Know the cost before you commit.

Stipula turns a plain description of an AI project into a build, usage and support cost, routes it to the right approver, holds the spend to a monthly cap set from what they approve, and records what it came to. It runs on a Linux server you provide and holds your AI provider keys. One of our turnkey solutions: the gap an assessment finds, closed on your own site, installed by our DevOps engineers and supported by the tecnov8 desk.

4
stages, from a plain description to a recorded cost: estimate, approve, enforce, review
20 min
to install on one Linux server, with four answers
30 days
the full product with no licence, on your own server
2
editions: Standard, and Intelligence with the AI assistant and model recommender

The problem

When the board asks what AI is costing, you need a number.

AI projects start in more than one place at once, on personal cards and shadow keys, and no single person can say what the total is or whether any of it was approved. The two usual answers leave you no better off.

No figure

Say yes without a figure

Someone signs off a project on a guess. The real build, usage and support cost only turns up later, and by then the spend has already happened.

Blocked

Say no to everything

Refuse AI across the board and the projects run anyway, off the books, where the spend cannot be seen, approved or capped.

Stipula

Cost it, then decide

A plain description becomes a real cost, routes to the right approver by what it costs and how sensitive its data is, and is held to a monthly cap set from what you approved. Every project seen and signed for in one place.

Discovered at the quarter, not the start

Spend on personal cards and shadow keys surfaces on an expense claim or a renewal notice, long after anyone could have held it to a cap. The bill is real whether or not the project was ever approved.

Nowhere to look for the total

When the spend is split across a dozen tools and as many invoices, there is no one place a budget holder can see the whole AI bill. Stipula gives you one number for every project, the approver who signed it off and what it has spent against its cap.

Asked who approved it?

With Stipula every estimate, the approver it went to and the decision made are kept, so you can show how each AI project was signed off.

How it works

One appliance on a Linux server you provide.

People work in a browser over HTTPS, signed in through your directory. Stipula runs as a small set of containers on one Linux server: the application that estimates, approves and records, and an AI gateway that holds your provider keys and meters spend against each cap, with two isolated databases. It reaches out only to the AI providers and the directory you point it at, email and webhooks if you switch them on, and a daily update check that can be switched off.

The application

Where estimates are built, routed and approved, and where the review records what each project came to. A guided form turns a plain description into a build, usage and support cost, with no technical knowledge needed. It holds the project database.

The AI gateway

Holds your AI provider keys and never hands them out. Each person on an approved project gets their own virtual key, and the gateway meters every pound against the project's cap, live, with a monthly ceiling and budget alerts before it is reached. Its database is kept apart from the application's.

The console

Where your people work, signed in through your directory: estimates, approvals, the cap and the record, in one browser tab. Budget holders see committed spend this financial year, the yearly run-rate, the pipeline value and live AI spend.

What it works out

Build cost, one-off
Usage cost, a year
Support cost, a year
Build or buy verdict

What stays on your server

Projects and estimates
Provider keys
The audit log
Encrypted backups

Everything Stipula needs runs on your server. With the AI assistant on, the descriptions and documents you allow go to your AI provider, and nothing else leaves. The guided form needs no technical knowledge; the detailed form sets the models, the coding assistant seats and the assumptions yourself, when you want to.

Estimate, approve, enforce, review

Four stages, in the order they happen.

Every AI project goes the same way, in one place. A plain description becomes a cost, the cost is approved, the approval becomes a cap the gateway holds, and the review records what it really took.

Stage	What happens	What you get
Estimate	Answer who will use the project and how often, how demanding it is, and whether to build or buy. The guided form needs no technical knowledge.	A one-off build cost, a yearly usage cost and a yearly support cost, and a build or buy verdict once a supplier's quote is added.
Approve	Submit the estimate and it routes automatically, by its first-year cost and its data classification, to the people who should decide. Nobody can approve their own.	A decision from the right approver, and who made it, kept with a PDF approval pack.
Enforce	The approved figures set a monthly spending cap. Each project member gets their own AI key, and the gateway meters every pound against the cap.	Live spend against the cap, with optional alerts at 50, 80 and 100 percent.
Review	90 and 180 days after approval, the owner records what the project actually cost and whether it delivered.	A better estimate next time, and estimates you can compare side by side.

