TrueForge

TrueForge is an open-source agent harness built by TrueFoundry, released under the MIT License. It handles the full agent execution loop — model calls, MCP tools, skills, sandboxing, human-in-the-loop approvals, context management, and session state — and exposes everything through a chat UI, an HTTP API with a TypeScript SDK, and an embeddable React UI SDK. The GitHub repository reports 4,373 stars and 276 forks as of late August 2026.

What It Is

TrueForge sits between a user goal and the final agent response as the runtime orchestration layer. Rather than just generating text, it runs the full execution loop: planning, tool routing and execution, context management for long tasks, security boundaries like sandboxing and human-in-the-loop approvals, and session state that survives reconnects and restarts. The project describes itself as "the open-source agent harness — the runtime layer that turns an LLM into a working agent."

Architecture and Key Components

TrueForge has three main parts:

• TrueForge Server — runs the agent loop, streams every step back to the client, pauses for human approval on sensitive actions, and persists sessions in SQLite (local) or Postgres (hosted).
• HTTP API + TypeScript SDK — a REST + Server-Sent Events API with an OpenAPI spec, plus the @truefoundry/trueforge-core and @truefoundry/trueforge-sdk npm packages for programmatic agent control.
• Chat UI + UI SDK — a bundled browser interface served on the same port as the server, plus @truefoundry/trueforge-ui, a customizable React component library for embedding the chat experience in your own app.

The server treats the sandbox as a tool rather than running the agent inside a sandbox permanently. A sandbox is provisioned only when the agent actually needs to execute code, which allows one server to run many agents concurrently and keeps turns that don't need code execution cheaper and faster.

Deployment Model

TrueForge supports two modes with identical agent features:

• Local mode — a single process on your machine, launched with npx @truefoundry/trueforge, backed by SQLite with no extra infrastructure. Intended for personal use only; not hardened for production or shared internet access.
• Hosted mode — Docker Compose or Helm deployment backed by Postgres and Redis, supporting multiple replicas behind a load balancer. Redis peers replicas so streams and cancellations follow the client across replicas. Supports optional OIDC login for admin vs. user roles.

Harness Capabilities

TrueForge ships a set of built-in capabilities that go beyond basic tool calling:

• MCP tools — remote MCP servers with header auth or OAuth, including in-chat authorization flows
• Skills — git-backed SKILL.md instruction packs loaded on demand in the sandbox
• Human checkpoints — tool approval gates, ask-user-questions, and Generative UI in chat
• Context engineering — subagents, deferred tool loading, Code Mode, large-result offloading, and compaction for long tasks
• Any model provider — OpenAI, Anthropic, Google Gemini, and any OpenAI-compatible endpoint, configured from YAML catalogs

Update: @truefoundry/trueforge-ui@0.2.4

The latest published release is @truefoundry/trueforge-ui@0.2.4, published on August 19, 2026. The repository was last pushed on August 25, 2026, indicating active development. The project also includes a benchmarking suite comparing TrueForge against Claude Managed Agents and deepagents on the same tasks, tools, and model, with results reproducible from the benchmark/ directory in the repository.