Customers

Only written permission puts a name on this page.

Nobody has given us that permission yet, so there are no logos here, no testimonials, and no anonymised “a leading bank”. What stands in their place is three situations built entirely out of behaviour that has shipped, with the part AGI declines to do written next to the part it does.

Try AGI Web See what a plan includes

Worked scenarios

A scenario that only lists benefits would be an advertisement.

None of these is a customer. Each is a configuration we have built, named down to the command that proves it: the surface it runs on, the mechanics underneath, the plan limit it meets, and the work it hands back to you. Follow the link under any of them and you can check the claim against the product yourself.

Engineering · Local mode
A team whose source cannot reach a third party
An NDA forbids sending this repository to an external model provider, which disqualifies most AI coding tools on the question of where inference happens, long before anyone gets to judge how well they answer.
- Desktop keeps one server URL per runtime — Ollama, LM Studio, llama.cpp, and vLLM — and fills its model picker from whatever that server answers with.
- agi models status names every local server it can reach and the models sitting on each. A base URL that is not loopback never gets a request built for it.
- No AGI account takes part, and a finished run prints “no cost — local model” because there was nothing to bill.
- The Local Mode plan tier is priced at zero and puts no cap on work you run against your own hardware.
What it does not do. AGI supplies neither the hardware nor the weights, and a model you host will lose to a frontier hosted model on the hardest problems. That is capability traded for containment, and it is a trade you have to want.
Desktop · Linux assets · v1.2.0CLI · Coming soon
How Local mode works
AGI · local model · ollama(local)in 3120 · out 640 · $0.0000 · ctx 9%
› agi models status
Local model servers
ollama running http://localhost:11434
models:
- qwen2.5-coder:14b
› agi exec --provider ollama "who logs a raw tenant id?"
src/ledger.rs logs it twice unredacted: the retry
branch and the error mapper.
cost: Tokens: 3120 in / 640 out (no cost — local model)
Default · local · effort:Medium · sandbox: seatbelt
Consulting · BYOK
A firm that charges model spend back to a client
Every engagement has to carry its own AI cost, and the firm will not put a platform vendor between itself and the invoice it has to justify line by line to the client paying it.
- Each surface writes the key into its own platform credential store, so one engagement’s key stays on the one machine that engagement runs from.
- Requests reach the provider endpoint directly, so the usage lands on the firm’s own provider account and shows up on the firm’s own invoice.
- The command above prints the same catalog rate card the CLI estimates from, so the arithmetic behind a cost line is inspectable before a run starts.
- --max-budget-usd ends a run once estimated spend passes a ceiling you set, which keeps a runaway loop off a client’s bill.
- The Local Mode + BYOK plan tier is priced at zero, which leaves the provider invoice as the only bill in the arrangement.
What it does not do. Per-client attribution comes out of your provider’s own billing tooling, one key or project per engagement. The number the CLI prints is an estimate from a local rate table; AGI never sees your invoice and writes no per-client spend report for you.
Desktop · CLI · VS CodeNo AGI charge on this route
How BYOK works
IT services · Unattended runs
A provider repeating the same checks across client estates
Overnight alert triage, certificate expiry sweeps, and ticket intake repeat across dozens of client environments, and doing them by hand makes coverage a function of headcount.
- agi --daemon reads ~/.agiworkforce/triggers.json and runs what it finds there: cron schedules, webhook endpoints, and filesystem watchers.
- An unattended session holds three tools — read_file, search_files, and list_directory. Anything mutating is denied outright, because no one is sitting at the prompt to refuse it.
- Each firing writes one JSON record into ~/.agiworkforce/daemon-logs, restricted to its owner, with known secret patterns redacted out of the prompt and the response first.
- Run this on managed cloud instead and the gates change: AGI Work needs Pro or above, and scheduled tasks are capped per plan at 2 on Basic and 5 on Pro.
What it does not do. A trigger can read and report; it cannot edit a file or run a shell command, so remediation still lands on an engineer. You also have to write the runbook each trigger follows, because AGI will not discover your clients’ processes on your behalf.
CLI · Coming soonComing soon
How the CLI runs unattended

Our policy on names

A name appears here only after someone has put it in writing.

