TrueForge Explained [What It Changes]

TL;DR

• TrueForge is a free, MIT-licensed agent harness from TrueFoundry that runs the plan-call-tool-repeat loop on your own servers, with any model you bring, instead of renting that loop from a single model vendor.

• As of August 19, 2026, TrueFoundry published a 14-task DevRev Enterprise-Bench comparison in which TrueForge plus GLM-5.2 averaged $2.90 per run against $11.80 for Claude Managed Agents on Opus 4.8, and still landed about 30% cheaper when both sides used Opus 4.8.

• A two-truck HVAC shop, a 10-person agency, or a solo clinic should care because the bill is no longer locked to one frontier model once the tools (CRM, inbox, certificates, checklists) already exist as Model Context Protocol connections.

• The honest limit: the published savings are vendor-run, local npx mode is not production, and governance (SSO, budgets, audit) is a separate paid gateway, not a free gift of the MIT license.

Key Takeaways

• The constraint that broke is model lock-in at the harness layer: the loop, sandbox, approvals, and context packing were bundled with one vendor's model and markup.

• TrueForge splits that layer into a core server, an HTTP API with a TypeScript SDK, and a chat UI, then lets you point the same agent at OpenAI, Anthropic, and 20+ other models.

• Context waste is the cost lever: delayed MCP schema loading, subagents, file offload of fat tool results, code-mode processing, and a default 50,000-token compaction threshold.

• Self-hosting pays when volume is high and cheaper models can finish the same tickets; managed runtimes can still be cheaper for small, spiky workloads.

• Treat a harness swap as a model-and-loop change, not a rebuild of form-to-CRM, certificate collection, or work-instruction tracking you already run.

What TrueForge is

TrueForge is the open-source runtime that sits around a language model, calls your tools, pauses for human approval, keeps session state, and keeps the context window from eating the budget — so the agent does the work instead of only writing a plan.

That sentence is the whole product. The model still reasons. The harness is what opens the file, hits the API, sandboxes the code, and remembers the last 40 turns.

If you run a two-truck HVAC company, the expensive part of "AI help" is not a chatbot greeting. It is the agent that reads a service ticket, pulls the last invoice, checks a parts PDF, and drafts the customer note. If that loop is glued to one vendor's managed agent, every messy ticket is billed like a frontier research task. A 10-person marketing agency sees the same pattern when an agent copies a form fill into the CRM and then writes a follow-up. A solo clinic sees it when an agent files a lab PDF and updates the chart. In each case the tools already exist. What you are overpaying for is the rented loop.

TrueFoundry, a San Francisco machine-learning infrastructure company founded in 2021, open-sourced TrueForge under the MIT License and put the code on GitHub. The local start command is npx @truefoundry/trueforge. Model keys, MCP servers, and the sandbox are bring-your-own. A paid hosted path and the company's AI Gateway sit beside that free core; they are not the same product.

For shops that already move documents through US Tech Automations workflows, TrueForge is a candidate for the model-and-loop slot, not a reason to redraw the intake path.

What shipped in mid-August 2026

TrueFoundry's engineering post is dated 18 August 2026. VentureBeat filed the news at 4:50 p.m. PT on 19 August 2026. InfoWorld followed on 20 August 2026.

The comparison target is Anthropic's Claude Managed Agents: a hosted runtime for long-running agents on the Claude platform, described on Anthropic's site as part of the company's model-and-product stack. TrueForge's pitch is the opposite shape. You run the harness. You pick the model. You keep the MCP connections when the model changes.

According to InfoWorld, TrueForge supports OpenAI and Anthropic models as well as more than 20 additional models, and developers can bring their own MCP servers and API keys. The same article notes a hosted version with usage-based pricing.

The license is not a slogan. The Open Source Initiative's MIT text grants use, copy, modify, merge, publish, distribute, sublicense, and sell rights, with the copyright notice kept and the software provided "AS IS". That is why a company can fork TrueForge, embed it, or ship it inside a commercial product without asking TrueFoundry for a seat.

The public repo, as loaded from GitHub, listed 5.1k stars, 352 forks, and 482 commits on main. The public repo shows 5.1k GitHub stars. Docs live at trueforge.dev. Local mode stores data in SQLite and is not an internet-facing service. Hosted mode uses PostgreSQL and Redis, started with Docker Compose or a Helm chart.

Why the bill was stuck

An agent that only chats is cheap. An agent that works is a loop: the model plans, the harness calls a tool, the tool returns a blob, the blob goes back into the prompt, and the cycle repeats until the task is done or someone hits stop. Every extra token in that loop is money. If the harness is proprietary and the model is proprietary, you cannot send the easy tickets to a cheaper model without rebuilding the agent.

