Authentication
Every /v1/* request needs a Devotel API key in the Authorization header:
Authorization: Bearer dvt-...Key format
dvt- followed by 40 lowercase hex characters, for example
dvt-91159742a1b2c3d4e5f60718293a4b5c6d7e8f90. Malformed keys are rejected before any lookup, so a
typo comes back as invalid_api_key rather than a slow failure.
How keys are stored
Only sha256(key) and the first 8 characters are stored. The dashboard shows you dvt-9115… for
identification; the full value exists exactly once, in the dialog that created it.
Lost a key? Revoke it and create a new one — it cannot be recovered or resent.
Managing keys
At chat.devotel.com/keys you can create keys, see each one’s prefix, creation date, last use and rate limits, and revoke any of them. Revocation takes effect immediately — the gateway drops its cached copy of the key as part of revoking it.
A revoked key returns 403 with code key_revoked, which is distinct from 401
invalid_api_key. That difference is deliberate: 403 means “this key existed and no longer
works”, so your retry logic can stop rather than assume a transient failure.
Model access
By default a key may use every generally available model — you do not have to choose, and an existing key’s access does not change when new models are added.
Creating a key offers two modes:
- Specific models (the default) — search the catalogue and add models as chips, starting with Kimi K3 already chosen. The list is exhaustive: once a key has one, anything not on it is refused, including models added later. Clear every chip to fall back to every generally available model.
- Cost-optimized (Auto) — every unrestricted model is included as a removable chip; the key is
restricted to
autoplus whatever remains in that pool.autoroutes each request to the cheapest model that can handle it, escalates to a stronger one if the first attempt fails, stays on the same model for the rest of a conversation, and bills whichever model actually answered — see Auto Routing for the routing rules. Removing every tool-calling or vision-capable model from the pool does not block key creation, but a request that needs that capability will then be refused; the key creation page warns you inline before you save.
Editing an existing key’s model list reuses the same picker, and a key created before this UI existed still renders correctly — an unrestricted key opens on Specific models with no chips selected, exactly its original meaning.
A few models are restricted and are not part of the general catalogue. They are granted per key by an administrator, and until then they behave as if they do not exist.
Asking for a model your key cannot use returns the same 404 model_not_found as asking for a
model that does not exist, and GET /v1/models lists only what your key can reach. This is
deliberate — it means the error cannot be used to discover which models exist — so if a model
you expected to work returns 404, check your key’s access before assuming the id is wrong.
IP allowlists
Under “Security” on the key creation and edit forms, you can optionally restrict a key to a list of source IPs — one IPv4/IPv6 address or CIDR range per line, for example:
203.0.113.10
10.0.0.0/24
2001:db8::/32Leave it empty and the key works from any source IP, which is the default and the meaning every
key created before this feature already has. When a list is set, a request from any other address
gets 401 with code ip_not_allowed, and the response never repeats your IP back to you. If a
key you expect to work keeps failing this way, the keys page shows the source IP and time of the
most recent rejection, so a typo’d range is easy to spot.
The allowlist is checked against the address our edge network sees the request come from, not
against any X-Forwarded-For header you send — that header can be set to anything by the caller,
so it is never trusted for this check.
Keeping keys safe
- Keys carry full access to your account’s API usage and are billed to you. Treat them like passwords.
- Never put a key in browser or mobile code. Anything shipped to a user’s device is readable by that user. Call the API from your own server and keep the key in its environment.
- Use one key per application or environment. Then revoking a leaked key affects one thing.
- Rotate by creating the new key, deploying it, and revoking the old one — in that order.
Errors
| Status | code | Meaning |
|---|---|---|
| 401 | missing_api_key | No Authorization header, or not Bearer |
| 401 | invalid_api_key | Malformed key, or no such key |
| 401 | ip_not_allowed | The key has an IP allowlist and this request’s source IP is not on it |
| 403 | key_revoked | The key exists but has been revoked |
Full reference: Rate Limits & Errors.
Reasoning effort defaults
A key can carry a default reasoning_effort, set on the
API keys page or by an administrator. It applies to every request
made with that key that does not specify one itself.
The precedence is, in order:
reasoning_effortin the request — always wins, including"none".- The key’s default — used when the request omits the field.
- Nothing — if neither is set, the field is left off entirely and the model applies its own default.
A key default is a convenience, not a ceiling: any request can override it in either direction. Use it to make the cheap choice the automatic one for an agent or a coding assistant, then raise it per request where it pays for itself. See Chat Completions for what the values do.