diff --git a/README.md b/README.md index ea91ebd..5908022 100644 --- a/README.md +++ b/README.md @@ -46,8 +46,9 @@ Instead of spending days wiring together LLMs, tools, and execution environments ## 📑 Table of Contents - [🧰 Available Out of the Box](#-available-out-of-the-box) - [🤖 Agents](#-agents) - - [📦 Local Tools (Zero External Dependencies)](#-local-tools-zero-external-dependencies) - - [🌐 MCP Servers (Context Superpowers)](#-mcp-servers-context-superpowers) + - [📦 Local Tools](#-local-tools) + - [🌐 MCP Servers](#-mcp-servers) + - [🧠 LLM Providers](#-llm-providers) - [🚀 Quick Start (Zero to Agent in 60s)](#-quick-start-zero-to-agent-in-60s) - [🛠️ Build Your Own Agent](#️-build-your-own-agent) - [🏗️ Architecture](#️-architecture) @@ -62,293 +63,73 @@ Instead of spending days wiring together LLMs, tools, and execution environments ### 🤖 Agents - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
AgentPurposeMCP ServersLocal Tools
developerCode Master: Read, search & edit code.webfetchAll codebase tools below
travel-coordinatorTrip Planner: Orchestrates agents.kiwi-com-flight-search
webfetch		Uses 3 sub-agents
chefChef: Recipes from your fridge.webfetch-
newsNews Anchor: Aggregates top stories.webfetch-
travelFlight Booker: Finds the best routes.kiwi-com-flight-search-
simpleChat Buddy: Vanilla conversational agent.--
github-pr-reviewerPR Reviewer: Reviews diffs, posts inline comments & summaries.- -
- View tools - get_pr_diff
- get_pr_comments
- post_review_comment
- post_general_comment
- reply_to_review_comment
- get_pr_metadata -
-
whatsappWhatsApp Agent: Bidirectional WhatsApp communication (personal account).webfetch
duckduckgo-search		-
+The framework includes several pre-built agents for common use cases: -
-📱 WhatsApp Agent Setup +| Agent | Purpose | +|-------|---------| +| `developer` | Code Master: Read, search & edit code | +| `travel-coordinator` | Trip Planner: Orchestrates agents | +| `chef` | Chef: Recipes from your fridge | +| `news` | News Anchor: Aggregates top stories | +| `travel` | Flight Booker: Finds the best routes | +| `simple` | Chat Buddy: Vanilla conversational agent | +| `github-pr-reviewer` | PR Reviewer: Reviews diffs, posts inline comments & summaries | +| `whatsapp` | WhatsApp Agent: Bidirectional WhatsApp communication | -The WhatsApp agent enables bidirectional communication through your personal WhatsApp account using QR code authentication. +📖 **See [docs/agents.md](docs/agents.md)** for detailed information about each agent, including configuration options and usage examples. -**Requirements:** -- Go 1.21+ and Git (for WhatsApp backend) -- Python 3.13+ -- A configured LLM provider (see environment variables below) +--- -**Configuration:** -```bash -# 1. Copy example config -cp agentic-framework/config/whatsapp.yaml.example agentic-framework/config/whatsapp.yaml - -# 2. Edit config/whatsapp.yaml with your settings: -# - model: "claude-sonnet-4-6" # Your LLM model -# - privacy.allowed_contact: "+34 666 666 666" # Your phone number (only this number can interact) -# - channel.storage_path: "~/storage/whatsapp" # Where to store session data -# - mcp_servers: ["web-fetch", "duckduckgo-search"] # Optional: MCP servers to use -``` +### 📦 Local Tools -**Usage:** -```bash -# Start the WhatsApp agent -bin/agent.sh whatsapp --config config/whatsapp.yaml +Fast, zero-dependency tools for working with local codebases: -# With custom settings (overrides config file) -bin/agent.sh whatsapp --allowed-contact "+1234567890" --storage ~/custom/path +| Tool | Capability | +|------|------------| +| `find_files` | Fast search via `fd` | +| `discover_structure` | Directory tree mapping | +| `get_file_outline` | AST signature parsing | +| `read_file_fragment` | Precise file reading | +| `code_search` | Fast search via `ripgrep` | +| `edit_file` | Safe file editing | -# Customize MCP servers -bin/agent.sh whatsapp --mcp-servers "web-fetch,duckduckgo-search" -bin/agent.sh whatsapp --mcp-servers none # Disable MCP +📖 **See [docs/tools.md](docs/tools.md)** for detailed documentation of each tool, including parameters and examples. -# Verbose mode for debugging -bin/agent.sh whatsapp --verbose -``` +--- -**First Run:** -1. Scan the QR code displayed in your terminal -2. Wait