term and summary memory.
• 🖼 Multimodal Support: Ready for vision, text, and more.
• 🔧 Extensible Custom Actions: Easily script dynamic agent behaviors in Go (interpreted, no compilation!).
• 🛠 Fully Customizable Models: Use your own models or integrate seamlessly with LocalAI.
• 📊 Observability: Monitor agent status and view detailed observable updates in real-time.

# Clone the repository
git clone https://github.com/mudler/LocalAGI
cd LocalAGI

# CPU setup (default)
docker compose up

# NVIDIA GPU setup
docker compose -f docker-compose.nvidia.yaml up

# Intel GPU setup (for Intel Arc and integrated GPUs)
docker compose -f docker-compose.intel.yaml up

# AMD GPU setup
docker compose -f docker-compose.amd.yaml up

# Start with a specific model (see available models in models.localai.io, or localai.io to use any model in huggingface)
MODEL_NAME=gemma-3-12b-it docker compose up

# NVIDIA GPU setup with custom multimodal and image models
MODEL_NAME=gemma-3-12b-it \
MULTIMODAL_MODEL=moondream2-20250414 \
IMAGE_MODEL=flux.1-dev-ggml \
docker compose -f docker-compose.nvidia.yaml up

Now you can access and manage your agents at http://localhost:8080

Still having issues? see this Youtube video: https://youtu.be/HtVwIxW3ePg

🆕 LocalAI is now part of a comprehensive suite of AI tools designed to work together:

	LocalAI LocalAI is the free, Open Source OpenAI alternative. LocalAI act as a drop-in replacement REST API that's compatible with OpenAI API specifications for local AI inferencing. Does not require GPU.
	LocalRecall A REST-ful API and knowledge base management system that provides persistent memory and storage capabilities for AI agents.

LocalAGI supports multiple hardware configurations through Docker Compose profiles:

• No special configuration needed
• Runs on any system with Docker
• Best for testing and development
• Supports text models only

• Requires NVIDIA GPU and drivers
• Uses CUDA for acceleration
• Best for high-performance inference
• Supports text, multimodal, and image generation models
• Run with: docker compose -f docker-compose.nvidia.yaml up
• Default models:
  • Text: gemma-3-4b-it-qat
  • Multimodal: moondream2-20250414
  • Image: sd-1.5-ggml
• Environment variables:
  • MODEL_NAME: Text model to use
  • MULTIMODAL_MODEL: Multimodal model to use
  • IMAGE_MODEL: Image generation model to use
  • LOCALAI_SINGLE_ACTIVE_BACKEND: Set to true to enable single active backend mode

• Supports Intel Arc and integrated GPUs
• Uses SYCL for acceleration
• Best for Intel-based systems
• Supports text, multimodal, and image generation models
• Run with: docker compose -f docker-compose.intel.yaml up
• Default models:
  • Text: gemma-3-4b-it-qat
  • Multimodal: moondream2-20250414
  • Image: sd-1.5-ggml
• Environment variables:
  • MODEL_NAME: Text model to use
  • MULTIMODAL_MODEL: Multimodal model to use
  • IMAGE_MODEL: Image generation model to use
  • LOCALAI_SINGLE_ACTIVE_BACKEND: Set to true to enable single active backend mode

You can customize the models used by LocalAGI by setting environment variables when running docker-compose. For example:

# CPU with custom model
MODEL_NAME=gemma-3-12b-it docker compose up

# NVIDIA GPU with custom models
MODEL_NAME=gemma-3-12b-it \
MULTIMODAL_MODEL=moondream2-20250414 \
IMAGE_MODEL=flux.1-dev-ggml \
docker compose -f docker-compose.nvidia.yaml up

# Intel GPU with custom models
MODEL_NAME=gemma-3-12b-it \
MULTIMODAL_MODEL=moondream2-20250414 \
IMAGE_MODEL=sd-1.5-ggml \
docker compose -f docker-compose.intel.yaml up

