Skip to main content
OmniCoreAgent OmniCoreAgent

Give your agent real work. Keep control.

Every action checked before it runs. Every run survives a crash without silently redoing anything. Every step on the record. An agent that does real work (writes code, runs commands, changes systems, calls APIs) needs more around it than a loop and tools. OmniCoreAgent is an agent runtime for Python. It includes the harness, the loop, tools, context, workspace and sub-agents around the model, and adds what running an agent for real needs: a policy on every action, a sandbox as the boundary for its commands, runs that survive a crash, budgets, and a record of every step. You use it from Python, from the command line, or as an HTTP server: one agent object, from a first script to a governed background worker, and onto a benchmark. How it fits together walks one run through every layer.
How a run works: your app calls OmniCoreAgent, which governs every action with a policy, a budget and a sandbox, talks to the model and to tools, and keeps the evidence of the run.

It stays in bounds

Before each action, the policy decides: allow it, ask a person, or refuse it, and names the rule; the policy is on by default. Budgets on tokens, cost, model calls and tool calls, once you set them, are checked before a call is made, not added up after.

It picks up where it stopped

A run pauses for a person or a top-up and continues where it stopped; with a durable store, even after a crash, in another process. It never silently redoes an action it already started.

It's all on the record

Every step is kept: what the model saw, what it did, what the policy decided and who approved, what it cost, and outcomes that arrive later, to review, evaluate and train on.
And what every runtime for real work needs: sandboxes that are not given your keys, human approval, MCP, sub-agents, scheduled runs, an HTTP server, and the same agent on Harbor and Terminal-Bench. How it compares.

Start in three steps

1

Install

Requires Python 3.12–3.14; check with python --version. On 3.10 or 3.11, the install stops and says so.
2

Set your model key

3

Build something

Quickstart — 5 minutes

An agent, a tool, a memory, and the evidence of its run.

Tour — 15 minutes

A policy that asks a person, a sandbox, a budget.

See it

The second line is the run’s trajectory: the step, the tool the model chose, its arguments, and exactly what the model received back.
This agent is governed by default: the permissive-dev policy allowed its tool, and would have refused reading raw secrets, shell commands on the host, unrestricted network and package installs. There is no budget until you set one. The guardrail, privacy redaction and the trace are on too; see the defaults.

Find your way

Tools

Your Python functions as tools, in parallel batches.

MCP servers

stdio, SSE and streamable HTTP, with OAuth.

Code mode

The model writes a program that calls your tools.

Skills

Packaged instructions and scripts the agent can use.

Memory

Sessions in memory, Redis, Postgres or MongoDB.

Workspace files

Files for notes and artifacts, local, S3 or R2.

Context

Long tasks without running out of context.

Sub-agents

Workers the lead agent spawns and reads back.

Models

OpenAI, Anthropic, Gemini, Groq, DeepSeek, Ollama and more.

Events and streaming

Follow a run as it happens.

Using an AI coding agent?

Use these docs with AI tools

Ask AI, llms.txt, copy as Markdown, the docs MCP server, Cursor and VS Code.

AGENTS.md

A map of the repository for coding agents: what lives where, and how to work in it.