for WhatsApp to authenticate -3. Send a message from your configured phone number -4. Agent will respond automatically - -**Privacy & Security:** -- 🔒 Only processes messages from the configured contact -- 🔒 Group chat messages are automatically filtered (not sent to LLM) -- 🔒 All data stored locally (no cloud storage of conversations) -- 🔒 Messages from other contacts are silently ignored -- 🔒 Message deduplication prevents reprocessing - -**Configuration Options:** -- `model`: LLM model to use (defaults to provider default) -- `mcp_servers`: MCP servers for web search and content fetching -- `privacy.allowed_contact`: Only this phone number can interact with the agent -- `privacy.log_filtered_messages`: Log filtered messages for debugging -- `channel.storage_path`: Directory for WhatsApp session and database files -- `features.group_messages`: Currently disabled by default for privacy +### 🌐 MCP Servers -
+Model Context Protocol servers for extending agent capabilities: -### 📦 Local Tools (Zero External Dependencies) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ToolCapabilityExample
find_filesFast search via fd*.py finds Python files
discover_structureDirectory tree mappingUnderstands project layout
get_file_outlineAST signature parsing (Python, TS, Go, Rust, Java, C++, PHP)Extracts classes/functions
read_file_fragmentPrecise file readingfile.py:10:50
code_searchFast search via ripgrepGlobal regex search
edit_fileSafe file editingInserts/Replaces lines
+| Server | Purpose | +|--------|---------| +| `kiwi-com-flight-search` | Search real-time flights | +| `webfetch` | Extract clean text from URLs & web search | +| `duckduckgo-search` | Web search via DuckDuckGo | -
-📝 Advanced: edit_file Formats +📖 **See [docs/mcp-servers.md](docs/mcp-servers.md)** for details on each server and how to add custom MCP servers. -**RECOMMENDED: `search_replace` (no line numbers needed)** -```json -{"op": "search_replace", "path": "file.py", "old": "exact text", "new": "replacement text"} -``` +--- -**Line-based operations:** -`replace:path:start:end:content` | `insert:path:after_line:content` | `delete:path:start:end` +### 🧠 LLM Providers -
+The framework supports **11 LLM providers** out of the box, covering 90%+ of the market: -### 🌐 MCP Servers (Context Superpowers) - - - - - - - - - - - - - - - - - - - - - -
ServerPurposeAPI Key Needed?
kiwi-com-flight-searchSearch real-time flights🟢 No
webfetchExtract clean text from URLs & web search🟢 No
+| Provider | Type | Use Case | +|----------|------|----------| +| **Anthropic** | Cloud | State-of-the-art reasoning (Claude) | +| **OpenAI** | Cloud | GPT-4, GPT-4.1, o1 series | +| **Azure OpenAI** | Cloud | Enterprise OpenAI deployments | +| **Google GenAI** | Cloud | Gemini models via API | +| **Google Vertex AI** | Cloud | Gemini models via GCP | +| **Groq** | Cloud | Ultra-fast inference | +| **Mistral AI** | Cloud | European privacy-focused models | +| **Cohere** | Cloud | Enterprise RAG and Command models | +| **AWS Bedrock** | Cloud | Anthropic, Titan, Meta via AWS | +| **Ollama** | Local | Run LLMs locally (zero API cost) | +| **Hugging Face** | Cloud | Open models from Hugging Face Hub | ---- - -### 🧠 Supported LLM Providers - -The framework supports **10+ LLM providers** out of the box, covering 90%+ of the LLM market: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ProviderTypeUse Case
AnthropicCloudState-of-the-art reasoning (Claude)
OpenAICloudGPT-4, GPT-4.1, o1 series
Azure OpenAICloudEnterprise OpenAI deployments
Google GenAICloudGemini models via API
Google Vertex AICloudGemini models via GCP
GroqCloudUltra-fast inference
Mistral AICloudEuropean privacy-focused models
CohereCloudEnterprise RAG and Command models
AWS BedrockCloudAnthropic, Titan, Meta via AWS
OllamaLocalRun LLMs locally (zero API cost)
Hugging FaceCloudOpen models from Hugging Face Hub
- -**Provider Priority:** Anthropic > Google Vertex > Google GenAI > Azure > Groq > Mistral > Cohere > Bedrock > HuggingFace > Ollama > OpenAI (fallback) +📖 **See [docs/llm-providers.md](docs/llm-providers.md)** for detailed setup instructions, environment variables, and provider comparison. --- @@ -409,89 +190,43 @@ bin/agent.sh chef -i "I have chicken, rice, and soy sauce. What can I make?"