That is the lock-in InfoWorld asked analysts about. Pareekh Jain of Pareekh Consulting said TrueForge gives enterprises more control and less vendor lock-in, including routing simple tasks to cheaper or open-weight models. Lian Jye Su of Omdia said the split is useful in regulated industries because the buyer can attach their own budget, access, and observability controls — and that the trade-off is running more of the stack yourself.

The constraint that broke is practical. Model Context Protocol gave tools a USB-C-style plug: one way to attach calendars, files, CRMs, and search. Once tools are portable, the remaining hostage is the loop. TrueForge is a bid to make that loop portable too.

MCP itself is still moving. According to InfoWorld's July 2026 protocol report, the release candidate was scheduled for July 28 and drops protocol-level sessions in favor of a stateless design so MCP services scale like other cloud apps. A harness that already treats MCP servers as bring-your-own plugs is aligned with that direction. It does not freeze the protocol for you.

Open-weight models are the other half of the break. Computerworld described IT leaders trying smaller open models to control cost, customize behavior, and keep data off a competitor's training pile. TrueForge's GLM-5.2 bench is that story applied to agents: same tasks, cheaper weights, only if the harness will actually call them.

How the harness works

TrueFoundry's launch post splits TrueForge into three pieces: a core server that runs the agent loop (stream each step, pause on sensitive actions, compact context, persist sessions); an HTTP API and TypeScript SDK so you can drive the same loop from your own app; and a chat UI you can theme or embed.

Most harnesses keep the agent inside a sandbox for the whole run. TrueForge treats the sandbox as a tool and spins one up only when the agent needs to execute code or touch files. Launch partners listed on that post include Daytona for the sandbox, which lists $0.0504/h per vCPU and $200 of free compute — Daytona's prices, not TrueForge's, and they sit on top of model tokens.

Context engineering is where TrueForge tries to spend less: delayed MCP schema loading, subagents, file offload of fat tool results, code-mode processing, and auto-compaction. According to VentureBeat, the documentation sets the default compaction threshold at 50,000 tokens.

Quickstart docs require Node.js 22.14 or newer. Local mode has no login by default and must stay on localhost. Hosted mode adds Postgres, Redis, Docker Compose or Helm.

A 10-person agency that already pushes form fills into a CRM through US Tech Automations can keep that connector and only change which model sits behind the agent loop. The form-to-CRM automation roundup is the intake side of that split: TrueForge does not replace the form; it replaces the rented brain that files it.

The published bench

TrueFoundry ran TrueForge against Claude Managed Agents and LangChain's deepagents on 14 L1–L2 tasks from DevRev's Enterprise-Bench: cross-system jobs over a Salesforce-style CRM, a Jira-style tracker, and a Drive-style store, via three MCP servers. DevRev publishes the suite as vendor-neutral, with a public task set and an independent LLM judge. TrueFoundry says every harness got the same tasks, tools, a fresh session per task, and a blind judge; a task counts as solved only if it meets every criterion.

Cost on the marketing bench used list token prices of $5 / $25 per 1M in/out for Opus 4.8 and $0.73 / $2.28 for GLM-5.2, cache-aware.

The more precise table on TrueForge's benchmarking docs reports means over n = 3 trials:

Sources: TrueForge benchmarking; task suite described at DevRev Enterprise-Bench.

The company blog rounds the same story as ~11 / 14 solved, $11.8 vs $8.5 vs $2.9, with wall-clock latency of 40 min for TrueForge on Opus against 63 min for Claude Managed Agents and 64 min for deepagents. According to that TrueFoundry benchmark post, TrueForge paired with GLM-5.2 averaged $2.9 per run versus $11.8 on Claude Managed Agents. GLM-5.2 completed the bench at $2.90 a run. Same-model TrueForge runs cost about 30% less.

VentureBeat reports $2.90 compared to $11.80, and $8.50 vs $11.80 on Opus 4.8. InfoWorld repeats those dollars and warns that TrueFoundry ran the bench, so the results have yet to be independently validated on larger production workloads. The docs also say TrueForge planned 19 tool calls per task versus 32 and 40 for the other two, and used about 40% of Claude Managed Agents' tokens.

deepagents is LangChain's open LangGraph harness. OpenAI Codex is the coding-focused CLI in VentureBeat's table, not a general enterprise agent server. VentureBeat lists Claude Managed Agents as Claude tokens plus $0.08 per running session-hour, and Codex-related spend around $100–$200 per developer per month. Those are VentureBeat's table figures, not an invoice.

