TrueForge: an MIT agent harness that runs 30-75% cheaper
TrueFoundry open-sourced TrueForge, an MIT-licensed agent harness. Benchmarked at $2.90 vs $11.80 against Claude Managed Agents. The loop is the part you should own.
The model gets the headlines. The harness gets the work. TrueFoundry open-sourced TrueForge on August 18 under an MIT license — the runtime loop that sits between a language model and an agent that actually does something. It is the layer most teams rent without noticing, and this open source agent harness makes the case that renting it is the expensive part.
What actually happened
TrueForge runs the loop: model calls, MCP tool integration, session state, sandboxed code execution through Daytona, human approval checkpoints, and context compaction. It ships as a core server, an HTTP API with a TypeScript SDK, and a chat UI. Local mode runs on SQLite; hosted mode runs Postgres and Redis. You start it with npx @truefoundry/trueforge, or deploy it with Docker Compose or a Helm chart.
The cost claim comes from a benchmark TrueFoundry ran on DevRev's Enterprise-Bench, which tests multi-step tool use across CRM, issue tracking, and document systems. Per VentureBeat's reporting, TrueForge completed 11 of 14 tasks. Running GLM-5.2, it cost $2.90. Claude Managed Agents running Opus 4.8 cost $11.80. TrueForge running the same Opus 4.8 cost $8.50 — 30% cheaper on identical model spend, which points at the harness rather than the model.
That benchmark is vendor-run. The reproduction code sits in the repo, which is more than most vendors offer, but run it on your own workload before you quote the number to anyone.
Why owning the agent harness matters for your business
Models are the commodity. The loop is your business logic. Which tools an agent may call, who approves a write, what gets compacted out of context, what lands in the audit log — that is not model behavior, it is policy. When it lives inside a managed agent product, changing vendors means rewriting your policy.
30% on the same model is the number to sit with. Identical model, identical task, a third less spend. The difference is loop efficiency: how much context gets resent, how many tool round-trips, how aggressively history is compacted. Managed platforms have no incentive to optimize that. You do.
MIT means you can fork it. Not "open weights with usage terms." Not "source available." MIT — fork it, modify it, self-host it, ship it inside a commercial product. That is the license that makes a dependency safe to build on.
Self-hosting is not free. Postgres, Redis, a sandbox provider, and someone who owns upgrades. For a five-person shop already running a database, that is marginal. If you run nothing today, a managed agent may still be the right call until volume justifies the operations work.
Key takeaways
- TrueForge is an MIT-licensed agent harness, open-sourced by TrueFoundry on August 18, 2026
- Vendor-run Enterprise-Bench results: $2.90 with GLM-5.2, $8.50 with Opus 4.8, vs $11.80 for Claude Managed Agents on Opus 4.8
- 30% cheaper on the same model points at loop efficiency, not model choice
- Ships with MCP tools, Daytona sandboxing, human approval checkpoints, and context compaction
- Run locally with
npx @truefoundry/trueforge; deploy via Docker Compose or Helm - Benchmark code is in the repo — reproduce it on your own workload before citing it
The harness is the part of your AI stack that should have your name on it. We build agent systems where the model is swappable and the approval rules, tool permissions, and audit log belong to you. See how we build vendor-agnostic agent systems, or look at what we've shipped.
Sources: TrueFoundry engineering blog, truefoundry/trueforge on GitHub, VentureBeat.
- #open-source
- #ai-agents
- #agent-harness
- #vendor-lock-in
- #ai-costs
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Meta AI's Mac app reads your screen — and trains on it
Meta's new Mac app does system-wide dictation, screen context, and Google Workspace access. Read the training-data terms before your team installs it.
Read itCursor Origin: keep your git remotes portable
Cursor launched Origin, an agent-scale git forge, the same week GitHub threw errors for seven hours. How to add a remote without handing over your codebase.
Read it