🔑 Required Environment Variables - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ProviderVariableRequired?Default Model
AnthropicANTHROPIC_API_KEY🟢 Yes*claude-haiku-4-5-20251001
OpenAIOPENAI_API_KEY🟢 Yes*gpt-4o-mini
Azure OpenAIAZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT⚪ Nogpt-4o-mini
Google GenAIGOOGLE_API_KEY⚪ Nogemini-2.0-flash-exp
Google Vertex AIGOOGLE_VERTEX_PROJECT_ID⚪ Nogemini-2.0-flash-exp
GroqGROQ_API_KEY⚪ Nollama-3.3-70b-versatile
Mistral AIMISTRAL_API_KEY⚪ Nomistral-large-latest
CohereCOHERE_API_KEY⚪ Nocommand-r-plus
AWS BedrockAWS_PROFILE or AWS_ACCESS_KEY_ID⚪ Noanthropic.claude-3-5-sonnet-20241022-v2:0
OllamaOLLAMA_BASE_URL⚪ Nollama3.2
Hugging FaceHUGGINGFACEHUB_API_TOKEN⚪ Nometa-llama/Llama-3.2-3B-Instruct
- -**Model Override Variables** (optional): -- `ANTHROPIC_MODEL_NAME`, `OPENAI_MODEL_NAME`, `AZURE_OPENAI_MODEL_NAME`, `GOOGLE_GENAI_MODEL_NAME`, `GROQ_MODEL_NAME`, etc. - -> ⚠️ **Note:** Only one provider's API key is required. The framework auto-detects which provider to use based on available credentials. +Only one provider's API key is required. The framework auto-detects which provider to use based on available credentials. + +```bash +# Anthropic (Recommended) +ANTHROPIC_API_KEY=sk-ant-your-key-here + +# OpenAI +OPENAI_API_KEY=sk-your-key-here + +# Google GenAI / Vertex +GOOGLE_API_KEY=your-google-key +GOOGLE_VERTEX_PROJECT_ID=your-project-id + +# Groq +GROQ_API_KEY=gsk-your-key-here + +# Mistral AI +MISTRAL_API_KEY=your-mistral-key-here + +# Cohere +COHERE_API_KEY=your-cohere-key-here + +# Azure OpenAI +AZURE_OPENAI_API_KEY=your-azure-key +AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com + +# AWS Bedrock +AWS_PROFILE=your-profile + +# Ollama (Local, no API key needed) +OLLAMA_BASE_URL=http://localhost:11434 + +# Hugging Face +HUGGINGFACEHUB_API_TOKEN=your-hf-token +``` + +📖 **See [docs/llm-providers.md](docs/llm-providers.md)** for detailed environment variable configurations, model overrides, and provider comparison.
@@ -633,7 +368,7 @@ bin/agent.sh developer -i "Hello" -v # 📜 Access logs (same location as local) tail -f agentic-framework/logs/agent.log -# 📱 Run the WhatsApp agent (requires config) +# 📱 Run the WhatsApp agent (requires config - see docs/agents.md) agentic-run whatsapp --config config/whatsapp.yaml # 📱 Run WhatsApp with custom settings diff --git a/docs/agents.md b/docs/agents.md new file mode 100644 index 0000000..9ee7e6c --- /dev/null +++ b/docs/agents.md @@ -0,0 +1,245 @@ +# Available Agents + +This document provides detailed information about all available agents in the Agentic Framework. + +## Agent Overview + +| Agent | Purpose | MCP Servers | Local Tools | +|-------|---------|-------------|-------------| +| `developer` | **Code Master:** Read, search & edit code. | `webfetch` | All codebase tools | +| `travel-coordinator` | **Trip Planner:** Orchestrates agents. | `kiwi-com-flight-search`, `webfetch` | Uses 3 sub-agents | +| `chef` | **Chef:** Recipes from your fridge. | `webfetch` | - | +| `news` | **News Anchor:** Aggregates top stories. | `webfetch` | - | +| `travel` | **Flight Booker:** Finds the best routes. | `kiwi-com-flight-search` | - | +| `simple` | **Chat Buddy:** Vanilla conversational agent. | - | - | +| `github-pr-reviewer` | **PR Reviewer:** Reviews diffs, posts inline comments & summaries. | - | Custom GitHub tools | +| `whatsapp` | **WhatsApp Agent:** Bidirectional WhatsApp communication (personal account). | `webfetch`, `duckduckgo-search` | - | + +--- + +## Developer Agent + +**Purpose:** The developer agent is a code master designed to read, search, and edit codebases. + +### Capabilities + +- **Codebase Exploration:** Navigate and understand project structures +- **Code Search:** Fast pattern matching using ripgrep +- **File Editing:** Safe modifications to source files +- **AST Analysis:** Understand code signatures in multiple languages + +### MCP Servers +- `webfetch` - For fetching documentation and web resources + +### Local Tools +- All available codebase tools (see [tools.md](tools.md) for details) + +### Usage Example +```bash +bin/agent.sh developer -i "Explain the architecture of this project" +bin/agent.sh developer -i "Find all functions that use the database connection" +``` + +--- + +## Travel Coordinator Agent + +**Purpose:** Orchestrates multiple agents to plan complex trips. + +### Capabilities + +- Coordinates with `travel` and `chef` sub-agents +- Manages complex travel planning workflows +- Integrates flight search with dining recommendations + +### MCP Servers +- `kiwi-com-flight-search` - Real-time flight data +- `webfetch` - For travel-related web content + +### Architecture +Uses 3 specialized sub-agents working together via LangGraph. + +### Usage Example +```bash +bin/agent.sh travel-coordinator -i "Plan a weekend trip from Madrid to Paris" +``` + +--- + +## Chef Agent + +**Purpose:** Suggests recipes based on ingredients you have available. + +### Capabilities + +- Recipe suggestions based on available ingredients +- Fetches recipes from online sources +- Adapts recipes to your ingredient constraints + +### MCP Servers +- `webfetch` - For fetching recipe information + +### Usage Example +```bash +bin/agent.sh chef -i "I have chicken, rice, and soy sauce. What can I make?" +``` + +--- + +## News Agent + +**Purpose:** Aggregates and summarizes top news stories. + +### Capabilities + +- Fetches latest news from multiple sources +- Summarizes key developments +- Filters by topics when requested + +### MCP Servers +- `webfetch` - For fetching news content + +### Usage Example +```bash +bin/agent.sh news -i "What are today's top tech stories?" +``` + +--- + +## Travel Agent + +**Purpose:** Finds the best flight routes for your journey. + +### Capabilities + +- Real-time flight search +- Route optimization +- Price comparison + +### MCP Servers +- `kiwi-com-flight-search` - Real-time flight data + +### Usage Example +```bash +bin/agent.sh travel -i "Find flights from Madrid to Barcelona next weekend" +``` + +--- + +## Simple Agent + +**Purpose:** A basic conversational agent for general chatting. + +### Capabilities + +- Natural conversation +- General knowledge (via LLM) +- No specialized tools + +### MCP Servers +- None + +### Usage Example +```bash +bin/agent.sh simple -i "Tell me a joke" +``` + +--- + +## GitHub PR Reviewer Agent + +**Purpose:** Automatically reviews pull requests by analyzing diffs and posting comments. + +### Capabilities + +- **Diff Analysis:** Reviews code changes line by line +- **Inline Comments:** Posts specific feedback on problematic code +- **Summary Comments:** Provides overall PR assessment +- **Thread Responses:** Engages in review conversations + +### Local Tools +- `get_pr_diff` - Retrieve the pull request diff +- `get_pr_comments` - Get existing comments on a PR +- `post_review_comment` - Post inline review comments +- `post_general_comment` - Post overall PR feedback +- `reply_to_review_comment` - Reply to review comment threads +- `get_pr_metadata` - Fetch PR metadata (title, author, etc.) + +### Usage Example +```bash +bin/agent.sh github-pr-reviewer -i "Review PR #123 for bugs and style issues" +``` + +--- + +## WhatsApp Agent + +The WhatsApp agent enables bidirectional communication through your personal WhatsApp account using QR code authentication. + +### Requirements +- Go 1.21+ and Git (for WhatsApp backend) +- Python 3.13+ +- A configured LLM provider + +### Configuration + +```bash +# 1. Copy example config +cp agentic-framework/config/whatsapp.yaml.example agentic-framework/config/whatsapp.yaml + +# 2. Edit config/whatsapp.yaml with your settings: +# - model: "claude-sonnet-4-6" # Your LLM model +# - privacy.allowed_contact: "+34 666 666 666" # Your phone number (only this number can interact) +# - channel.storage_path: "~/storage/whatsapp" # Where to store session data +# - mcp_servers: ["web-fetch", "duckduckgo-search"] # Optional: MCP servers to use +``` + +### Usage + +```bash +# Start the WhatsApp agent +bin/agent.sh whatsapp --config config/whatsapp.yaml + +# With custom settings (overrides config file) +bin/agent.sh whatsapp --allowed-contact "+1234567890" --storage ~/custom/path + +# Customize MCP servers +bin/agent.sh whatsapp --mcp-servers "web-fetch,duckduckgo-search" +bin/agent.sh whatsapp --mcp-servers none # Disable MCP + +# Verbose mode for debugging +bin/agent.sh whatsapp --verbose +``` + +### First Run +1. Scan the QR code displayed in your terminal +2. Wait for WhatsApp to authenticate +3. Send a message from your configured phone number +4. Agent will respond automatically + +### Privacy & Security +- 🔒 Only processes messages from the configured contact +- 🔒 Group chat messages are automatically filtered (not sent to LLM) +- 🔒 All data stored locally (no cloud storage of conversations) +- 🔒 Messages from other contacts are silently ignored +- 🔒 Message deduplication prevents reprocessing + +### Configuration Options +| Option | Description | +|--------|-------------| +| `model` | LLM model to use (defaults to provider default) | +| `mcp_servers` | MCP servers for web search and content fetching | +| `privacy.allowed_contact` | Only this phone number can interact with the agent | +| `privacy.log_filtered_messages` | Log filtered messages for debugging | +| `channel.storage_path` | Directory for WhatsApp session and database files | +| `features.group_messages` | Currently disabled by default for privacy | + +### MCP Servers +- `webfetch` - For fetching web content +- `duckduckgo-search` - For web search capabilities + +--- + +## Creating Custom Agents + +To create your own agent, see the [Build Your Own Agent](../README.md#️-build-your-own-agent) section in the main README. diff --git a/docs/llm-providers.md b/docs/llm-providers.md new file mode 100644 index 0000000..fa4b43d --- /dev/null +++ b/docs/llm-providers.md @@ -0,0 +1,273 @@ +# LLM Providers + +This document provides detailed information about all LLM providers supported by the Agentic Framework. + +## Provider Overview + +| Provider | Type | Use Case | API Key Required? | +|----------|------|----------|-------------------| +| **Anthropic** | Cloud | State-of-the-art reasoning (Claude) | Yes* | +| **OpenAI** | Cloud | GPT-4, GPT-4.1, o1 series | Yes* | +| **Azure OpenAI** | Cloud | Enterprise OpenAI deployments | No | +| **Google GenAI** | Cloud | Gemini models via API | No | +| **Google Vertex AI** | Cloud | Gemini models via GCP | No | +| **Groq** | Cloud | Ultra-fast inference | No | +| **Mistral AI** | Cloud | European privacy-focused models | No | +| **Cohere** | Cloud | Enterprise RAG and Command models | No | +| **AWS Bedrock** | Cloud | Anthropic, Titan, Meta via AWS | No | +| **Ollama** | Local | Run LLMs locally (zero API cost) | No | +| **Hugging Face** | Cloud | Open models from Hugging Face Hub | No | + +**Provider Priority:** Anthropic > Google Vertex > Google GenAI > Azure > Groq > Mistral > Cohere > Bedrock > HuggingFace > Ollama > OpenAI (fallback) + +--- + +## Environment Variables + +### Anthropic + +```bash +ANTHROPIC_API_KEY=sk-ant-your-key-here +ANTHROPIC_MODEL_NAME=claude-haiku-4-5-20251001 # Optional +``` + +**Default Model:** `claude-haiku-4-5-20251001` + +--- + +### OpenAI + +```bash +OPENAI_API_KEY=sk-your-key-here +OPENAI_MODEL_NAME=gpt-4o-mini # Optional +``` + +**Default Model:** `gpt-4o-mini` + +--- + +### Azure OpenAI + +```bash +AZURE_OPENAI_API_KEY=your-azure-key +AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com +AZURE_OPENAI_MODEL_NAME=gpt-4o-mini # Optional +AZURE_OPENAI_API_VERSION=2024-02-15-preview # Optional +``` + +**Default Model:** `gpt-4o-mini` + +--- + +### Google GenAI + +```bash +GOOGLE_API_KEY=your-google-key +GOOGLE_GENAI_MODEL_NAME=gemini-2.0-flash-exp # Optional +``` + +**Default Model:** `gemini-2.0-flash-exp` + +--- + +### Google Vertex AI + +```bash +GOOGLE_VERTEX_PROJECT_ID=your-project-id +GOOGLE_VERTEX_LOCATION=us-central1 # Optional +GOOGLE_VERTEX_MODEL_NAME=gemini-2.0-flash-exp # Optional +``` + +**Default Model:** `gemini-2.0-flash-exp` + +--- + +### Groq + +```bash +GROQ_API_KEY=gsk-your-key-here +GROQ_MODEL_NAME=llama-3.3-70b-versatile # Optional +``` + +**Default Model:** `llama-3.3-70b-versatile` + +--- + +### Mistral AI + +```bash +MISTRAL_API_KEY=your-mistral-key-here +MISTRAL_MODEL_NAME=mistral-large-latest # Optional +``` + +**Default Model:** `mistral-large-latest` + +--- + +### Cohere + +```bash +COHERE_API_KEY=your-cohere-key-here +COHERE_MODEL_NAME=command-r-plus # Optional +``` + +**Default Model:** `command-r-plus` + +--- + +### AWS Bedrock + +```bash +AWS_PROFILE=your-profile +# OR +AWS_ACCESS_KEY_ID=your-access-key +AWS_SECRET_ACCESS_KEY=your-secret-key +AWS_REGION=us-east-1 # Optional +BEDROCK_MODEL_NAME=anthropic.claude-3-5-sonnet-20241022-v2:0 # Optional +``` + +**Default Model:** `anthropic.claude-3-5-sonnet-20241022-v2:0` + +--- + +### Ollama + +```bash +OLLAMA_BASE_URL=http://localhost:11434 +OLLAMA_MODEL_NAME=llama3.2 # Optional +``` + +**Default Model:** `llama3.2` + +**Requirements:** Ollama must be running locally with the specified model