USTA analysis: 100 identical bench-priced runs

USTA analysis (inputs are the vendor-published unit costs above; multiplier is 100 identical runs, shown so you can redo the arithmetic):

Sources for unit costs: TrueFoundry benchmark post ($8.5 / $2.9 / $11.8, reported by VentureBeat as $8.50 / $2.90 / $11.80). 100-run totals and deltas are $11.80×100, $8.50×100, $2.90×100, then CMA minus each row.

Same-model savings on 100 runs: $1,180 − $850 = $330 (28%). Cheaper-model savings: $1,180 − $290 = $890 (75.4%), the ratio TrueFoundry markets as "75%." That is bench unit price times a round count, not a production forecast. Test the ratio on your own MCP tools.

A manufacturer collecting supplier certificates can run that test without a new intake product. Keep the certificate-of-analysis collection path, point an agent at the same files, and compare token bills. Teams already routing those documents through US Tech Automations workflows will plug this in as a model swap, not a rebuild.

Who shipped it

TrueFoundry's three co-founders — Nikunj Bajaj, Abhishek Choudhary, and Anuraag Gutgutia — previously worked at Meta and WorldQuant, per VentureBeat. Gutgutia told VentureBeat customers asked for a vendor-neutral way to launch managed agents, and that TrueForge should sit alongside other harnesses while traffic still flows through TrueFoundry's gateway.

According to Intel Capital's 6 February 2025 announcement, TrueFoundry raised $19 million in Series A funding led by Intel Capital, with Eniac Ventures, Peak XV's Surge, Jump Capital, and angels including Gokul Rajaram and Mohit Aron, bringing total financing to about $21 million. The same release cited 4X year-over-year customer-base growth and 1,000+ clusters for ML workloads.

According to VentureBeat's January 2026 TrueFailover story, TrueFoundry was processing more than 10 billion requests per month through its AI Gateway. According to TrueFoundry's AI Gateway page, the gateway connects 1,600+ models, with a free tier of 50k requests per month — the commercial control plane, not the MIT harness.

On 24 June 2026 TrueFoundry acquired UK MLOps vendor Seldon AI. Seldon's site now folds the brand into TrueFoundry and cites 25,000+ MLOps professionals around Seldon Core. VentureBeat names NetApp as a beta user and Automattic as an early user.

What changes for a small or mid-size team

The change is not "install Kubernetes this week." The change is that the agent loop can live in a repo you can fork, on a laptop you already have, with the model as a config field.

For a two-truck HVAC shop, that looks like a dispatcher agent that reads voicemail, matches a customer in the CRM, and drafts a schedule SMS. Keep the CRM MCP server, try GLM-5.2 on routine tickets, and reserve a frontier model for ugly warranty disputes. The state of small-business automation is full of those split-brain jobs: most volume is boring; the tail is not.

For a 10-person agency, an executive-assistant style agent that files briefs and turns a form into a CRM record can keep the executive-assistant automation patterns. The harness behind them does not have to stay rented.

For a plant that already tracks work-instruction acknowledgements or changeover checklists, an agent that nags missing sign-offs is a loop over files and a tracker. That loop should not be rebuilt when the model vendor changes price. Shops tracking those acknowledgements in US Tech Automations do not need a new ticket system to test a cheaper harness. Start local with npx, keep it on localhost, and only move to Compose or Helm when more than one person must share sessions.

Honest limits

Open source is not governance. Gutgutia told VentureBeat that if you use only the open-source harness, you still need to put the right controls in it or in front of another internal system. SSO, budgets, PII masking, and unified traces are what the AI Gateway sells. The MIT license will not invent an audit log for a clinic.

Local mode is a personal tool. GitHub and the quickstart both say there is no login by default and data lives in a local SQLite file. Do not expose it.

The 75% figure is a vendor bench on 14 tasks, three trials in the docs table, one blind LLM judge, list prices, cache-aware. InfoWorld says it has not been independently validated on larger production workloads. Your MCP servers, retries, approvals, and sandbox minutes will move the number.

Self-hosting adds people cost. Jain told InfoWorld that managed services may remain cheaper for smaller or unpredictable workloads because the provider absorbs overhead. Su said token consumption is likely the largest share of agent operating cost, so self-hosting gets more attractive when you can actually use cheaper open-weight models and already have the engineering to run them. A solo clinic with a handful of agent runs a week may not beat Claude Managed Agents on total cost of ownership even if the token line looks better. Daytona vCPU rates, Helm, Postgres, and Redis sit beside tokens. "The harness is free" is true and incomplete.

Su also told InfoWorld the industry has not yet standardized context, routing, and tool use enough for the harness to be a fully independent layer. Jain does not expect a single winner in the way containers converged. MCP helps portability. It does not finish the job.

