Appearance
Guardrails
Guardrails holds every limit on the model calls your account makes through Kadmo, on one page: what may be spent, where the models may run, which models may be called, what Kadmo checks in your chats, and how fast calls may come. Budgets, regions, models and the call rate hold for every lane alike: Ask Kadmo, Slack, Microsoft Teams, Goals, chat agents and your agents' jobs. Content rules apply to the chats.
Where it is
Open Settings, then Guardrails under Configure, right after BYOK (app.kadmo.ai/guardrails). The command palette (⌘K) finds it as Guardrails, and its two parts as Budgets and Data residency.
- Every member can read the page. Only admins change it; a member reads You have read-only access — an account admin sets budgets, regions and content rules., every box disabled and no Set budget or Change.
- There are no guardrail objects to create and nothing to assign: the scopes are the Whole account and each workspace. A limit is set where it is read.
The page has five panels, in this order.
Budgets
The account's budgets: Whole account, one row per workspace, and No workspace. Each budget shows what was spent against its limit, a month or in total, and a status at 80 % and at the limit. Set budget on a row without one, or Change on a row that has one, opens the Budget dialog. How budgets work, and what a member sees when one is reached, is on Models and credit › Budgets.
Every link to a budget lands here: the banner, Needs you, the run dialog, the job page and the budget e-mails. Settings › Billing keeps one line that reads the budgets (Budgets · €237.20 of €400 a month, checkout-service capped) and links here.
Data residency
Where the account's models may run. Each row has three boxes, EU, US and Global, and at its right what they leave and how many models a call there may pick.
| Row | What it sets |
|---|---|
| Whole account | The regions every call of the account may run in. All three by default. |
| A workspace | Narrows the account's regions for that workspace's calls. A workspace that sets none reads as the account, its boxes drawn in the inherited look. A region the account does not allow is disabled in every workspace row. |
A box saves as you change it and a toast confirms Regions saved. A row keeps at least one region: unchecking the last one is refused, because a workspace with none reads the account's, and an account with none allows every region.
The words at the right read the regions that apply: every region, EU only, EU and US, followed by the number of models a call there may pick: models Kadmo serves here, your plan includes and no admin switched off, in those regions. Which providers run in which region is on kadmo.ai/pricing; EU only can read 0 models while no EU provider is served.
A row warns in words when its regions leave out the default model's region (Claude Sonnet 5.5, the default model, runs in the US — calls here that use it are refused until a model served in the EU is picked.) or leave no region at all.
What a call meets. Kadmo checks every call against the account's regions and the workspace's, before anything leaves Kadmo. A call on a model outside them is refused with the reason and the fix:
| Case | What the caller reads |
|---|---|
| The account leaves the model's region out | This account allows EU models only; claude-sonnet-5 runs in the US. Pick a model served in the EU, or change the account's AI regions on Guardrails. |
| The workspace leaves it out | This workspace allows EU models only; claude-sonnet-5 runs in the US. Pick a model served in the EU, or change the workspace's AI regions on Guardrails. |
| No region is left | No region is left for this workspace: the account and the workspace allow no region in common. Allow at least one on Guardrails. |
Which plans include data residency is on kadmo.ai/pricing. On a plan without it, the panel reads Business and up and shows the regions with every box disabled; budgets and the models' switches work on every plan.
Models
One row, Available models, with the providers Kadmo serves here and where they run, the default model, and how many models are switched on. It opens Settings › Models, where an admin switches models on and off and sets the default.
A model an admin switched off is refused for every lane, pinned agents included: Claude Opus 5.5 is switched off on this account — pick another model. The model pickers leave it out, as before.
Content rules
What Kadmo checks in your chats — Ask Kadmo, Slack, Microsoft Teams, Goals and chat agents — and what it does with what it finds. Kadmo reads every chat message before it leaves for the model. For every account it counts what the rules find and changes nothing: the panel's baseline reads Kadmo's baseline: audit. An admin can have a group, or one of the account's own patterns, mask or block instead.
| Group | Detectors | Actions |
|---|---|---|
| Personal data | e-mail address, phone number, card number, US Social Security number, Canadian SIN (off by default) | Audit · Mask · Block |
| Keys and secrets | Anthropic, OpenAI and OpenRouter keys, Mistral keys where they are named, AWS access keys, private keys | Audit · Mask · Block |
- Audit counts the finding and changes nothing.
- Mask replaces what it finds in the message before it leaves Kadmo: an e-mail address reads
<EMAIL>, a card<CREDIT_CARD>, a key<API_KEY>, an own pattern's match its name in capitals. Tool calls, attachments and the rest of the call are left as sent. - Block refuses the message with the rule's name, and nothing reaches the model (below).
The rules read what a chat sends to the model, never the model's answers. Your agents' jobs and a member's own client, such as Claude Code with a Kadmo key, send their prompts to the model as they are: content rules do not apply to them.
Each group opens to a switch per detector. The detectors check what they find: a card number passes the card networks' checksum, a US SSN and a Canadian SIN are written with their separators, so an order number is not taken for either.
Your own patterns. Add pattern opens a dialog with a Name, the Pattern and its Action; a pattern reads a chat's message, as the groups do. Try it matches the pattern against a text you type, in your browser, and sends nothing. An account keeps at most 25 patterns; Remove takes one off. A pattern is refused, with the reason, when it looks ahead or behind, refers back to a group, sets a flag such as (?i) (write both cases in a class instead, as [Kk][Dd]), uses a POSIX or Unicode class, repeats more than 1,000 times, repeats a group that holds a repeat or a choice of its own (such as (a+)+ or (a|ab)+; repeat the characters instead, as [ab]+), matches empty text, is longer than 256 characters, or is not a valid pattern.
Every change saves as you make it and a toast confirms it (Content rules saved, Pattern added, Pattern removed). A line under each group or pattern that masks or blocks says what it does. Both apply within seconds.
What the rules found. Each group and each of your patterns shows how many messages it matched this month, such as 12 this month or none this month, and a line under your patterns gives the month's total, removed patterns included (Nothing found this month when there is none). A message that matches two rules counts once for each. The counts carry no text. They follow the workspace you are in (· in checkout-service), and the whole account when you are in none.
What a blocked message meets. Nothing reaches the model, and the chat shows the rule's sentence: Blocked by your Guardrails: Keys and secrets. Remove it and send again. — the group's name, or the pattern's. The message is not charged. In Ask Kadmo a blocked message stays in the thread for you to read, and the next message is sent without it.
What the rules do not read. Attachments, images and the results of tools are not checked, and neither are the model's answers, your agents' jobs, a member's own client, or calls an account makes with its own key outside Kadmo. Voice is not checked. The model's earlier answers in a conversation are not read again as part of the message.
On a plan without content rules, the panel reads Business and up and shows the baseline, every action Audit.
Call rate
The ceilings in force, read-only: how many calls a minute each key may make, with its burst, and Kadmo's ceiling for every caller together. A call over a ceiling waits and retries.
Related
- Models and credit: credit, budgets, the model tiers and the LLM API's errors.
- Account and settings: Settings › Models and Billing.
- Workspaces: the scope a workspace's regions and budget limit.