available. + +--- + +### Hugging Face + +```bash +HUGGINGFACEHUB_API_TOKEN=your-hf-token +HUGGINGFACEHUB_MODEL_NAME=meta-llama/Llama-3.2-3B-Instruct # Optional +``` + +**Default Model:** `meta-llama/Llama-3.2-3B-Instruct` + +--- + +## Provider Comparison + +### Anthropic (Claude) +- **Pros:** Best reasoning capabilities, excellent code understanding, strong safety guardrails +- **Cons:** Higher cost per token +- **Best for:** Complex reasoning, code analysis, multi-step tasks + +### OpenAI (GPT) +- **Pros:** Widely available, good general capabilities +- **Cons:** Can be more expensive than alternatives +- **Best for:** General-purpose tasks, when Anthropic key not available + +### Google GenAI / Vertex AI (Gemini) +- **Pros:** Fast, good multimodal capabilities, competitive pricing +- **Cons:** API changes more frequently +- **Best for:** Cost-sensitive workloads, Google Cloud users + +### Groq +- **Pros:** Extremely fast inference (low latency) +- **Cons:** Uses open-source models (may have different quality profiles) +- **Best for:** Real-time applications, speed-critical tasks + +### Ollama +- **Pros:** Free (after initial compute), privacy, no API limits +- **Cons:** Requires local compute resources, quality depends on model +- **Best for:** Development, privacy-sensitive work, offline use + +--- + +## Setup Instructions + +### 1. Copy the Environment Template + +```bash +cp .env.example .env +``` + +### 2. Configure Your Provider + +Edit `.env` and add credentials for your preferred provider. You only need one provider configured: + +```bash +# Choose ONE of the following: + +# Anthropic (Recommended) +ANTHROPIC_API_KEY=sk-ant-your-key-here + +# OR OpenAI +OPENAI_API_KEY=sk-your-key-here + +# OR Google +GOOGLE_API_KEY=your-google-key + +# OR Groq +GROQ_API_KEY=gsk-your-key-here + +# OR Ollama (Local, no API key needed) +OLLAMA_BASE_URL=http://localhost:11434 +``` + +### 3. Verify Configuration + +```bash +# List available agents to verify the framework is working +bin/agent.sh list + +# Run a simple agent +bin/agent.sh simple -i "Hello, can you hear me?" +``` + +--- + +## Model Selection + +The framework uses provider priority to automatically select the best available provider. You can override this with model-specific environment variables: + +```bash +# Force a specific model +ANTHROPIC_MODEL_NAME=claude-sonnet-4-6 +OPENAI_MODEL_NAME=gpt-4o +GROQ_MODEL_NAME=llama-3.3-70b-versatile +``` + +Or use the model selection in the agent configuration: + +```python +# In your agent configuration or .env file +MODEL_NAME="claude-sonnet-4-6" # Will use the provider's model +``` + +--- + +## Switching Providers + +To switch between providers without changing code: + +1. Update your `.env` file with the new provider's credentials +2. Remove or comment out the old provider's credentials +3. Run your agent - it will auto-detect the active provider + +The framework automatically detects which provider to use based on available environment variables. + +--- + +## Troubleshooting + +### No provider detected +Ensure at least one provider's API key is set in your `.env` file. + +### Ollama connection refused +Make sure Ollama is running: `ollama serve` + +### Rate limits +Consider using a different provider or Ollama for local inference. + +### Model not found +Check the model name is correct for your provider and that the model is available in your region/subscription. diff --git a/docs/mcp-servers.md b/docs/mcp-servers.md new file mode 100644 index 0000000..b5d02fd --- /dev/null +++ b/docs/mcp-servers.md @@ -0,0 +1,148 @@ +# MCP Servers + +This document provides detailed information about the MCP (Model Context Protocol) servers available in the Agentic Framework. + +## Server Overview + +| Server | Purpose | API Key Needed? | +|--------|---------|-----------------| +| `kiwi-com-flight-search` | Search real-time flights | 🟢 No | +| `webfetch` | Extract clean text from URLs & web search | 🟢 No | +| `duckduckgo-search` | Web search via DuckDuckGo | 🟢 No | + +--- + +## kiwi-com-flight-search + +Real-time flight search integration using the Kiwi.com API. + +### Description +Provides access to live flight data including prices, schedules, and routing information. Used by the `travel` and `travel-coordinator` agents. + +### Capabilities +- Search flights between cities +- Get real-time pricing +- Compare different routes +- Find optimal travel dates + +### Example Usage +```python +# The travel agent automatically uses this server when querying flights +# Example agent input: +# "Find flights from Madrid to Barcelona for next weekend" +``` + +### Configuration +No API key required. The server connects directly to Kiwi.com's public API. + +--- + +## webfetch + +Fetches and extracts clean text from web pages and performs web searches. + +### Description +Retrieves web content and converts it to readable, clean text format. Also supports web search functionality for finding relevant pages. + +### Capabilities +- Fetch any URL and extract readable content +- Perform web searches +- Convert HTML to clean markdown/text +- Follow redirects automatically + +### Example Usage +```python +# Fetch a specific page +webfetch("https://example.com/article") + +# Web search (via the MCP server) +websearch("python best practices 2025") +``` + +### Configuration +No API key required. Uses publicly available web fetching and search APIs. + +--- + +## duckduckgo-search + +Web search integration using DuckDuckGo. + +### Description +Provides privacy-focused web search capabilities without tracking. + +### Capabilities +- Web search +- Instant answers +- No tracking or personalization +- Anonymous results + +### Example Usage +```python +# Perform a web search +duckduckgo_search("latest tech news") + +# Search for specific information +duckduckgo_search("how to install python on mac") +``` + +### Configuration +No API key required. Uses DuckDuckGo's public search API. + +--- + +## Adding Custom MCP Servers + +To add a new MCP server to the framework: + +### 1. Add Server Configuration + +Edit `src/agentic_framework/mcp/config.py`: + +```python +DEFAULT_MCP_SERVERS = { + "my-server": { + "command": "npx", + "args": ["-y", "@modelcontextprotocol/server-myserver"], + "env": { + # Server-specific environment variables + } + }, + # ... existing servers +} +``` + +### 2. Register Server with Agent + +```python +from agentic_framework.core.langgraph_agent import LangGraphMCPAgent +from agentic_framework.registry import AgentRegistry + +@AgentRegistry.register("my-agent", mcp_servers=["my-server"]) +class MyAgent(LangGraphMCPAgent): + @property + def system_prompt(self) -> str: + return "You have access to my custom MCP server." +``` + +### 3. Test the Server + +```bash +# Test with your agent +bin/agent.sh my-agent -i "Use my-server to do something" +``` + +--- + +## Using MCP Servers in Agents + +MCP servers are automatically loaded when specified in the agent registration: + +```python +@AgentRegistry.register("my-agent", mcp_servers=["webfetch", "duckduckgo-search"]) +class MyAgent(LangGraphMCPAgent): + # The agent now has access to web fetch and search capabilities + pass +``` + +The LLM automatically discovers available tools from the MCP servers and uses them as needed based on the conversation context. diff --git a/docs/tools.md b/docs/tools.md new file mode 100644 index 0000000..6133d58 --- /dev/null +++ b/docs/tools.md @@ -0,0 +1,227 @@ +# Local Tools + +This document provides detailed information about all local tools available in the Agentic Framework. + +## Tool Overview + +| Tool | Capability | Example | +|------|------------|---------| +| `find_files` | Fast search via `fd` | `*.py` finds Python files | +| `discover_structure` | Directory tree mapping | Understands project layout | +| `get_file_outline` | AST signature parsing (Python, TS, Go, Rust, Java, C++, PHP) | Extracts classes/functions | +| `read_file_fragment` | Precise file reading | `file.py:10:50` | +| `code_search` | Fast search via `ripgrep` | Global regex search | +| `edit_file` | Safe file editing | Inserts/Replaces lines | + +--- + +## find_files + +Fast file search using the `fd` tool. + +### Description +Recursively searches for files matching a pattern in the project directory. + +### Parameters +- `pattern`: Glob pattern to match (e.g., `*.py`, `test_*.ts`) + +### Example +```bash +# Find all Python files +find_files "*.py" + +# Find test files +find_files "test_*.py" +``` + +--- + +## discover_structure + +Maps the directory structure of the project. + +### Description +Provides a tree-like view of the project's directory structure, helping agents understand the codebase layout. + +### Parameters +- `path`: Optional starting path (defaults to project