Modular hardware-in-the-loop · Phase 1 running end to end
A slot-based test bench that lets AI agents, CI pipelines and engineers drive real microcontroller hardware, with timing kept on the module to well under a microsecond, fault injection on every module and a full audit trail. Priced like a good oscilloscope, not a car.
The missing link
Agentic development stops where the hardware starts. The bench gives an agent real stimulus, hardware timestamps and fault injection through MCP, behind the same safety rails a person would have.
An agent token can observe. Driving anything needs more.
Scoped signals, risk classes, an action budget and an expiry, approved by whoever holds the bench.
MCP is the protocol, not the security boundary. Voltage and timing limits are enforced on the module.
Who did what, under whose lease and grant, written to the audit log.
One clock
Each module stamps its edges, bytes and faults against the shared 10 MHz clock. A run reads like a logic analyser trace, and a reaction on the module shows up as a measured gap.
Illustrative run, not to scale. The 0.83 µs reaction is measured on the dev board; the rest follows the plan's timing.
What it solves
Industrial modular HIL is quote-only, commonly tens to hundreds of thousands of dollars, and tied to one vendor's toolchain.
→A modular bench targeted at $3–8k, under a manager's approval limit. Tests are plain pytest.
Flaky USB hubs, ports that rename, relay boards and scripts that only one person understands.
→One daemon, cards found from their ID EEPROM, hot-plug, a lease per bench and a safe state on any loss.
PREEMPT_RT's worst case is tens of microseconds and not guaranteed. A round trip through USB is milliseconds.
→Timing lives on the module MCUs. The host defines the plan and is never in the timing path.
Power loss during a flash write, brown-out, a dropped radio link: rarely tested, and when it is, by hand.
→Fault injection on every module, fired by a rule on the module the moment a condition occurs.
Giving a coding agent a relay and a power supply is either useless or dangerous.
→MCP tools generated from the capability registry, with grants, budgets, approvals and firmware limits.
A test passed on some bench, with some firmware, with some calibration. Which one?
→Plan hash, firmware version, bench ID and calibration go with every run. Runs that differ are marked not comparable.
Go deeper
Scroll through the chassis, one module, the backplane and the device under test, full screen.
Open the tour → 02Three tiers, the timing modes, what's new, and the module lineup.
Read the architecture → 03From agentic loops to production lines, with real test code and a comparison.
See where it fits → 04What's built, what's measured, what's next, and the design partner program.
Where we are →Design partner program
We're looking for 5–10 firmware teams who ship at least monthly and test on hardware by hand or with a DIY rig today. Paid pilots, early hardware and a direct line to the people building it.
Talk to us →