We publish a customer name only with written permission, and we do not fall back on anonymised descriptions specific enough to identify someone who never agreed to be identified. Until a customer says yes in writing, this page stays exactly as you are reading it.

Permission

We will ask you before your name ever appears here.

AGI Web is available now, and every other surface reports its own state on the download page rather than in a claim on this one. If AGI earns a place in how you work and you are willing to say so in writing, this page will finally carry something other than worked examples.

Tell us we can use your name

Nothing on this page is a customer reference.

Only written permission puts a name on this page.

A scenario that only lists benefits would be an advertisement.

Engineering · Local mode

A team whose source cannot reach a third party

An NDA forbids sending this repository to an external model provider, which disqualifies most AI coding tools on the question of where inference happens, long before anyone gets to judge how well they answer.

Desktop keeps one server URL per runtime — Ollama, LM Studio, llama.cpp, and vLLM — and fills its model picker from whatever that server answers with.
agi models status names every local server it can reach and the models sitting on each. A base URL that is not loopback never gets a request built for it.
No AGI account takes part, and a finished run prints “no cost — local model” because there was nothing to bill.
The Local Mode plan tier is priced at zero and puts no cap on work you run against your own hardware.

What it does not do. AGI supplies neither the hardware nor the weights, and a model you host will lose to a frontier hosted model on the hardest problems. That is capability traded for containment, and it is a trade you have to want.

Desktop · Linux assets · v1.2.0CLI · Coming soon

How Local mode works

AGI · local model · ollama(local)in 3120 · out 640 · $0.0000 · ctx 9%

› agi models status

Local model servers

ollama running http://localhost:11434

models:

- qwen2.5-coder:14b

› agi exec --provider ollama "who logs a raw tenant id?"

src/ledger.rs logs it twice unredacted: the retry

branch and the error mapper.

cost: Tokens: 3120 in / 640 out (no cost — local model)

Default · local · effort:Medium · sandbox: seatbelt

Consulting · BYOK

A firm that charges model spend back to a client

Every engagement has to carry its own AI cost, and the firm will not put a platform vendor between itself and the invoice it has to justify line by line to the client paying it.

Each surface writes the key into its own platform credential store, so one engagement’s key stays on the one machine that engagement runs from.
Requests reach the provider endpoint directly, so the usage lands on the firm’s own provider account and shows up on the firm’s own invoice.
The command above prints the same catalog rate card the CLI estimates from, so the arithmetic behind a cost line is inspectable before a run starts.
--max-budget-usd ends a run once estimated spend passes a ceiling you set, which keeps a runaway loop off a client’s bill.
The Local Mode + BYOK plan tier is priced at zero, which leaves the provider invoice as the only bill in the arrangement.

What it does not do. Per-client attribution comes out of your provider’s own billing tooling, one key or project per engagement. The number the CLI prints is an estimate from a local rate table; AGI never sees your invoice and writes no per-client spend report for you.

Desktop · CLI · VS CodeNo AGI charge on this route

How BYOK works

IT services · Unattended runs

A provider repeating the same checks across client estates

Overnight alert triage, certificate expiry sweeps, and ticket intake repeat across dozens of client environments, and doing them by hand makes coverage a function of headcount.

agi --daemon reads ~/.agiworkforce/triggers.json and runs what it finds there: cron schedules, webhook endpoints, and filesystem watchers.
An unattended session holds three tools — read_file, search_files, and list_directory. Anything mutating is denied outright, because no one is sitting at the prompt to refuse it.
Each firing writes one JSON record into ~/.agiworkforce/daemon-logs, restricted to its owner, with known secret patterns redacted out of the prompt and the response first.
Run this on managed cloud instead and the gates change: AGI Work needs Pro or above, and scheduled tasks are capped per plan at 2 on Basic and 5 on Pro.

What it does not do. A trigger can read and report; it cannot edit a file or run a shell command, so remediation still lands on an engineer. You also have to write the runbook each trigger follows, because AGI will not discover your clients’ processes on your behalf.

CLI · Coming soonComing soon

How the CLI runs unattended

We will ask you before your name ever appears here.

Nothing on this page is a customer reference.