True Forge: Run Open-Source Managed Agents on Your Own Hardware
True Forge is an MIT-licensed, self-hosted agent harness that replaces Claude and Gemini managed agents. Here's how it works and how to set it up.
What is True Forge?
True Forge is an open-source, MIT-licensed agent execution harness built by True Foundry. It does the same job as Claude managed agents or Gemini’s agent APIs, running the loop of sessions, tool calls, sandboxing, and approvals, but you host it yourself, on your own hardware or your own cloud account, with whatever model you choose. It ships as a TypeScript repo you install locally with a single command, uses a local SQLite database by default, and exposes your agents through a chat UI, an HTTP API, and a TypeScript SDK.
TL;DR
- True Forge is a self-hosted alternative to managed agents from Anthropic and Google, giving you the same “define an agent, let the platform run the loop” experience without sending your data or configuration to a vendor’s servers.
- The harness, not the model, is often what drives agent costs, because it controls how many tool calls happen, what goes into the context window, and how much token bloat accumulates over a session.
- A benchmark run by True Foundry on 14 enterprise tasks showed True Forge using around 3.8 million tokens versus roughly 10 million for the same tasks on Claude managed agents, using the identical Opus model, a claimed 30% cost reduction with comparable accuracy.
- Switching to a cheaper open model like GLM inside True Forge pushed total cost down further, reportedly by around 75% compared to the original Claude managed agent run.
- Because it’s MIT licensed and self-hosted, True Forge avoids vendor lock-in on models, data retention policies, and infrastructure, which matters for teams that can’t get hosted agent platforms through procurement or compliance.
- It supports local models through Ollama and LM Studio, plus hosted providers like Claude, Gemini, DeepSeek, and others, along with MCP connectors and importable skills.
- The chat interface in True Forge is a test bench, not the product itself, the real value is the runtime you expose via API for other applications, teammates, or internal tools to call.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How does True Forge differ from other agent frameworks?
There are roughly three categories of “agent harness” in circulation right now, and the term gets used loosely enough to cause real confusion.
The first is the coding harness: Claude Code, Codex CLI, OpenCode, and similar tools. These wrap a model with a file system, a shell, and a set of tools, and run the loop for you, but they’re purpose-built for writing code.
The second is the assemble-it-yourself framework: LangGraph, Deep Agents, Agent SDKs, ADK. These give you the raw pieces and expect you to wire up state persistence, approvals, and the execution loop yourself. Powerful, but often tedious when you just want a working harness rather than a framework to build one from scratch.
The third category, the newest, is the hosted managed agent: Claude managed agents, Gemini’s agent APIs. You define an agent, a model (almost always the provider’s own), a set of tools, and a system prompt, and the provider runs everything else: sessions, tool calls, sandboxes, approvals. Convenient, but until recently this category meant exactly one thing: you were tied to that provider’s infrastructure and, usually, their models.
True Forge sits in this third category but breaks the vendor tie. It’s the same “let the platform run the loop” experience, but it’s a repo you clone and run, with any model plugged in.
Why does the harness matter more than the model?
Most people default to thinking about agent cost purely in terms of price-per-token and which model they picked. That misses where the money actually goes: the harness decides how many tool calls get made, in what order, and what accumulates in the context window over a session. Since you’re billed for what sits in that context window, a harness with tool bloat (too many tool definitions loaded at once, no progressive disclosure of information) runs up token counts fast, independent of which model is doing the reasoning.
True Foundry tested this directly. They took 14 tasks from DevRev’s enterprise benchmark suite, tasks requiring an agent to join data across a CRM, an issue tracker, a document store, and multiple MCP servers using the same prompt, and ran three trials per task with a blind LLM judge scoring the outputs. They deliberately kept the model constant (Opus) across three different harnesses to isolate the harness’s effect.
The result: the same model, the same task, and comparable accuracy, but True Forge used around 3.8 million tokens per run on average versus roughly 10 million tokens for Claude managed agents, a claimed 30% cost reduction from harness efficiency alone. When they then swapped in a cheaper open model like GLM, total cost dropped further, by a claimed 75% relative to the original Claude managed agent baseline. Strip out the model swap and just look at same-model, same-answer-quality: that gap is entirely down to how the harness manages tool calls and context.
