Skip to content

feat: add a Nebius Token Factory example provider profile #4856

Open

Description

@amrrs

User Story

I use OpenShell to run coding and research agents in sandboxes, and my models are served by Nebius Token Factory through its OpenAI-compatible API at https://api.tokenfactory.nebius.com/v1. Today I have to hand-write a provider profile for it, and the obvious copy of the openai profile with the host swapped would hand my agents a key that can also create dedicated endpoints and fine-tuning jobs on my account.

Problem Statement

providers/ has no example profile for Nebius Token Factory. Operators who use it have to write one from scratch and decide on their own which parts of the Token Factory API a sandboxed agent should reach.

Impact / Why This Matters

A Token Factory API key works on more than inference. On the same host it can list, upload, and delete files and datasets, start fine-tuning jobs and batch operations, and create, update, or delete dedicated endpoints, which are billable GPU deployments. A profile that grants the whole host, like the existing openai, nvidia, and deepinfra examples, lets a sandboxed agent use the injected key for all of that. Operators who don't read the API reference closely will not notice.

Proposed Design

Add providers/nebius-token-factory.yaml, following the format of the other example profiles:

id: nebius-token-factory
display_name: Nebius Token Factory
category: inference
inference_capable: true
credentials:
  - name: api_key
    env_vars: [NEBIUS_API_KEY]
    auth_style: bearer
    header_name: authorization
endpoints:
  - host: api.tokenfactory.nebius.com
    port: 443
    protocol: rest
    enforcement: enforce
    rules:
      - allow: { method: POST, path: "/v1/chat/completions" }
      - allow: { method: POST, path: "/v1/completions" }
      - allow: { method: POST, path: "/v1/embeddings" }
      - allow: { method: POST, path: "/v1/rerank" }
      - allow: { method: POST, path: "/v1/responses" }
      - allow: { method: GET, path: "/v1/models" }
binaries: [/usr/bin/curl, /usr/local/bin/curl]

Workflow:

openshell provider profile import -f providers/nebius-token-factory.yaml --global
openshell provider create --name nebius --type nebius-token-factory --credential NEBIUS_API_KEY
openshell sandbox create --provider nebius -- curl -sS https://api.tokenfactory.nebius.com/v1/models

NEBIUS_API_KEY matches the variable name in the Token Factory documentation.

Acceptance Criteria

  • providers/nebius-token-factory.yaml lints and imports, and has the same header block as the other examples (client binaries, reference layout, credential scope, endpoint access, smoke test).
  • A sandbox with the provider attached can call chat completions (streaming included), completions, embeddings, rerank, the Responses API, and the model list.
  • In that sandbox, requests to the files, datasets, fine-tuning, operations, and dedicated-endpoint APIs are denied by the proxy.
  • No telemetry or provider-table changes are needed, in line with Remove the inference provider table and provider profile telemetry buckets #3906.

Alternatives Considered

  • Leave it to operators to copy the openai profile and change the host. That gives agents the whole Token Factory API, as described above.
  • Grant /v1/** read-write. This is simpler, but it still allows file, dataset, fine-tuning, and operation calls, and it would allow new management routes added under /v1 later.
  • An endpointless profile bound per sandbox from policy. That suits per-deployment hosts, but Token Factory has a single public host, so a fixed endpoint is simpler for operators.

Agent Investigation

  • Profiles are import-only examples under providers/. Adding one changes only the YAML and the expected-ID list in list_provider_profiles_returns_built_in_profile_categories, as in feat(providers): add Oracle Cloud Infrastructure Generative AI provider profile #3904.
  • I listed the routes from Token Factory's published OpenAPI spec (https://api.tokenfactory.nebius.com/openapi.json). Inference routes are the ones above. Management routes are under /v1/files, /v1/datasets, /v1/fine_tuning, /v1/operations, and /v0/dedicated_endpoints.

Activity

  1. github-actions commented on Nov 20, 2026

    @github-actions

    This issue has had no activity for 28 days and is now marked stale. It may be closed in 14 days if there is no further activity. Comment or remove the state:stale label to keep it open.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    state:staleInactive item at risk of automatic closure.state:triage-neededOpened without agent diagnostics and needs triage

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions