# Introduction

> Foundry keeps AI coding agents working through long, multi-day software jobs. It runs your tasks one at a time, checks every result with local tests, and…

Source: https://bifrost5.com/foundry/docs/introduction/

# What is Foundry?

Information

The open-source edition runs everything locally. Production-Grade adds what a team needs around it.

- Steer your runs from web, phone or chat
- Never lose a run: encrypted cloud backup
- See finished work securely before you merge
- Plan faster, or bring your own AI tool

[See editions](/foundry/docs/editions)

Foundry keeps AI coding agents working through long, multi-day software jobs. It runs your tasks one at a time, checks every result with local tests, and records what happened, so the work survives crashes and restarts.

## The Problem

When you chain AI agents together on long multi-day coding projects, they fail due to:

- **Context degradation**: After multiple tasks, agent drift occurs. They forget earlier context.

- **Manual recovery burden**: If execution halts or crashes, resuming requires human intervention.

- **Cost spiral**: Without validation, agents make redundant calls and burn tokens.

- **No audit trail**: You can't review what happened or why it failed.

## The Solution

Foundry acts as a **deterministic supervisor** for agentic chains. It orchestrates tasks, validates outputs relentlessly using local non-AI checks, maintains context across repetitions with a sliding window, and keeps costs bounded.

## Why Foundry is Different

- **Headless Worker Abstraction (BYOK)**: Wraps standard CLIs & APIs (Cursor, Copilot, Gemini, Ollama) and drives them as autonomous worker nodes. Leverage your providers' context engines: Foundry manages the session.

- **Strict Determinism**: Foundry runs the tasks you define and does not plan for itself. Removing planning from the worker agent stops scope creep.

- **Local-First Hybrid Validation**: Free, local deterministic checks (AST, regex, file capacity) run *before* ever asking an LLM. The LLM is a "Judge of Last Resort."

- **Crash-Proof Infrastructure**: State persisted to DragonflyDB after every atomic step. Kill the process, upgrade, restart: `resume` exactly where it left off.

### Gotchas

The open-source edition is a complete local setup, with some limits to know before you start:

- **It does not plan.** Foundry runs the tasks you give it. Write them yourself or have your own AI tool draft them first.

- **It runs on one machine.** State lives locally, so there is no cloud backup or sync. If the machine is lost, so is the history.

- **You drive it from the dashboard or the CLI.** There is no mobile app, chat control or direct hook for your own AI tool to pull work.

- **You see results as files and test output.** There are no hosted previews of the finished work.

- **Support is community-based.** There are no response-time commitments.

- **Checks are only as good as your tests.** You pay your own providers, and it needs Docker for state.

The [Production-Grade edition](/foundry/docs/editions) covers these.

### Status

- **Public Beta**: Core control loop and state persistence are stable.

- **License**: MIT for Open-Source Variant Core · Commercial White-Label for Production-Grade Edition

- **Model**: Bring Your Own Keys (BYOK). You supply your own provider subscriptions.