Candid · Flagship product
AI agents that get cheaper as they work.
Candid is designed to get faster and cheaper on work it has done before, to take a change of mind partway through a long task without starting over, and to run as small models on ordinary hardware, even offline.
01What it will do
Five outcomes.
Gets cheaper with experience
Status: In researchMost AI pays full price for every task, even one it has solved a thousand times. Candid is designed to get faster and cheaper on the kinds of work it has already done.
Correctable mid-task
Status: In researchChange your mind halfway through a large job, say a big software build, and Candid updates what needs to change, checks what else that affects, and carries on. No starting over. No quietly breaking earlier work.
Runs where people are
Status: In researchSmall versions designed to run on ordinary laptops, offline, so data can stay on the device and costs stay predictable. Larger versions in the cloud or on a company’s own servers.
Tested, not just described
Status: PlannedA virtual lab for electronics, robotics and engineering experiments, with key predictions checked against real measurements on a physical bench, and every result published with what’s needed to reproduce it. The bench doesn’t exist yet.
Open where it matters
Status: PlannedWe plan to release small models openly.
02Who it’s for
Built for people who pay for compute, or can’t.
Developers and software teams
Long software builds you can redirect halfway through, without starting over.
Engineers, hardware startups and universities
Experiments run in simulation, with key predictions checked against real measurements.
Schools and learners
Small models designed for the ordinary laptops people already have, working offline.
Organisations that keep data on their own infrastructure
Versions designed to run on your own servers, so data stays where it belongs.
03Where it runs
Where people already are.
On a laptop, offline
Small versions designed for ordinary laptops. Data can stay on the device and costs stay predictable.
In the cloud
Larger versions for heavier work.
On your own servers
For organisations that need to keep data on their own infrastructure.
Designed forWindowsmacOSLinux
04Roadmap
What we plan to show, and when.
Status: Planned Timing depends on funding. These are plans, not results.
Nov 2026
- A physical lab bench.
- The first simulated prediction compared with a real measurement.
Dec 2026 – Jan 2027
- First agent demo: it takes a mid-task change on a large software project and finishes faster than starting over.
- The first small model running offline on a laptop.
Feb – Mar 2027
- Private beta.
- A robot-arm task learned in simulation and repeated on a real arm.
- Measurements of repeated tasks getting cheaper.
Apr 2027
- Public demo.
- An openly released small model.
- Reproducible experiment packs.
05Business model
How it will pay for itself.
- A usage-based API for teams.
- Licences for self-hosted versions.
- Lab services for engineering teams and universities.
- Education packages.
- Small models, free and open.
Candid has no customers or revenue today. This is the intended model, not a description of current business.
06Design partners
Help shape it before it exists.
We are looking for a small number of teams to talk to while Candid is designed: teams running long software projects, engineering groups and universities that run experiments, and organisations whose data has to stay on their own infrastructure.
07How it works
The technical approach.
This page describes what Candid is designed to do, not how. That is deliberate while the work is unpublished. Candid is its own architecture, to be built from scratch once the lab is funded.
Technical depth
We’re happy to walk researchers, programmes and investors through our technical approach in confidence. Get in touch.