Psych Runtime
GitHub

Agent runtime for Python

Build AI agents
that finish the work.

Create agents and multi-step workflows that use your tools, wait for people, and recover from interruptions. Psych runs inside your Python application, with your model and infrastructure.

See how it works Start building
Inside a runInteractive overview
Psych Runtime
01 / THE RUNTIME

Turn decisions into actions.

Psych calls the tools your agent is allowed to use and returns their results to the model. The loop continues until the agent finishes.

Your modelChooses the action
PsychRuns the work
Your toolsDo the action
Your store Steps & results
Model chooses a toolEach step becomes part of the run history.
  1. 01Model chooses a tool
  2. 02Psych checks access
  3. 03Tool executes
  4. 04Result returns to model

The work between model calls

From a model decision
to a finished run.

A model chooses what to do next. Psych checks what it can access, executes the action, records the result, and carries the run forward. Your application stays in charge of permissions, people, and deployment.

Agents

Define the instructions, model, tools, and access once. Psych runs the model and tool loop for each request.

Workflows

Combine tools, agents, and nested workflows into named steps. A recovered run continues after its last completed step.

Tools

Connect Python functions, HTTP endpoints, MCP servers, or other agents. Access is checked again on every turn.

Memory

Remember facts across runs for the same tenant and end user. Keep retrieval behind the system you already use.

Inside your application

Keep your product.
Add agent execution.

Import Psych into the application you already run. Keep your interface, authentication, database, and deployment. Supply the model, tools, and storage that fit your product.

Explore the architecture

The practical questions

Before you
start building.

What is Psych Runtime?

Psych is an embeddable Python runtime for AI agents and workflows. It runs the model and tool loop, records progress, pauses when your application needs a person, and exposes the result through a small public API.

What can I build with it?

Use Psych behind customer-facing agents, internal automation, or a product where your own users create agents. It supplies execution. Your application supplies the interface, identity, and business rules.

How do workflows work?

A workflow sequences named tool, agent, or nested workflow steps. Completed step outputs are read from the run log, so recovery continues after the last completed step instead of running it again.

How do I deploy it?

Run Psych inside your service and worker processes. Both use the same Store. You choose and operate the model provider, database, sandbox, and telemetry that your deployment needs.

Is it ready for production?

Psych is pre-alpha at 0.x. The runtime exists and is tested, but the public API can change between minor versions. Start with a bounded use case you can test end to end.

Build with Psych

Your first agent
starts with one tool.

Build your first agent

Pre-alpha. The API is still evolving. Start with a guided example and a use case you can test.