TrueFoundry releases TrueForge agent harness under MIT license

• TrueForge is an MIT-licensed agent harness that can be forked and shipped commercially, according to kimmonismus's launch thread.
• The runtime has a local SQLite mode and a hosted Docker Compose or Helm mode backed by Postgres and Redis, as the launch thread describes.
• The core loop handles tool calls, context, subagents, approval checkpoints, sandboxed execution, and traces, which kimmonismus's overview lists as the harness layer.
• Model routing spans OpenAI, Anthropic, Google, Kimi, GLM, and DeepSeek providers, according to kimmonismus's provider list.
• TrueFoundry reported roughly 11 solved tasks out of 14 for GLM-5.2 on TrueForge and Opus 4.8 on Claude Managed Agents in a comparison that kimmonismus's launch thread summarized, at $2.9 rather than $11.8 per run.

The launch post exposes a core server, TypeScript SDK, and embeddable chat UI rather than a single black-box agent endpoint. Its benchmark post used three MCP servers and a blind LLM judge; sandboxed code execution remains a Daytona dependency in kimmonismus's deployment note.

TrueForge separates its runtime into three surfaces in the launch post:

1. Core server: Runs the agent loop, streams steps, pauses sensitive actions for approval, compacts context, uses subagents, and persists sessions across reconnects.
2. HTTP API and TypeScript SDK: Exposes the agent from application code through @truefoundry/trueforge-sdk.
3. Chat UI and UI SDK: Supplies a complete UI that can be used as-is, themed, or embedded.

The project is published in the public GitHub repository under the MIT License.

Local mode is a single process with SQLite and no additional infrastructure. The hosted configuration uses Docker Compose or Helm, plus Postgres and Redis; the official post adds replicas behind a load balancer and OIDC login for team deployments.

TrueFoundry says the agent features are the same in both modes, with the deployment infrastructure as the difference.

TrueForge takes user-supplied model keys, MCP servers, and SKILL.md files. Its documented provider path covers direct OpenAI, Anthropic, and Google integrations alongside open-weight options such as Kimi, GLM, and DeepSeek; the benchmark post says any OpenAI-compatible API can be used.

The launch partners named by TrueFoundry include Together AI, Fireworks AI, Alibaba, and web providers Exa, Tavily, Parallel Web, and Bright Data. MiniMax_AI's announcement said MiniMax joined the ecosystem on day one.

TrueForge provisions a sandbox only when an agent needs to execute code. The launch post says that leaves non-code turns outside the sandbox path, while allowing a server to support concurrent agents.

The sandbox is the exception to the self-hosted story at launch: kimmonismus's deployment note says it currently runs through Daytona, with a self-hosted option planned.

TrueFoundry evaluated its harness against Claude Managed Agents and deepagents on DevRev's Enterprise-Bench: 14 cross-system tasks using the same Salesforce-style CRM, Jira-style tracker, Drive-style document store, three MCP servers, and a fresh session per task. The benchmark methodology says an LLM judge scored answers blind against task criteria, with no partial credit.

Its reported mean results were:

• Claude Managed Agents, Opus 4.8: about 11/14 solved, $11.8 per run, 10M tokens, and 63 minutes.
• TrueForge, Opus 4.8: about 11/14 solved, $8.5 per run, 3.8M tokens, and 40 minutes.
• TrueForge, GLM-5.2: about 11/14 solved, $2.9 per run, and 3.7M tokens.
• deepagents, Opus 4.8: about 10/14 solved, $21 per run, 16.5M tokens, and 64 minutes.

The cost calculation used actual token usage at stated list prices and accounted for cache hits. TrueFoundry attributes the same-model gap to a smaller prompt payload, fewer tool calls, and compaction of history and large tool responses.

The self-hosted mode leaves model and MCP credentials with the operator. TrueFoundry's AI Gateway tier adds RBAC, budgets, PII guardrails, credential rotation, and unified traces; kimmonismus's sandbox note identifies the first three controls as enterprise features.

The project launched as an open-source beta, with general availability described as a few weeks away by . The direct repository link is in kimmonismus's repository post.