Not a Coding Tool but an Agent Runtime: Running Your Own Agent Layer with TrueForge

NeuralNine frames TrueForge from the first minute as infrastructure, not as another coding assistant next to Pi or OpenCode. The product is an MIT-licensed harness from TrueFoundry that supplies the runtime layer for general-purpose agents. I find the framing useful because it sets the right expectation: this is the backbone of an agentic system, not a chat window.

The video then explains what a harness actually covers, and I think this is its best two minutes. A language model by itself only maps input tokens to output tokens, while an agent also needs tool calls, MCP connections, code sandboxes, approvals from a human and waiting states. TrueForge owns that surrounding loop: reasoning, acting, reading results and continuing until the job is done. Once you see the loop as the product, the comparison with Claude Managed Agents starts to make sense.

Before the demo, the presenter puts a disclosure on the table: the video is a paid collaboration with TrueForge, though the opinions are his own. I credit that openness, and it matters for everything that follows because the cost claims originate from the vendor. The invitation is to try everything shown for free and to star the repository, so at least the viewer can verify the basics without paying.

The benchmark comes from an X post and covers 14 tasks from DevRev's Enterprise-Bench, spanning CRM records, issue trackers and document stores. Three configurations are compared: Claude Managed Agents with Opus 4.8, TrueForge with the same Opus 4.8, and TrueForge with the open GLM-5.2 model. Reported completion is 11 of 14 in each comparable run, with per-task cost at 11.80 dollars managed, 8.50 dollars on TrueForge with the same model, and 2.90 dollars with the open model. That is roughly 30 percent cheaper on equal models and 75 percent cheaper with the open model.

TrueFoundry itself is worth a sentence of context. It is a San Francisco machine-learning infrastructure company founded in 2021 by former Meta engineers, and it already sells enterprises a paid gateway for model access, credentials and budgets. TrueForge sits above that gateway as the open execution loop, which explains the business logic: give away the harness, keep the control plane. Forbes describes the same split, with the 50 percent saving headline needing qualification.

Installation is deliberately boring, which I mean as praise. You copy an NPX command from the GitHub repository, run it once, and it downloads the package, creates a local SQLite database and serves a page on localhost. The browser view that opens is your control panel, and from there you attach whatever models you like. A local-first start like this lowers the trial barrier to nearly zero.

Model setup covers the big providers plus a self-hosted escape hatch. You can paste an OpenAI key, do the same for Anthropic or Gemini, or register a custom provider with a base URL pointing at Ollama. The video walks through key creation on the provider console and then shows the models appearing inside TrueForge. That vendor-neutral shelf, including local models, is the core promise of the whole project.

The chat view that greets you is explicitly a laboratory, not the product. You open a new chat, toggle connectors such as Linear, Notion or Exa web search, and experiment with prompts before anything is permanent. A visible save button marks the boundary between playing and shipping. I like this separation because it forces you to prove the setup works before you name it and expose it.

The first full example builds a stock research assistant. A throwaway test prompt asks for Nvidia financials, the agent fans out into tool calls and a sub-agent, and it visibly digs through 10-Q and 10-K style filings before summarizing. Then comes the key lesson: none of that chat history becomes the agent. The saved definition is a short system prompt plus the chosen model, connectors and skills, after which a bare MSFT input produces a fresh Microsoft rundown.

The payoff section moves the agent out of the browser and into Python. There is no Python SDK, only a TypeScript one, so the demo uses plain requests against the local API: open a session with the agent name, keep the returned session id, post a user message and read back a server-sent-events stream. The script filters the stream for model message deltas and prints them live, and a Tesla query returns streamed findings while the dashboard shows each agent step. A short program like this turns the harness into a genuine runtime layer for your own apps.

The second agent raises the stakes with a Daytona code sandbox. After creating a full-access API key and pasting it into the sandbox settings, the agent can execute Python instead of merely describing numbers. Asked for Meta financials as a PDF with Matplotlib plots, it searches the web, runs the code remotely and hands back a downloadable report with generated charts. Saving that as a second agent with charting baked into its instructions makes every later ticker, such as Nvidia, follow the same pipeline automatically.

Skills get their own compact demonstration through algorithmic art. Without any skill attached, a creative request yields a plain SVG or HTML file; with the dedicated skill enabled, the agent first reads the skill instructions and produces a visibly more professional piece with adjustable parameters. The contrast is the whole argument for skills in miniature: packaged know-how beats raw generation. GitHub-importable skills plus custom MCP servers are what turn the trivial web-search demo into something proprietary.

The closing stretch returns to vendor neutrality with local models. A custom provider entry aimed at the standard Ollama base URL, a manually entered model id from the Qwen family with around 27 billion parameters, and pasted context and output limits are enough to bring a home machine into the same workflow. The presenter runs it from his own workstation hardware and gets answers through the identical loop that earlier served the OpenAI model. Everything open, everything owned: that is the note the video ends on, next to a reminder that the project lives or dies by community stars.