Signal vs Speculation

Demonstrated fact (sourced): TrueFoundry released TrueForge under MIT in mid-August 2026, with a public GitHub repo, npx local mode, and Compose/Helm hosted mode. It published a 14-task Enterprise-Bench comparison in which TrueForge matched Claude Managed Agents' solved-task band at lower listed token cost, including a GLM-5.2 configuration around $2.90–$3.0 per run versus $11.8. The company had previously raised a $19 million Series A (total financing about $21 million), launched a failover product on a gateway that it said processed more than 10 billion requests per month, and acquired Seldon AI in June 2026. MCP is a public standard whose mid-2026 revision is moving toward stateless scaling. Analysts quoted in InfoWorld said model flexibility can cut token cost and that self-hosting is a better fit for high-volume, well-staffed workloads than for small, spiky ones.

Our read: If the published token ratios hold on real SMB tools — CRM plus inbox plus PDF store — shops that already have MCP-shaped connectors will keep two model keys on purpose. The first savings will come from routing "file this PDF" work to open-weight models, not from rewriting the agent. Teams with no owner for keys, logs, and a Postgres box will stay on a fully managed assistant, and they should. We expect TrueFoundry to keep the MIT harness thin so the gateway stays the paid layer; that is the business model Gutgutia described. We do not expect the 75% bench ratio to appear intact on a 2-truck HVAC dispatch agent in year one, because sandbox time, retries, and approvals were not the whole of that $11.80. A 20–40% token cut on same-model work is a fair internal test target, because that is the band the Opus-vs-Opus row already shows.

Frequently asked questions

What is TrueForge?

TrueForge is TrueFoundry's MIT-licensed agent harness: the runtime that loops a model through tools, sandboxes, approvals, and session state on infrastructure you control. It is not a new model. It is the layer around the model.

How much cheaper is TrueForge than Claude Managed Agents?

On TrueFoundry's write-up, TrueForge plus GLM-5.2 averaged about $2.90 per run against $11.80 for Claude Managed Agents on Opus 4.8 (about 75% less), and the same Opus 4.8 model averaged about $8.50–$8.6 (about 30% less). Those are vendor-published bench means, not your invoice.

Can a small shop run TrueForge without a platform team?

You can run local mode with one npx command, SQLite, and Node.js 22.14+, on localhost only. Sharing it with a team means Compose or Helm, Postgres, Redis, and login. If nobody owns those pieces, a managed agent will likely stay cheaper in calendar time.

Does the MIT license include enterprise SSO and budgets?

No. The MIT license covers the software. Gutgutia told VentureBeat the open-source harness does not inherit company access policies by itself. Budgets, RBAC, and traces are the AI Gateway product.

Which models can I point TrueForge at?

InfoWorld reports OpenAI and Anthropic plus more than 20 others; GitHub lists OpenAI, Anthropic, Google Gemini, catalog providers, or any OpenAI-compatible endpoint. GLM-5.2 is the open-weight example in the published bench.

Is the 75% savings independently verified?

No. InfoWorld states that TrueFoundry conducted the benchmark and that results have yet to be independently validated across larger production workloads.

How is TrueForge different from deepagents or Codex?

deepagents is LangChain's LangGraph harness. Codex is OpenAI's terminal coding agent. TrueForge positions itself as a general-purpose production harness with a local-to-Helm path and sandbox-as-a-tool.

Glossary

• TrueForge: TrueFoundry's open-source agent harness (MIT), the runtime loop around a model.

• Agent harness: Software that plans, calls tools, sandboxes code, enforces approvals, and keeps session context so a model can finish a task.

• Claude Managed Agents: Anthropic's hosted, Claude-centric runtime for long-running agents; the comparison target in TrueFoundry's bench.

• Model Context Protocol (MCP): An open standard for connecting AI apps to external tools and data, often described as USB-C for models.

• Context compaction: Trimming or summarizing a long agent transcript so later turns do not replay every prior token; TrueForge's documented default threshold is 50,000 tokens.

• Sandbox-as-a-tool: Provisioning an isolated code/file environment only when the agent needs it, instead of keeping every turn inside a sandbox.

• AI Gateway: TrueFoundry's commercial control plane for model and MCP access, credentials, budgets, and traces; separate from the MIT harness.

• Enterprise-Bench: DevRev's public suite of cross-system enterprise agent tasks; TrueFoundry used 14 L1–L2 items in its published comparison.

TrueForge does not ask you to throw away the workflows that already file forms, certificates, and checklists. It asks whether the loop around those workflows still needs to be rented from the same company that sells the model. If you want that loop on rails you already operate, start from agentic workflows on the platform and keep the harness swap as a config change, not a project name.