# With custom actions directory
LOCALAGI_CUSTOM_ACTIONS_DIR=/app/custom-actions docker compose up

If no models are specified, it will use the defaults:

• Text model: gemma-3-4b-it-qat
• Multimodal model: moondream2-20250414
• Image model: sd-1.5-ggml

Good (relatively small) models that have been tested are:

• qwen_qwq-32b (best in co-ordinating agents)
• gemma-3-12b-it
• gemma-3-27b-it

• ✓ Ultimate Privacy: No data ever leaves your hardware.
• ✓ Flexible Model Integration: Supports GGUF, GGML, and more thanks to LocalAI.
• ✓ Developer-Friendly: Rich APIs and intuitive interfaces.
• ✓ Effortless Setup: Simple Docker compose setups and pre-built binaries.
• ✓ Feature-Rich: From planning to multimodal capabilities, connectors for Slack, MCP support, LocalAGI has it all.

Explore detailed documentation including:

• Installation Options
• REST API Documentation
• Connector Configuration
• Agent Configuration

LocalAGI supports environment configurations. Note that these environment variables needs to be specified in the localagi container in the docker-compose file to have effect.

Download ready-to-run binaries from the Releases page.

Requirements:

• Go 1.20+
• Git
• Bun 1.2+

# Clone repo
git clone https://github.com/mudler/LocalAGI.git
cd LocalAGI

# Build it
cd webui/react-ui && bun i && bun run build
cd ../..
go build -o localagi

# Run it
./localagi

LocalAGI can be used as a Go library to programmatically create and manage AI agents. Let's start with a simple example of creating a single agent:

import (
    "github.com/mudler/LocalAGI/core/agent"
    "github.com/mudler/LocalAGI/core/types"
)

// Create a new agent with basic configuration
agent, err := agent.New(
    agent.WithModel("gpt-4"),
    agent.WithLLMAPIURL("http://localhost:8080"),
    agent.WithLLMAPIKey("your-api-key"),
    agent.WithSystemPrompt("You are a helpful assistant."),
    agent.WithCharacter(agent.Character{
        Name: "my-agent",
    }),
    agent.WithActions(
        // Add your custom actions here
    ),
    agent.WithStateFile("./state/my-agent.state.json"),
    agent.WithCharacterFile("./state/my-agent.character.json"),
    agent.WithTimeout("10m"),
    agent.EnableKnowledgeBase(),
    agent.EnableReasoning(),
)

if err != nil {
    log.Fatal(err)
}

// Start the agent
go func() {
    if err := agent.Run(); err != nil {
        log.Printf("Agent stopped: %v", err)
    }
}()

// Stop the agent when done
agent.Stop()

import (
    "github.com/mudler/LocalAGI/core/state"
    "github.com/mudler/LocalAGI/core/types"
)

// Create a new agent pool
pool, err := state.NewAgentPool(
    "default-model",           // default model name
    "default-multimodal-model", // default multimodal model
    "image-model",            // image generation model
    "http://localhost:8080",  // API URL
    "your-api-key",          // API key
    "./state",               // state directory
    "http://localhost:8081", // LocalRAG API URL
    func(config *AgentConfig) func(ctx context.Context, pool *AgentPool) []types.Action {
        // Define available actions for agents
        return func(ctx context.Context, pool *AgentPool) []types.Action {
            return []types.Action{
                // Add your custom actions here
            }
        }
    },
    func(config *AgentConfig) []Connector {
        // Define connectors for agents
        return []Connector{
            // Add your custom connectors here
        }
    },
    func(config *AgentConfig) []DynamicPrompt {
        // Define dynamic prompts for agents
        return []DynamicPrompt{
            // Add your custom prompts here
        }
    },
    func(config *AgentConfig) types.JobFilters {
        // Define job filters for agents
        return types.JobFilters{
            // Add your custom filters here
        }
    },
    "10m", // timeout
    true,  // enable conversation logs
)

// Create a new agent in the pool
agentConfig := &AgentConfig{
    Name: "my-agent",
    Model: "gpt-4",
    SystemPrompt: "You are a helpful assistant.",
    EnableKnowledgeBase: true,
    EnableReasoning: true,
    // Add more configuration options as needed
}

err = pool.CreateAgent("my-agent", agentConfig)

// Start all agents
err = pool.StartAll()

// Get agent status
status := pool.GetStatusHistory("my-agent")

// Stop an agent
pool.Stop("my-agent")

// Remove an agent
err = pool.Remove("my-agent")

LocalAGI provides two powerful ways to extend its functionality with custom actions:

LocalAGI supports custom actions written in Go that can be defined inline when creating an agent. These actions are interpreted at runtime, so no compilation is required.

You can also place custom Go action files in a directory and have LocalAGI automatically load them. Set the LOCALAGI_CUSTOM_ACTIONS_DIR environment variable to point to a directory containing your custom action files. Each .go file in this directory will be automatically loaded and made available to all agents.

Example setup:

# Set the environment variable
export LOCALAGI_CUSTOM_ACTIONS_DIR="/path/to/custom/actions"

# Or in docker-compose.yaml
environment:
  - LOCALAGI_CUSTOM_ACTIONS_DIR=/app/custom-actions

Directory structure:

custom-actions/
├── weather_action.go
├── file_processor.go
└── database_query.go

Each file should contain the three required functions (Run, Definition, RequiredFields) as described below.

When creating a new Agent, in the action sections select the "custom" action, you can add the Golang code directly there.

Custom actions in LocalAGI require three main functions:

1. Run(config map[string]interface{}) (string, map[string]interface{}, error) - The main execution function
2. Definition() map[string][]string - Defines the action's parameters and their types
3. RequiredFields() []string - Specifies which parameters are required

Note: You can't use additional modules, but just use libraries that are included in Go.

Here's a practical example of a custom action that fetches weather information:

import (
    "encoding/json"
    "fmt"
    "net/http"
    "io"
)

type WeatherParams struct {
    City    string `json:"city"`
    Country string `json:"country"`
}

type WeatherResponse struct {
    Main struct {
        Temp     float64 `json:"temp"`
        Humidity int     `json:"humidity"`
    } `json:"main"`
    Weather []struct {
        Description string `json:"description"`
    } `json:"weather"`
}

func Run(config map[string]interface{}) (string, map[string]interface{}, error) {
    // Parse parameters
    p := WeatherParams{}
    b, err := json.Marshal(config)
    if err != nil {
        return "", map[string]interface{}{}, err
    }
    if err := json.Unmarshal(b, &p); err != nil {
        return "", map[string]interface{}{}, err
    }

    // Make API call to weather service
    url := fmt.Sprintf("http://api.openweathermap.org/data/2.5/weather?q=%s,%s&appid=YOUR_API_KEY&units=metric", p.City, p.Country)
    resp, err := http.Get(url)
    if err != nil {
        return "", map[string]interface{}{}, err
    }
    defer resp.Body.Close()

    body, err := io.ReadAll(resp.Body)
    if err != nil {
        return "", map[string]interface{}{}, err
    }

    var weather WeatherResponse
    if err := json.Unmarshal(body, &weather); err != nil {
        return "", map[string]interface{}{}, err
    }

    // Format response
    result := fmt.Sprintf("Weather in %s, %s: %.1f°C, %s, Humidity: %d%%", 
        p.City, p.Country, weather.Main.Temp, weather.Weather[0].Description, weather.Main.Humidity)

    return result, map[string]interface{}{}, nil
}

func Definition() map[string][]string {
    return map[string][]string{
        "city": []string{
            "string",
            "The city name to get weather for",
        },
        "country": []string{
            "string", 
            "The country code (e.g., US, UK, DE)",
        },
    }
}

func RequiredFields() []string {
    return []string{"city", "country"}
}

Here's another example that demonstrates file system operations:

import (
    "encoding/json"
    "fmt"
    "os"
    "path/filepath"
)

type FileParams struct {
    Path    string `json:"path"`
    Action  string `json:"action"`
    Content string `json:"content,omitempty"`
}

func Run(config map[string]interface{}) (string, map[string]interface{}, error) {
    p := FileParams{}
    b, err := json.Marshal(config)
    if err != nil {
        return "", map[string]interface{}{}, err
    }
    if err := json.Unmarshal(b, &p); err != nil {
        return "", map[string]interface{}{}, err
    }

    switch p.Action {
    case "read":
        content, err := os.ReadFile(p.Path)
        if err != nil {
            return "", map[string]interface{}{}, err
        }
        return string(content), map[string]interface{}{}, nil
        
    case "write":
        err := os.WriteFile(p.Path, []byte(p.Content), 0644)
        if err != nil {
            return "", map[string]interface{}{}, err
        }
        return fmt.Sprintf("Successfully wrote to %s", p.Path), map[string]interface{}{}, nil
        
    case "list":
        files, err := os.ReadDir(p.Path)
        if err != nil {
            return "", map[string]interface{}{}, err
        }
        
        var fileList []string
        for _, file := range files {
            fileList = append(fileList, file.Name())
        }
        
        result, _ := json.Marshal(fileList)
        return string(result), map[string]interface{}{}, nil
        
    default:
        return "", map[string]interface{}{}, fmt.Errorf("unknown action: %s", p.Action)
    }
}

func Definition() map[string][]string {
    return map[string][]string{
        "path": []string{
            "string",
            "The file or directory path",
        },
        "action": []string{
            "string",
            "The action to perform: read, write, or list",
        },
        "content": []string{
            "string",
            "Content to write (required for write action)",
        },
    }
}

func RequiredFields() []string {
    return []string{"path", "action"}
}

To use custom actions, add them to your agent configuration:

1. Via Web UI: In the agent creation form, add a "Custom" action and paste your Go code
2. Via API: Include the custom action in your agent configuration JSON
3. Via Library: Add the custom action to your agent's actions list

LocalAGI supports both local and remote MCP servers, allowing you to extend functionality with external tools and services.

The Model Context Protocol (MCP) is a standard for connecting AI applications to external data sources and tools. LocalAGI can connect to any MCP-compliant server to access additional capabilities.

Local MCP servers run as processes that LocalAGI can spawn and communicate with via STDIO.

{
  "mcpServers": {
    "github": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "-e",
        "GITHUB_PERSONAL_ACCESS_TOKEN",
        "ghcr.io/github/github-mcp-server"
      ],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "<YOUR_TOKEN>"
      }
    }
  }
}

Remote MCP servers are HTTP-based and can be accessed over the network.

You can create MCP servers in any language that supports the MCP protocol and add the URLs of the servers to LocalAGI.

1. Via Web UI: In the MCP Settings section of agent creation, add MCP servers
2. Via API: Include MCP server configuration in your agent config

• Security: Always validate inputs and use proper authentication for remote MCP servers
• Error Handling: Implement robust error handling in your MCP servers
• Documentation: Provide clear descriptions for all tools exposed by your MCP server
• Testing: Test your MCP servers independently before integrating with LocalAGI
• Resource Management: Ensure your MCP servers properly clean up resources

The development workflow is similar to the source build, but with additional steps for hot reloading of the frontend:

# Clone repo
git clone https://github.com/mudler/LocalAGI.git
cd LocalAGI

cd webui/react-ui

# Install dependencies
bun i

# Compile frontend (the build directory needs to exist for the backend to start)
bun run build

# Start frontend development server
bun run dev

Then in separate terminal:

cd LocalAGI

# Create a "pool" directory for agent state
mkdir pool

# Set required environment variables
export LOCALAGI_MODEL=gemma-3-4b-it-qat
export LOCALAGI_MULTIMODAL_MODEL=moondream2-20250414
export LOCALAGI_IMAGE_MODEL=sd-1.5-ggml
export LOCALAGI_LLM_API_URL=http://localai:8080
export LOCALAGI_LOCALRAG_URL=http://localrecall:8080
export LOCALAGI_STATE_DIR=./pool
export LOCALAGI_TIMEOUT=5m
export LOCALAGI_ENABLE_CONVERSATIONS_LOGGING=false
export LOCALAGI_SSHBOX_URL=root:root@sshbox:22

