Documentation

How it works

Sorin sits between your agents and the services they call (LLMs, GitHub, and other tools). Your upstream credentials are stored in Sorin — your code only ever uses a scoped Sorin agent key.

Every agent action is authorized against your org's permission rules, logged, and optionally held for human approval before execution.

Installation

pip install sorin-sdk

Requires Python 3.8+.

Agent keys

Get your agent key from the Sorin dashboard after creating an agent. Your upstream API credentials (GitHub, Anthropic, OpenAI) are stored in Sorin — your code only ever sees the Sorin agent key.

MCP Server

Connect Sorin to any MCP-compatible AI coding tool. The SDK ships with a sorin mcp install CLI that handles setup for you.

pip install sorin-sdk
sorin mcp install --key <your-agent-key>

Restart Claude Code after running the command to connect.

Using Cursor, Windsurf, or VS Code? Pass --json to print the JSON config block to paste into your editor's MCP settings instead:

sorin mcp install --key <your-agent-key> --json

GitHub Actions

GitHub operations through Sorin are authorized against your org's permission rules, logged, and optionally held for human approval. Your upstream GitHub token never leaves Sorin — your code only uses the scoped agent key.

from sorin import SorinClient

sorin = SorinClient(agent_key="<your-agent-key>")

# Read a file
file = sorin.github.read_file("your-org", "your-repo", "README.md")

# Push a file to a branch
sorin.github.push_file("your-org", "your-repo", "notes.md", "# Hello", "add notes", "my-feature")

# Open a PR
pr = sorin.github.create_pr("your-org", "your-repo", "My PR", "body", "my-feature")

LLM Inference

One-line swap — replace your existing Anthropic or OpenAI client with the Sorin drop-in. Every inference call is authenticated, logged, and governed. No other code changes required.

from sorin import SorinLLM

client = SorinLLM(agent_key="<your-agent-key>")

# Identical to anthropic.Anthropic().messages.create(...)
response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}]
)

Permissions & Approvals

Every agent has a permission set that controls which actions it can take and which require human approval. Actions not explicitly permitted are blocked. For sensitive actions, Sorin pauses execution and notifies an approver — the agent resumes only after explicit approval or denial.