Authorize
checks the agent's ceilings and has the action risk-scored.
Ferrule is the gateway your agents call to act in your own tools.
The agent drafted standups and opened pull requests for six weeks.
Then procurement asked what it costs at forty people, and where content is processed.
Nobody had either answer in writing, and the pilot ended there.
A model risk-scores every action and a rate card prices it before it runs. A log line records it after.
€0.004
per executed action, on a published rate card
EU only
executed and risk-scored in Sweden, never copied out
10,000 actions
free, no card, metered and logged like paid ones
Compare
Against CORE, on procurement's questions
| What a buyer checks | Ferrule | CORE |
|---|---|---|
| Cost at 40 people | ✓€0.004 per action, published | ✕Usage, storage and tier fees, unpublished |
| Processing region | ✓Stockholm and Azure Sweden Central | ✕None named |
| Model providers | ✓Azure OpenAI, zero retention, to score risk | ✕OpenAI, Anthropic or Google Gemini |
| No-training commitment | ✓Yes; baselines fit one tenant's data only | ✕Google user data only |
| Apps out of the box | ✓New MCPs on request | ✓50+ through one MCP endpoint |
Scope
Buy CORE if you need memory or planning
- ✕We do not model you - no knowledge graph or entity resolution.
- ✕We do not plan - even your agent's bad plans run as asked.
- ✓Only on request - we run nothing against your tools on our own schedule.
Pricing
€0.004 per action
One rate for every tool and operation.
€0.004
per action
- €0.0028 above two million actions a month, applied automatically
- Refused, rate-limited and failed third-party calls are not charged
- No seat price, minimum, annual prepayment or onboarding fee
- First 10,000 actions free, no card, and you see what they would have cost.
- 99.9% monthly availability; a month below it is credited automatically.
- Rolling monthly with no notice period; logs export even after you leave.
Why it works this way
The public rate card loses us the largest deals
A published rate is easy to undercut. Some deals a negotiated quote would win, we lose on arithmetic.
An unpriced platform never reaches procurement, only a sales cycle that takes a quarter.
The terms name the region too: Stockholm for the gateway and log, Azure Sweden Central for both models.
Features
Four things the gateway does
Each one leaves a record you can export.
Price
quotes the public rate card before the action runs.
Execute
makes the call with your grant and reports the result.
Log
records the call, risk score, price, authorizer, result and trace id for 400 days.
The AI layer
Two models, and neither can make an action happen
A risk classifier inside Authorize, and anomaly detection over the Log.
A routine action and a harmful one can share an endpoint and parameter shape. Only a model can read what the parameters say.
Risk classifier - scores tool, operation and parameters for risk and intent.
Over your risk ceiling - refused, or held for your named person.
Learning - that person's decisions tune only your tenant's risk threshold.
Anomaly detection - flags an agent's new tool, new volume or rising refusals.
Never approve - a fooled model passes only what your ceilings allow.
Runs in Sweden Central - Azure OpenAI with Azure Machine Learning.