TrueForge: Run LLM Agents with a Managed Runtime

Overview

TrueForge is a self-hostable runtime for building and running LLM agents. It handles the execution loop and surrounding infrastructure, such as model calls, MCP tools, session state, sandboxed execution, and human approvals, so developers do not have to wire each capability together themselves.

It can be used through a bundled chat interface, an HTTP API with a TypeScript SDK, or an embeddable UI SDK. Local mode uses SQLite for personal use; hosted mode is intended for teams and multi-replica deployments, using Postgres and Redis.

Key Features

• Connect models from OpenAI, Anthropic, Google Gemini, and other catalog providers, or use an OpenAI-compatible endpoint.
• Configure remote MCP servers, including header authentication and OAuth.
• Load git-backed SKILL.md instruction packs on demand in the sandbox.
• Run code and file operations in an isolated sandbox, provisioned when needed.
• Add human checkpoints for tool approval and user questions, with Generative UI in chat.
• Manage context with subagents, deferred tool loading, Code Mode, large-result offloading, and compaction.
• Use the included chat UI, TypeScript SDK, HTTP API, or embeddable UI SDK.

Use Cases

• A developer prototyping an agent can try local mode without setting up hosted infrastructure.
• A team building an internal assistant can connect its chosen models and MCP tools, then use the chat UI and approval checkpoints for interactive work.
• An application developer can integrate agent sessions and turns through the HTTP API and TypeScript SDK, or embed the UI SDK in an existing product.
• A team running saved agents on a recurring cadence can use schedules for unattended tasks, with hosted deployment for shared use.

Getting Started

Run the local quickstart with:

npx @truefoundry/trueforge@latest

See the README and Quickstart for setup options and configuration details.

Alternatives

• deer-flow: DeerFlow focuses on long-running research, coding, and content tasks coordinated by subagents, while TrueForge provides a broader agent runtime with approvals and chat.
• openai-agents-js: OpenAI Agents JS is a TypeScript SDK for embedding agents in applications, while TrueForge also provides a runtime interface for sessions, sandboxing, and approvals.
• ruflo: Ruflo centers on orchestrating Claude Code and Codex teams with shared memory, while TrueForge provides a general runtime for agents, sessions, and approvals.
• OpenHuman: OpenHuman offers a Rust-based harness across desktop, browser, and terminal, while TrueForge focuses on runtime components including approvals and sandboxing.

Considerations

• The README requires Node.js 22.14 or newer.
• Local mode is intended for personal use on localhost. It has no login by default and stores data in a local SQLite file, so the project advises against using it as a production or internet-facing deployment.
• Hosted mode requires Postgres and Redis, adding infrastructure and deployment work.
• Daytona is the sandbox provider available today; additional providers are described as planned.
• The repository lists 104 open issues, so review current project activity and open issues before adopting it.