Guided or detailed

The guided form asks a few plain questions. The detailed form sets the models, the coding assistant seats and the assumptions yourself.

Build or buy

Add a supplier's quote and Stipula says build, buy or too close to call, on the build, usage and support figures.

Compare side by side

Put estimates next to each other, with the verdict and the break-even point, and download the reports.

A sample project

One plain description, costed, approved and capped.

A customer support chatbot for the sales team, walked through the stages. The figures are the maker's worked sample, not a quote.

Describe
A customer support chatbot for the sales team
In plain words: who will use it, how often, how demanding it is, and whether to build or buy.
Estimate
A one-off build cost and two yearly costs
Stipula works out the build cost, the yearly usage cost and the yearly support cost from the description.
Verdict
Build or buy, decided on the figures
Set against the supplier's quote, the figures for this project come down on build.
Build
Approve
Routed to the right approvers
By its cost and how sensitive its data is. The approvers sign off the number, and the decision is kept.
Approved
Cap
A monthly cap of £678, metered live
Each project member gets their own AI key. £403 spent this month against the cap, with budget alerts before it is reached.
Within cap
Review
What it actually took, recorded
At 90 and 180 days the owner records the real cost and whether it delivered, so the next estimate is better.

£36,976

The one-off build cost in the sample, worked out from the description.

£6,510

A year of usage, metered through the gateway.

£9,660

A year of support, costed in from the start.

A sample. The project, its supplier's quote, the figures and the name are fictional (Northwind); every stage is something Stipula does today.

On the record

Every estimate, decision and pound, kept in one place.

The audit history is tamper-evident: rows sealed in a keyed hash chain, signed anchors, and a full check every day. The estimate, the routing, the decision and who made it are all kept.

step 01
Estimate submitted
Build, usage and support cost
step 02
Routed to the approver
By cost and data sensitivity
step 03
Approved, and who decided
The decision is kept
step 04
Cap set and metered
Live, against the gateway

Illustration. The estimate, the routing, the decision and who made it are kept, and the audit history records what was done.

A decision that is kept

Every estimate carries its build, usage and support cost, and the approval, and who made it, is held on the record. Nobody can approve an estimate they created or submitted, and the attempt is recorded. A board and audit pack PDF covers spend, approvals, overrides and data-protection events.

Spend held to the cap

The approved figures become the cap. The gateway meters every pound against it, live, with a monthly ceiling and budget alerts at 50, 80 and 100 percent, and a capped project's top-up goes to an approver, never the owner.

Your data on your side

Two isolated databases that cannot reach each other. A PII guardrail is installed automatically when the server meets its requirements and runs by default, masking or blocking personal data in prompts before they reach an AI provider. It is tested every day. Only usage figures are shown, never prompt content.

Signed in to the same standard

Local accounts, Active Directory, Microsoft Entra ID, OpenID Connect and SAML, with SCIM provisioning and two-factor by default for administrators who use a password. Roles of user, viewer, admin and super admin; approving is assigned, not a role.

The appliance holds your keys, keeps a tamper-evident record and takes updates with an automatic rollback.

Yours to run

On your server, holding your keys.

Four steps from a spare server to a costed first project. tecnov8 does them with you, or for you.

Copy one installer package to a server

One signed installer package onto a 64-bit Linux (amd64) server with systemd and Docker. It is checked against the Stipula release key before anything installs.

Run the installer

About 20 minutes, four answers: the hostname, the TLS choice, an admin email and whether to run the guardrail. Or unattended, from an answers file.

Accept, then activate

A super admin accepts the licence agreement. The 30 day evaluation runs from first start; when you buy, a super admin downloads an activation request and tecnov8 sends the licence file back.

Add your first project

Build an estimate from a plain description, and route it for approval.

