TrueForge
A general-purpose open-source agent runtime framework released by TrueFoundry under the MIT license. It supports tool calling, MCP server connections, code sandboxing, and human-in-the-loop mechanisms, and can connect to multiple LLM providers, positioning itself as a low-cost alternative to managed agent services.
Timeline (last 90 days)
TrueForge is an open-source agent framework that can be launched via the command npx truefoundry/trueforge and supports integration with OpenAI, Anthropic, Gemini, or custom model providers.
Without any tools connected, a model acting as a research agent can only answer based on its training data, still believes the current time is around 2025, and cannot access the internet.
After the author connected the Exa web search MCP server to TrueForge, the agent produced a research summary via real web searches that included models such as DeepSeek R1, MiniMax, GLM, Mistral, and others.
In the benchmark, using the open-source model GLM with TrueForge reduced costs by as much as 75% while maintaining accuracy.
On first run, TrueForge downloads dependencies and creates a local SQLite database; the service runs on localhost and provides a web interface.
TrueForge supports connecting to OpenAI, Anthropic, and Gemini, and can also add custom providers (such as Ollama local models) by specifying a base URL.
TrueForge's chat interface is a testing sandbox where temporary operations are not persisted as system prompts; a formal Agent is defined by the system prompt, model selection, connectors, and skills saved at creation time.
TrueForge is open-sourced by TrueFoundry under the MIT license; it is a general-purpose Agent runtime layer designed to replace Claude's managed agent service.
The TrueForge API call flow is: POST a request to the sessions endpoint to start a session and retrieve a session ID, then send a user message to the /{sid}/turns endpoint and receive a streaming response via SSE.
By connecting to a Daytona sandbox, TrueForge allows the Agent to actually execute Python code, for example generating charts with matplotlib and outputting a PDF report.
4 more timeline events
All Facts (13)
TrueForge is an open-source agent framework that can be launched via the command npx truefoundry/trueforge and supports integration with OpenAI, Anthropic, Gemini, or custom model providers.
50%UnverifiedWithout any tools connected, a model acting as a research agent can only answer based on its training data, still believes the current time is around 2025, and cannot access the internet.
50%UnverifiedIn the benchmark, switching the same model (Opus 4.x) from Claude's managed agent to TrueForge yielded roughly equivalent accuracy while reducing costs by approximately 30%.
50%UnverifiedIn the benchmark, using the open-source model GLM with TrueForge reduced costs by as much as 75% while maintaining accuracy.
50%UnverifiedOn first run, TrueForge downloads dependencies and creates a local SQLite database; the service runs on localhost and provides a web interface.
50%UnverifiedTrueForge supports connecting to OpenAI, Anthropic, and Gemini, and can also add custom providers (such as Ollama local models) by specifying a base URL.
50%UnverifiedTrueForge is open-sourced by TrueFoundry under the MIT license; it is a general-purpose Agent runtime layer designed to replace Claude's managed agent service.
50%UnverifiedTrueForge provides a REST API and a TypeScript SDK, but currently has no Python SDK.
50%UnverifiedThe TrueForge API call flow is: POST a request to the sessions endpoint to start a session and retrieve a session ID, then send a user message to the /{sid}/turns endpoint and receive a streaming response via SSE.
50%UnverifiedBy connecting to a Daytona sandbox, TrueForge allows the Agent to actually execute Python code, for example generating charts with matplotlib and outputting a PDF report.
50%UnverifiedTrueForge supports connecting to local models; it can connect to a Dell Pro Max running the Qwen model (equipped with the GB10, i.e., the Dell version of the DGX Spark) via the Ollama base URL (localhost:11434/v1).
50%UnverifiedTrueForge's chat interface is a testing sandbox where temporary operations are not persisted as system prompts; a formal Agent is defined by the system prompt, model selection, connectors, and skills saved at creation time.
50%UnverifiedThe benchmark published by the TrueForge team is based on 14 tasks from DevRel Enterprise Bench, involving interactions with systems such as CRMs, issue trackers, and documentation repositories.
50%