Skip to main content
A guardrail is a reusable usage policy. It sets:
  • Model access: which providers, LLMs and Expert Models can be called.
  • Member budget: how much each member can spend per day, week or month.
  • Rate limit: how many requests each member can make per second, minute, hour or day.
Create a guardrail once, then attach it to a role, a member or an API key.
Guardrails are part of the advanced organisation features. See plans and pricing for availability.

​
Create a guardrail

1

Open Guardrails

In the Eden AI dashboard, go to Settings, select Guardrails, then click Create Guardrail.
2

Basics

Give the guardrail a Name and, optionally, a Description. Under Assignment, pick the roles, members and API keys it applies to. You can also assign it later from Members or Roles.
3

Budget & rate limit

Set a member budget and a rate limit, or leave them empty for no limit.
4

Model access

Choose which providers and models can be called. See Model access.
5

Save

Click Create guardrail. You can change any part of it later from the guardrail’s page.

​
Model access

The Model access section: the Only allow mode, a Providers card, a Models card and the Current eligibility previewThe Model access section: the Only allow mode, a Providers card, a Models card and the Current eligibility preview
Pick a restriction mode, then list providers and models:
  • Providers: a provider covers everything it serves, its LLMs and its Expert Models. Block OpenAI here and no OpenAI model can be called.
  • Models: click Add models, then switch between LLMs and Expert models.
    • LLMs lists the models behind every OpenAI-compatible endpoint: chat, embeddings, images, speech, transcriptions and moderations. Use the Region filter to allow a model in one region only, for example the EU.
    • Expert models are grouped by task. Tick a provider to cover all its models for that task, or tick single models.
The Add models picker on Expert models, grouped by taskThe Add models picker on Expert models, grouped by task
Current eligibility, under the lists, shows what the guardrail allows today, by provider and by model, before you save.

​
What a request is checked against

  • LLM choices apply to LLM calls: chat completions, Responses, Messages, and the OpenAI-compatible /v3/embeddings, /v3/images, /v3/audio and /v3/moderations endpoints.
  • Expert model choices apply to Expert Model calls on /v3/universal-ai, and to the web_search hook on LLM calls, which searches through a web/search provider such as Linkup. A call that names no model is checked as the provider’s default model.
  • A request the guardrail doesn’t allow gets HTTP 403. The error names the refused model’s id, such as openai/gpt-5.5 or audio/tts/google/chirp3-hd.

​
Custom rules

A guardrail can also hold ordered rules that mix allow and block. The editor shows them under Custom rules and keeps them as they are. Each rule is a pattern on the model id: the first rule that matches decides, and the default applies when none does. A * also matches /: */gpt-4o matches openai/gpt-4o and any Expert Model id ending in /gpt-4o.

​
Member budget

The Budget and rate limit section: a monthly member budget with an alert at 80% and a rate limit of 60 requests per minuteThe Budget and rate limit section: a monthly member budget with an alert at 80% and a rate limit of 60 requests per minute
Under Budget & rate limit, set an Amount in USD and a period: daily, weekly or monthly. Turn on Budget alert to email the member, with a copy to the organization owner, when their spend reaches a share of the budget.
  • Per member, not shared. Each member the guardrail applies to gets their own budget, across all their API keys. It is not a shared pool for the team.
  • Resets each period. Spend resets at the start of each day, week or month (UTC).
  • Soft cap. The check happens before each request, so a request already running can take spend past the cap. The next request is the one that gets refused.
  • Real usage only. Sandbox calls and calls made with your own provider keys (BYOK) don’t count. Refunds credit it back.
The budget applies only when the guardrail is attached to a role or a member. On a guardrail attached to an API key it is ignored: use a token budget to cap a single key.

​
Rate limit

Set the number of Requests and a window: second, minute, hour or day. Requests over the limit are refused until the window passes. The limit applies to each member and replaces your plan’s rate limit, which it can’t exceed. Leave it empty to inherit the limit from the next level up.

​
Assigning a guardrail

A guardrail does nothing until it is attached. Attach it under Assignment on the guardrail’s page, or:
  • A role: from Settings → Roles & Permissions, so everyone with that role inherits the policy.
  • A member: from Settings → Members, to apply it to one person.
  • An API key: to constrain a single key, for example a shared or CI key.
Each role, member and API key holds one guardrail. When a request comes in, Eden AI resolves each limit on its own, the most specific level first:
Model access and the rate limit each come from the most specific level that sets them, so a key can override the model access and still inherit its role’s rate limit. The member budget skips the API key level: it comes from the member or the role, amount and period together, never mixed across levels.
Guardrails are configured in the dashboard; there is no API for them yet. To automate the per-key side of provisioning (budgets, expiry, attribution to a member), use the Management API.