Your model, your machine, every call on the record.
Herth builds AI agents for research labs and organizations with sensitive data. Each one runs entirely on hardware you own.
One run, as recorded
Every action the agent took, in order. One was blocked, and one waited for a yes. This run is a sample.
| Time | Action | Target | Status |
|---|---|---|---|
| 09:14:02 | load | ~/models/Qwen3.8-27B-Q4_K_M.gguf | ok |
| 09:14:07 | read | ~/lab/run041.csv | ok |
| 09:14:08 | read | ~/lab/protocol.md | ok |
| 09:14:15 | tool | fit_curve | ok |
| 09:14:31 | net | 203.0.113.9:443 | blocked |
| 09:14:32 | tool | plot_fit | ok |
| 09:14:47 | write | ~/lab/fit041.svg | ask |
| 09:15:03 | write | ~/lab/fit041.svg | ok |
| 09:15:03 | end | run complete | ok |
| 61 s | 9 rows, 1 ask, 1 blocked | 0 B sent |
The run ends with a file on your disk.
Here the agent read a set of measurements, fitted a curve to them and asked before it wrote the result. The figure is the fit it saved.
Nothing leaves. Everything is on the record.
The model runs on your machine.
So do the tools it calls and the record of what it did. Herth is built for local inference from the ground up, from the model to the harness.
You see every step.
You can see what the agent is doing at each step. That is how it earns your trust.
Your data stays where it is.
Sensitive lab and organizational data stays on your machine. In this run the agent reached for the network once, and the call was blocked.
Tool calls are checked before they run.
Each call is gated before it executes, and outputs are checked for hallucinations, so you can depend on the results.
For work that cannot leave the building
- Research labs
- Run the analysis next to the instrument. Unpublished data stays on the lab's own machine.
- Organizations with sensitive data
- Put an agent on records, contracts and internal documents without handing them to a vendor.
We build a toolkit for your field
Field toolkits sit on top of a base research toolkit that we build for any lab, and more fields are added per client.
- Chemistry
- Math
- Physics
- Statistics and plotting
Four phases, from your lab to a working agent
Every deployment goes through the same four phases, and ends with the whole agent benchmarked on your hardware.
- 1 PreparationWe learn the machine, the work and anything of yours the model could learn from.
- 2 ModelA strong open-weight model that fits your hardware. It can be fine-tuned, or distilled from a stronger one.
- 3 HarnessContext, gate, verify, tools and transparency turn a local model into a working agent.
- 4 FinalizationQuantized to fit your hardware, integrated with the harness and benchmarked there.
Get Herth
Tell us what you work on and what you would hand to an agent first. Every deployment starts with that conversation.