Is vendor lock-in a bigger issue than cost?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
For a lot of teams, yes. When your agent definition lives inside someone else’s hosted runtime, the system prompt, the tool wiring, the approval rules, and the session history all sit on that vendor’s product. That’s fine until you want to try a different model, a different sandbox, or a different configuration, and discover you can’t move any of it easily.
Data retention is a concrete example: if you need zero-day data retention, you can’t use Claude managed agents at all. Companies operating under compliance mandates that require on-premises or private-cloud deployment often can’t get a hosted agent platform through procurement, no matter how capable it is. Their fallback has historically been building the entire loop themselves, which tends to be slower and more expensive than either option.
Open-weight models changing the calculus here matters too. As open models get more capable, the same weights are available across multiple providers with compatible API endpoints, so choosing a self-hosted harness doesn’t mean settling for weaker intelligence. It means you can point the harness at whichever provider or in-house deployment fits your legal, cost, or performance requirements at any given moment, including fully local models run through Ollama or LM Studio.
How do you set up True Forge?
Setup is designed to be quick. True Forge installs locally with a single command from its documentation’s quick-start guide, and by default runs against a local SQLite database, so your data doesn’t leave the machine. There’s also a hosted mode for production use, supporting Postgres, Redis, and Docker, but the local, single-command path is the deliberate on-ramp for testing it out.
Once it’s running, you configure four things in the settings:
- Models: standard providers like Claude and Gemini are supported, alongside other open providers, or you can point it at a fully private API endpoint. Any model that exposes a compatible API can be wired in.
- Connectors: MCP servers and tools such as Tavily (search), DeepWiki, Notion, or Bright Data, or a custom MCP server running locally or in the cloud.
- Skills: importable via a skill.md file, giving the agent a defined capability (a Tavily research skill, for example) without hand-coding tool logic.
- Sandboxes: execution environments for the agent’s actions, an area still developing at time of writing.
The built-in chat interface lets you test that models, connectors, and skills are wired up correctly, and it will actually run as an agent (reasoning, tool calls, checking each connected piece) rather than acting as a simple chatbot. But this interface is explicitly a test bench, not the end product. The real deployment target is the runtime itself, accessible via HTTP API or the TypeScript SDK, so you can build an agent once and call it from Slack, a mobile app, or any internal tool without every consumer needing to know how the agent is configured.
Is True Forge worth using over a managed agent platform?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
It depends on what you’re optimizing for. If you want the fastest path to a working agent with no infrastructure to manage, a hosted managed agent platform still removes more friction. If you’re running agents at volume, need to control costs, can’t send data to a third party, or want the freedom to swap models without rewriting your integration, a self-hosted, open-source harness like True Forge addresses exactly those constraints, at the cost of running your own infrastructure.
Frequently Asked Questions
What license is True Forge released under?
True Forge is released under the MIT license, meaning it can be used, modified, and self-hosted without a paid tier or usage-based runtime billing.
Can True Forge run fully local models?
Yes. It integrates with local model runners like Ollama and LM Studio, in addition to hosted providers such as Claude, Gemini, DeepSeek, and other open model APIs.
Does True Forge require Claude or Gemini to work?
No. You can configure any model with a compatible API endpoint, including open-weight models hosted by third-party providers or run entirely on your own hardware.
How is True Forge different from LangGraph or similar frameworks?
LangGraph and similar tools give you building blocks and expect you to assemble the execution loop, state persistence, and approval logic yourself. True Forge runs that loop for you, similar to a managed agent platform, but self-hosted.
What’s the actual benefit of using True Forge over Claude managed agents?
Based on True Foundry’s own benchmark using the same model across harnesses, True Forge used substantially fewer tokens per task (about 3.8 million versus about 10 million) while maintaining comparable accuracy, plus it avoids data retention restrictions and vendor lock-in tied to a specific provider’s infrastructure.
