TrueForge is an open-source agent harness, that is, the runtime layer that turns an LLM into a working agent. It is developed by TrueFoundry and released under the MIT license. Its core value lies in this: building an agent is easy, but running it well is hard—you need streaming output, session persistence, tool servers, sandboxes, approvals, and a UI. TrueForge provides these capabilities out of the box.
Main capabilities:
-
Execution loop: TrueForge runs the agent's execution loop for you, including model calls, MCP tools, skills, sandboxes, approvals, context management, and session state.
-
Three ways of exposure: a chat UI, an HTTP API with a TypeScript SDK (@truefoundry/trueforge-sdk), and an embeddable UI SDK (@truefoundry/trueforge-ui).
-
Initial setup comes from a catalog: configure models, MCP servers, skills, and sandboxes once, and the agent picks from what is connected. Presets come from the included YAML catalog and can be customized.
-
Any model provider: supports OpenAI, Anthropic, Google Gemini, and other catalog providers, or any OpenAI-compatible endpoint.
-
MCP tools: supports remote MCP servers with header authentication or OAuth, including in-chat authorization.
-
Skills: git-based SKILL.md instruction packages, loaded on demand in the sandbox.
-
Sandbox as a tool: isolated code/file execution (currently Daytona, with plans to support more providers in the future), configured only when needed. Secrets remain in the harness.
-
Human checkpoints: tool approvals, asking the user questions, and generative UI in chat.
-
Context engineering: sub-agents, deferred tool loading, Code Mode, large-result offloading, and compaction.
-
Chat UI + SDK: use the built-in UI, automate via @truefoundry/trueforge-sdk, or embed @truefoundry/trueforge-ui.
Deployment methods:
TrueForge can scale up or down: local mode (single process + SQLite) or hosted mode (Postgres + Redis, Docker Compose, Helm, or Railway).
Quick start:
You can use the Quickstart guide to run TrueForge in multiple ways (locally, with Docker Compose, Kubernetes, or Railway), connect models, tools, and skills, and build your first reusable agent. To develop TrueForge in this repository, see CONTRIBUTING.md.
Limitations and caveats:
Local mode is only suitable for your machine and is a convenient way to try out TrueForge, not a production or internet-facing setup. By default there is no login, data is stored in a local SQLite file, and you should keep it on localhost. If you use local mode beyond this scope, the official project assumes no responsibility for data loss or unauthorized access. For shared or production deployments, use hosted mode.
In addition, TrueForge was benchmarked against Claude Managed Agents and deepagents on the same tasks, tools, and models; the project states it has the same accuracy at lower cost, reproducible from the benchmark/ directory, with the related article being Benchmarking. The project welcomes contributions (bug reports, feature and documentation fixes); fork PRs should modify only the source code, and maintainers will regenerate the SDK after merging. Security vulnerabilities should be reported according to SECURITY.md rather than as a public issue. The community has a Discord.
