Managing endpoints
Create and operate *.arkor.app endpoints: targets, auth modes, API keys, run retention, and lifecycle.
An endpoint is a dedicated URL of the form https://<slug>.arkor.app that serves OpenAI-compatible chat completions. Each endpoint belongs to a project and is managed from the project's Endpoints tab. The wire protocol is documented in the Chat Completions API reference.
Creating an endpoint
Endpoints → New endpoint asks for four things.


Subdomain slug
The slug becomes the endpoint's subdomain. Rules:
- 2 to 50 characters, lowercase letters, digits, and hyphens.
- No leading or trailing hyphen.
- Globally unique across all of Arkor. A taken slug returns a conflict error.
- Reserved words (
www,api,app,admin,dashboard,docs,status, and similar) are rejected.
Target
What the endpoint serves. Four choices:
- A job's final adapter. The LoRA adapter produced by a completed training job, served on top of the job's base model.
- A specific checkpoint. A mid-run checkpoint of a training job, selected by step.
- Base model only. A supported base model with no adapter.
- All public models (multi-model). Every public model in the Arkor catalog. Callers pick one per request with the OpenAI
modelfield; omitting it uses the default model,GET /v1/modelslists the current ids, and newly published models appear on the endpoint automatically. (One-click endpoints from the home page use this target.)
Arkor Cloud validates the target at creation time: the job must belong to the same project, the adapter artifact must exist, and the base model must be registered for inference.
Auth mode
- Fixed API key (recommended): callers must present a key via
Authorization: Bearerorx-api-key. - No auth (public): anyone with the URL can call the endpoint.
Run retention
Controls whether requests and responses are recorded for Run history:
- Keep for N days (default: 7). The dashboard accepts 1 to 365 days.
- Unlimited — transcripts are kept until you delete them.
- Don't store — nothing is recorded for new requests; runs stored before a retention change keep their original lifecycle (see Run history).
API keys
Endpoints in fixed-key mode can have up to 20 enabled keys, each with a label (up to 80 characters, unique among enabled keys).


- The plaintext key is shown once, at creation. Only a hash and a display prefix are stored.
- Each key can be revoked individually. Revocation takes effect within about 30 seconds (the edge caches deployment state briefly).
- The list shows each key's label and prefix.
Hand out separate keys per consumer so you can revoke one without rotating the others.
Editing an endpoint
From the endpoint detail page you can, at any time:
- Change the auth mode or run retention.
- Disable / enable the endpoint. Disabled endpoints stop serving but keep their slug and configuration.
The target itself has no dashboard control today: retargeting the same URL to another job, checkpoint, or base model is done through the SDK's deployments API, which updates the endpoint in place without clients noticing.
Deleting an endpoint
Delete removes the endpoint with its keys and stored run history. Serving stops within the same roughly 30-second window as key revocation (edge instances briefly cache deployment state). The slug is normally free to register again right away; only when a large stored-run history has to be drained by the background sweeper does it stay reserved until that cleanup finishes.
Anonymous endpoints
Endpoints created without an account behave differently: one-click endpoints from the home page expire after 7 days (see One-click endpoint), and anonymous identities are capped in how many live endpoints they can hold. Account-owned endpoints never expire.