root) +- `max_depth`: Maximum depth to explore (default: 3) + +### Example +```bash +# Discover full project structure +discover_structure() + +# Discover specific directory +discover_structure("src/", max_depth=2) +``` + +--- + +## get_file_outline + +Extracts code signatures using AST parsing. + +### Description +Parses source files and extracts classes, functions, methods, and their signatures. Supports multiple languages: +- Python +- TypeScript/JavaScript +- Go +- Rust +- Java +- C++ +- PHP + +### Parameters +- `file_path`: Path to the file to analyze + +### Example +```bash +# Get outline of a Python file +get_file_outline("src/main.py") + +# Output example: +# class MyClass: +# def __init__(self, x: int) -> None +# def process(self) -> str +# def helper_function(data: list) -> dict +``` + +--- + +## read_file_fragment + +Reads specific portions of a file. + +### Description +Reads a range of lines from a file, allowing precise inspection of specific code sections. + +### Format +`file_path:start_line:end_line` + +### Parameters +- `file_path`: Path to the file +- `start_line`: Starting line number (1-indexed) +- `end_line`: Ending line number (inclusive) + +### Example +```bash +# Read lines 10-50 of a file +read_file_fragment("src/main.py:10:50") + +# Read a specific function (assuming it's on lines 25-40) +read_file_fragment("src/main.py:25:40") +``` + +--- + +## code_search + +Fast code search using `ripgrep`. + +### Description +Performs global regex-based searches across the codebase for patterns, function names, or specific code constructs. + +### Parameters +- `pattern`: Regular expression to search for +- `file_pattern`: Optional glob pattern to filter files (e.g., `*.py`) +- `context_lines`: Number of context lines to include (default: 2) + +### Example +```bash +# Search for a function name +code_search("def process_data") + +# Search in Python files only +code_search("database\.query", file_pattern="*.py") + +# Search with more context +code_search("TODO|FIXME", context_lines=5) +``` + +--- + +## edit_file + +Safe file editing with multiple operation modes. + +### Description +Modifies files with safe operations including insert, replace, and delete. Supports both line-based operations and search-replace mode. + +### Operation Modes + +#### RECOMMENDED: `search_replace` (no line numbers needed) + +Format: `{"op": "search_replace", "path": "file.py", "old": "exact text", "new": "replacement text"}` + +```json +{ + "op": "search_replace", + "path": "src/main.py", + "old": "def old_function():\n pass", + "new": "def new_function():\n return True" +} +``` + +#### Line-based Operations + +| Operation | Format | Description | +|-----------|--------|-------------| +| `replace` | `replace:path:start:end:content` | Replace lines start through end with content | +| `insert` | `insert:path:after_line:content` | Insert content after specified line | +| `delete` | `delete:path:start:end` | Delete lines start through end | + +### Examples + +```bash +# Search and replace (recommended) +edit_file('{"op": "search_replace", "path": "src/main.py", "old": "old_value", "new": "new_value"}') + +# Replace specific lines +edit_file("replace:src/main.py:10:15:print('new code')") + +# Insert after line 20 +edit_file("insert:src/main.py:20:\n# New comment\nprint('hello')") + +# Delete lines 30-35 +edit_file("delete:src/main.py:30:35") +``` + +### Safety Features +- Validates file exists before editing +- Checks line numbers are within bounds +- Preserves file permissions +- Creates backup if configured + +--- + +## Creating Custom Tools + +To add your own local tools to an agent: + +```python +from langchain_core.tools import StructuredTool +from agentic_framework.core.langgraph_agent import LangGraphMCPAgent +from agentic_framework.registry import AgentRegistry + +@AgentRegistry.register("my-agent") +class MyAgent(LangGraphMCPAgent): + @property + def system_prompt(self) -> str: + return "You are my custom agent." + + def local_tools(self) -> list: + return [ + StructuredTool.from_function( + func=self.my_tool, + name="my_tool", + description="Description of what your tool does", + ) + ] + + def my_tool(self, input_data: str) -> str: + # Your tool logic here + return f"Processed: {input_data}" +``` + +For more details, see the [Build Your Own Agent](../README.md#️-build-your-own-agent) section in the main README.