Skip to main content

Getting Started

sypho is a runtime and control plane for deploying and running production AI agents. Package an agent as a versioned deployment, run it inside an isolated sandbox, and control exactly what it can touch — tools, data, secrets, and spend — while every action is traced back to the control plane.

What is sypho?​

sypho lets you:

  • Bring your own framework — LangChain, LangGraph, CrewAI, or anything else that runs in a container. No proprietary SDK required.
  • Deploy to your own infrastructure — self-hosted control plane and runners, local or in your cloud
  • Scale across dev, staging, and production deployments
  • Access structured data through namespaced, versioned platform tools
  • Observe everything with full run traces, tool-call logs, and spend tracking

Quick Start​

1. Get the CLI​

curl -fsSL https://your-control-plane.example.com/downloads/install.sh | sh

See Installation for details, including installing just the runner.

2. Authenticate​

sypho login
# Prompts for your control-plane URL and platform token

3. Create and run an agent​

sypho init my-agent
cd my-agent

# Build the agent image
sypho build --version 0.1.0

# Deploy it
sypho deploy --version 0.1.0

# Run it
sypho run my-agent:0.1.0 --input '{"message": "Hello!"}'

Your agent is now running on your sypho control plane.

How It Works​

  1. Write your agent in any framework — see Bring Your Own Framework
  2. Declare a manifest (sypho.manifest.json) describing its entrypoint and what it's allowed to touch
  3. Build with the CLI (produces a Docker image)
  4. Deploy to your control plane (registers the manifest and image)
  5. Run via CLI or API — execution happens inside a network-isolated sandbox, enforced against the manifest

Key Concepts​

Agents​

An agent is a Docker image plus a manifest. The manifest declares:

  • Entrypoints — how the agent can be invoked
  • Tools — which tools (by exact name, glob, or pattern) the agent may call
  • Structured-data dependencies — which namespaces it can read or write

See the manifest reference for the full schema.

Deployments​

Organize your agents by environment:

  • dev — development and testing
  • staging — pre-production validation
  • prod — production workloads

Runs​

Each execution of your agent:

  • Created via CLI or API
  • Executes inside its own network-isolated sandbox
  • Tracked with logs and events, traceable back to its deployment version

Platform Tools​

Every agent gets access to namespaced, versioned structured data:

structured_data_save   — create or update an entry
structured_data_get    — fetch by key
structured_data_query  — search by metadata

A run's access to each namespace is fixed at creation — read or read_write, scoped to specific keys — and enforced on every call.

Agents also get secrets, Slack notifications, human-in-the-loop guidance, and sub-agent delegation — see Platform Tools for the full list.

Monitoring​

  • Real-time logs via CLI: sypho logs <run_id> --follow
  • Full run and tool-call traces in the control plane

Next Steps​