Server: 64-bit Linux (amd64) with systemd. Two CPU cores, 8 GB of memory, 10 GB with the guardrail, and 40 GiB (about 43 GB) of free disk.
Containers: A small set under Docker Engine with the Compose v2 plugin: the application, the AI gateway, the guardrail and one database service holding two isolated databases.
Network: HTTPS on ports 80 and 443, no other inbound ports. The certificate is internal, Let's Encrypt or your own. Outbound to the AI providers, your directory and the update service, which can be switched off.
Sign-in: Local account, Active Directory, Microsoft Entra ID or SAML, with SCIM provisioning. Two-factor required by default for administrators who sign in with a password.
AI gateway: Chat and completions, the Responses API, Anthropic Messages and embeddings; Azure OpenAI through API Management. Each key limited to the approved models and the monthly budget.
Backup: Encrypted daily backups on the server, kept 30 days, plus a backup command to schedule to an off-server share. Restore onto a replacement server with your vaulted keys and a replacement licence.
Updates: In the app: the release's signature is checked, a restore point is taken, and it switches back on its own if the health check fails. Or from a signed installer package.

Licensed to your installation

The licence is a file tied to the installation. A super admin downloads an activation request and installs the licence file that comes back.

Two editions. Standard covers the four stages. Intelligence adds the AI assistant, which fills an estimate in from a short description for you to check, and the model recommender, which picks the least costly model meeting the quality you need.

A late renewal is not an outage.

Evaluate
30 days, the full product, no licence.
Activate
Install the licence file.
Run
Your data on your server.
Update
In the app, signed, with automatic rollback.
Renew
A renewal request from the Licence page.
Grace
30 more days, fully working, then read-only with nothing lost.

If a renewal is late, Stipula keeps working fully for 30 days, then turns read-only: nothing is lost, and installing the licence brings it straight back. An evaluation that ends unlicensed opens only the Licence page until a licence is installed, with the data kept.

Try it

Start with 30 days.

Every new Stipula installation runs the full product for 30 days with no licence, Intelligence features included, on your own server, with your own projects.

Install it on a server

One signed installer package, about 20 minutes, four answers. tecnov8 can do this step with you on a call.

Build an estimate

Describe a project in plain words and get a build, usage and support cost, and with a supplier's quote, a build or buy verdict.

Approve and cap it

Route it for approval, then let the gateway hold the spend to the cap, live.

The tecnov8 ecosystem

One of our turnkey solutions, inside one ecosystem.

01 · Assess
A tecnov8 assessment finds the gap: AI projects running with no approved budget or spending cap, and nobody able to say what the total is.
The assessment menu →
02 · Close
Stipula closes it on your own site: every project costed, approved on the merits and held to a cap, with a record an auditor can check.
All turnkey solutions →
03 · Install
Our DevOps engineers stand up the appliance, connect it to your directory and your AI providers, and set the approval levels with your finance team.
How we install and run it →
04 · Support
App Guardian stands behind it. The tecnov8 desk is first line, so there is one number to call and one record of what happened.
Application support →

Your Stipula partner

Hardened, installed and supported by tecnov8.

Stipula is sold directly and through IT partners. tecnov8 is one. We harden it before it goes in, install it with you or for you, and the tecnov8 desk is first line support for as long as you run it. Book a demo, book a call or ask for the brochure below.

How we support it →

Microsoft, Active Directory, Azure and Entra are trademarks of the Microsoft group of companies. Docker is a trademark of Docker, Inc. Anthropic is a trademark of Anthropic, PBC. Linux is a registered trademark of Linus Torvalds. All other trademarks are the property of their respective owners. Facts as of Stipula 3, October 2026.

Delivered by tecnov8

Installed, rolled out and supported by one team.

01 · Prove

Proof of concept

A 30 day evaluation of the full product, Intelligence included, with no licence, on your own server.

02 · Roll out

Onboarding

Install the appliance from the signed installer package, connect it to your directory and your AI providers, set the approval levels with your finance team, and switch the gateway keys on project by project.

03 · Run

Support

The tecnov8 desk is first line support. You run Stipula yourself, with the desk behind you.

See Stipula on your own estate

Three ways to take the next step. Pick the one that suits where you are today.

Talk

Thirty minutes on your estate, where you are today and whether Stipula fits.

See

A live walk through of Stipula with an engineer, on a call.

Read

The Stipula brochure sent straight to your inbox, to share with your team.

Each product links to its own dedicated vendor page. Hardened, installed and supported by tecnov8.