# Start development server
go run main.go

Note: see webui/react-ui/.vite.config.js for env vars that can be used to configure the backend URL

Link your agents to the services you already use. Configuration examples below.

{
  "token": "YOUR_PAT_TOKEN",
  "repository": "repo-to-monitor",
  "owner": "repo-owner",
  "botUserName": "bot-username"
}

{
  "token": "Bot YOUR_DISCORD_TOKEN",
  "defaultChannel": "OPTIONAL_CHANNEL_ID"
}

{
  "botToken": "xoxb-your-bot-token",
  "appToken": "xapp-your-app-token"
}

{ 
  "token": "your-bot-father-token",
  "group_mode": "true",
  "mention_only": "true",
  "admins": "username1,username2"
}

{
  "server": "irc.example.com",
  "port": "6667",
  "nickname": "LocalAGIBot",
  "channel": "#yourchannel",
  "alwaysReply": "false"
}

{
  "smtpServer": "smtp.gmail.com:587",
  "imapServer": "imap.gmail.com:993",
  "smtpInsecure": "false",
  "imapInsecure": "false",
  "username": "user@gmail.com",
  "email": "user@gmail.com",
  "password": "correct-horse-battery-staple",
  "name": "LogalAGI Agent"
}

curl -X GET "http://localhost:3000/api/agents"

curl -X GET "http://localhost:3000/api/agent/my-agent/status"

curl -X POST "http://localhost:3000/api/agent/create" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-agent",
    "model": "gpt-4",
    "system_prompt": "You are an AI assistant.",
    "enable_kb": true,
    "enable_reasoning": true
  }'

curl -X DELETE "http://localhost:3000/api/agent/my-agent"

curl -X PUT "http://localhost:3000/api/agent/my-agent/pause"

curl -X PUT "http://localhost:3000/api/agent/my-agent/start"

curl -X GET "http://localhost:3000/api/agent/my-agent/config"

curl -X PUT "http://localhost:3000/api/agent/my-agent/config" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4",
    "system_prompt": "You are an AI assistant."
  }'

curl -X GET "http://localhost:3000/settings/export/my-agent" --output my-agent.json

curl -X POST "http://localhost:3000/settings/import" \
  -F "file=@/path/to/my-agent.json"

curl -X POST "http://localhost:3000/api/chat/my-agent" \
  -H "Content-Type: application/json" \
  -d '{"message": "Hello, how are you today?"}'

curl -X POST "http://localhost:3000/api/notify/my-agent" \
  -H "Content-Type: application/json" \
  -d '{"message": "Important notification"}'

curl -N -X GET "http://localhost:3000/api/sse/my-agent"

curl -X GET "http://localhost:3000/api/meta/agent/config"

{
  "name": "my-agent",
  "model": "gpt-4",
  "multimodal_model": "gpt-4-vision",
  "hud": true,
  "standalone_job": false,
  "random_identity": false,
  "initiate_conversations": true,
  "enable_planning": true,
  "identity_guidance": "You are a helpful assistant.",
  "periodic_runs": "0 * * * *",
  "permanent_goal": "Help users with their questions.",
  "enable_kb": true,
  "enable_reasoning": true,
  "kb_results": 5,
  "can_stop_itself": false,
  "system_prompt": "You are an AI assistant.",
  "long_term_memory": true,
  "summary_long_term_memory": false
}

MIT License — See the LICENSE file for details.

LOCAL PROCESSING. GLOBAL THINKING.
Made with ❤️ by mudler