Skip to main content
Glama
README.md
# 🐑 Shepherd MCP

MCP (Model Context Protocol) server for Shepherd - Debug your AI agents like you debug your code.

This MCP server allows AI assistants (Claude, Cursor, etc.) to query and analyze your AI agent sessions from multiple observability providers.

## Supported Providers

- **AIOBS** (Shepherd backend) - Native Shepherd observability
- **Langfuse** - Open-source LLM observability platform

## Installation

```bash
pip install shepherd-mcp
```

Or run directly with uvx:

```bash
uvx shepherd-mcp
```

## Configuration

### Environment Variables

#### AIOBS (Shepherd)

- `AIOBS_API_KEY` (required) - Your Shepherd API key
- `AIOBS_ENDPOINT` (optional) - Custom API endpoint URL

#### Langfuse

- `LANGFUSE_PUBLIC_KEY` (required) - Your Langfuse public API key
- `LANGFUSE_SECRET_KEY` (required) - Your Langfuse secret API key
- `LANGFUSE_HOST` (optional) - Custom Langfuse host URL (defaults to cloud.langfuse.com)

### .env File Support

shepherd-mcp automatically loads `.env` files from the current directory or any parent directory. This means if you have a `.env` file in your project root:

```bash
# .env
# AIOBS
AIOBS_API_KEY=aiobs_sk_xxxx

# Langfuse
LANGFUSE_PUBLIC_KEY=pk-lf-xxxx
LANGFUSE_SECRET_KEY=sk-lf-xxxx
LANGFUSE_HOST=https://cloud.langfuse.com
```

It will be automatically loaded when the MCP server starts.

### Claude Desktop

Add to your `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "shepherd": {
      "command": "uvx",
      "args": ["shepherd-mcp"],
      "env": {
        "AIOBS_API_KEY": "aiobs_sk_xxxx",
        "LANGFUSE_PUBLIC_KEY": "pk-lf-xxxx",
        "LANGFUSE_SECRET_KEY": "sk-lf-xxxx",
        "LANGFUSE_HOST": "https://cloud.langfuse.com"
      }
    }
  }
}
```

### Cursor

Add to your `.cursor/mcp.json`:

```json
{
  "mcpServers": {
    "shepherd": {
      "command": "uvx",
      "args": ["shepherd-mcp"],
      "env": {
        "AIOBS_API_KEY": "aiobs_sk_xxxx",
        "LANGFUSE_PUBLIC_KEY": "pk-lf-xxxx",
        "LANGFUSE_SECRET_KEY": "sk-lf-xxxx",
        "LANGFUSE_HOST": "https://cloud.langfuse.com"
      }
    }
  }
}
```

Or if installed via pip:

```json
{
  "mcpServers": {
    "shepherd": {
      "command": "shepherd-mcp",
      "env": {
        "AIOBS_API_KEY": "aiobs_sk_xxxx",
        "LANGFUSE_PUBLIC_KEY": "pk-lf-xxxx",
        "LANGFUSE_SECRET_KEY": "sk-lf-xxxx",
        "LANGFUSE_HOST": "https://cloud.langfuse.com"
      }
    }
  }
}
```

## Available Tools

### AIOBS (Shepherd) Tools

#### `aiobs_list_sessions`

List all AI agent sessions from Shepherd.

**Parameters:**
- `limit` (optional): Maximum number of sessions to return

**Example prompt:**
> "List my recent AI agent sessions from AIOBS"

#### `aiobs_get_session`

Get detailed information about a specific session including the full trace tree, LLM calls, function events, and evaluations.

**Parameters:**
- `session_id` (required): The UUID of the session to retrieve

**Example prompt:**
> "Get AIOBS session details for abc123-def456"

#### `aiobs_search_sessions`

Search and filter sessions with multiple criteria.

**Parameters:**
- `query` (optional): Text search (matches name, ID, labels, metadata)
- `labels` (optional): Filter by labels as key-value pairs
- `provider` (optional): Filter by LLM provider (e.g., 'openai', 'anthropic')
- `model` (optional): Filter by model name (e.g., 'gpt-4o-mini', 'claude-3')
- `function` (optional): Filter by function name
- `after` (optional): Sessions started after date (YYYY-MM-DD)
- `before` (optional): Sessions started before date (YYYY-MM-DD)
- `has_errors` (optional): Only return sessions with errors
- `evals_failed` (optional): Only return sessions with failed evaluations
- `limit` (optional): Maximum number of sessions to return

**Example prompts:**
> "Find all AIOBS sessions that used OpenAI with errors"
> "Search for sessions from yesterday that failed evaluations"

#### `aiobs_diff_sessions`

Compare two sessions and show their differences including:
- **Metadata**: Duration, labels, timestamps
- **LLM calls**: Count, tokens (input/output/total), average latency, errors
- **Provider/Model distribution**: Which providers and models were used
- **Function events**: Total calls, unique functions, function-specific counts
- **Trace structure**: Trace depth, root nodes
- **Evaluations**: Pass/fail counts and rates
- **System prompts**: Compare system prompts across sessions
- **Request parameters**: Temperature, max_tokens, tools used
- **Response content**: Content length, tool calls, stop reasons

**Parameters:**
- `session_id_1` (required): First session UUID to compare
- `session_id_2` (required): Second session UUID to compare

**Example prompt:**
> "Compare AIOBS sessions abc123 and def456"

---

### Langfuse Tools

#### `langfuse_list_traces`

List traces with pagination and filters. Traces represent complete workflows or conversations.

**Parameters:**
- `limit` (optional): Maximum results per page (default: 50)
- `page` (optional): Page number (1-indexed)
- `user_id` (optional): Filter by user ID
- `name` (optional): Filter by trace name
- `session_id` (optional): Filter by session ID
- `tags` (optional): Filter by tags
- `from_timestamp` (optional): Filter after timestamp
- `to_timestamp` (optional): Filter before timestamp

**Example prompt:**
> "List the last 20 Langfuse traces"

#### `langfuse_get_trace`

Get a specific trace with its observations (generations, spans, events).

**Parameters:**
- `trace_id` (required): The trace ID to fetch

**Example prompt:**
> "Get Langfuse trace details for trace-id-123"

#### `langfuse_list_sessions`

List sessions with pagination. Sessions group related traces together.

**Parameters:**
- `limit` (optional): Maximum results per page
- `page` (optional): Page number
- `from_timestamp` (optional): Filter after timestamp
- `to_timestamp` (optional): Filter before timestamp

**Example prompt:**
> "Show me Langfuse sessions from the last week"

#### `langfuse_get_session`

Get a specific session with its metrics and traces.

**Parameters:**
- `session_id` (required): The session ID to fetch

**Example prompt:**
> "Get Langfuse session details for session-123"

#### `langfuse_list_observations`

List observations (generations, spans, events) with filters.

**Parameters:**
- `limit` (optional): Maximum results per page
- `page` (optional): Page number
- `name` (optional): Filter by observation name
- `user_id` (optional): Filter by user ID
- `trace_id` (optional): Filter by trace ID
- `type` (optional): Filter by type (GENERATION, SPAN, EVENT)
- `from_timestamp` (optional): Filter after timestamp
- `to_timestamp` (optional): Filter before timestamp

**Example prompt:**
> "List all GENERATION type observations from Langfuse"

#### `langfuse_get_observation`

Get a specific observation with full details including input, output, usage, and costs.

**Parameters:**
- `observation_id` (required): The observation ID to fetch

**Example prompt:**
> "Get details for Langfuse observation obs-123"

#### `langfuse_list_scores`

List scores/evaluations with filters.

**Parameters:**
- `limit` (optional): Maximum results per page
- `page` (optional): Page number
- `name` (optional): Filter by score name
- `user_id` (optional): Filter by user ID
- `trace_id` (optional): Filter by trace ID
- `from_timestamp` (optional): Filter after timestamp
- `to_timestamp` (optional): Filter before timestamp

**Example prompt:**
> "Show me Langfuse scores for trace trace-123"

#### `langfuse_get_score`

Get a specific score/evaluation with full details.

**Parameters:**
- `score_id` (required): The score ID to fetch

**Example prompt:**
> "Get Langfuse score details for score-123"

---

### Legacy Tools (Deprecated)

For backwards compatibility, the following tools are still available but will be removed in a future version:

- `list_sessions` → Use `aiobs_list_sessions`
- `get_session` → Use `aiobs_get_session`
- `search_sessions` → Use `aiobs_search_sessions`
- `diff_sessions` → Use `aiobs_diff_sessions`

## Use Cases

### 1. Debugging Failed Runs

> "Show me all AIOBS sessions that had errors in the last 24 hours"

### 2. Performance Analysis

> "Compare AIOBS session abc123 with session def456 and tell me which one was more efficient"

### 3. Prompt Regression Detection

> "Find Langfuse traces with failed evaluations"

### 4. Cost Tracking

> "List Langfuse observations and summarize the total cost"

### 5. Session Inspection

> "Get the full trace tree for the most recent Langfuse trace and explain what happened"

### 6. Cross-Provider Analysis

> "Show me both AIOBS sessions and Langfuse traces from today"

## Development

### Setup

```bash
git clone https://github.com/neuralis/shepherd-mcp
cd shepherd-mcp
python -m venv venv
source venv/bin/activate
pip install -e ".[dev]"
```

### Running Tests

```bash
pytest
```

### Running Locally

```bash
export AIOBS_API_KEY=aiobs_sk_xxxx
export LANGFUSE_PUBLIC_KEY=pk-lf-xxxx
export LANGFUSE_SECRET_KEY=sk-lf-xxxx
python -m shepherd_mcp
```

### Publishing to PyPI

Releases are automatically published to PyPI via GitHub Actions when a release is created.

To publish manually:

```bash
# Build the package
pip install build twine
python -m build

# Upload to PyPI
twine upload dist/*
```

## Architecture

```
src/shepherd_mcp/
├── __init__.py          # Package exports
├── __main__.py          # Entry point
├── server.py            # MCP server with tool handlers
├── models/              # Data models
│   ├── __init__.py
│   ├── aiobs.py         # AIOBS-specific models
│   └── langfuse.py      # Langfuse-specific models
└── providers/           # Provider clients
    ├── __init__.py
    ├── base.py          # Base provider interface
    ├── aiobs.py         # AIOBS client implementation
    └── langfuse.py      # Langfuse client implementation
```

```
┌─────────────────┐     stdio      ┌─────────────────┐
│  Cursor/Claude  │ ◄────────────► │  shepherd-mcp   │
│    (Client)     │   stdin/stdout │   (subprocess)  │
└─────────────────┘                └────────┬────────┘
                                            │ HTTPS
                                  ┌─────────┴─────────┐
                                  │                   │
                                  ▼                   ▼
                         ┌─────────────┐     ┌─────────────┐
                         │ Shepherd API│     │ Langfuse API│
                         │   (AIOBS)   │     │   (Cloud)   │
                         └─────────────┘     └─────────────┘
```

## License

MIT

TDQS

B3.1/5.0

Scored across 18 tools

Disambiguation3/5

The tools are organized into two clear domains (AIOBS and Langfuse), but within each domain there is significant overlap. For example, aiobs_list_sessions and aiobs_search_sessions both retrieve sessions, with the latter adding filtering, which could cause confusion about which to use. Similarly, langfuse_list_sessions and langfuse_search_sessions serve very similar purposes. The deprecated tools further add redundancy without clear differentiation from their AIOBS counterparts.

Naming Consistency4/5

The naming follows a consistent prefix_verb_noun pattern (e.g., aiobs_diff_sessions, langfuse_get_observation), which is predictable and readable. However, there are minor deviations: some tools use 'list' while others use 'search' for similar operations, and the deprecated tools lack the prefix, breaking full consistency but not severely impacting usability.

Tool Count3/5

With 18 tools, the count is borderline high for the scope, which appears to be monitoring and analyzing AI agent sessions and traces. The inclusion of deprecated tools (6 out of 18) inflates the number unnecessarily, making the set feel heavy and cluttered, though the core functionality is well-represented.

Completeness4/5

The tool set provides comprehensive coverage for querying and analyzing sessions, traces, observations, and scores across two systems (AIOBS and Langfuse), with operations like get, list, search, and diff. Minor gaps might include update or delete operations, but these are likely not needed for the monitoring domain, and agents can work effectively with the provided read-oriented tools.

Maintenance

ActivityInactive
ResponsivenessNo issues