cursive-blink

The customer Python SDK and blink command-line tool for Blink. Upload supported checkpoints, request model deployments, submit supported evaluation campaigns, follow their progress, and download published results. You need a Blink service URL and a customer team API key supplied by your Blink operator.

This package does not run a Blink server, schedule GPUs, execute evaluation sandboxes, perform training or include model weights. Installation does not grant access to a service, model or compute. Available features depend on the service's configured capabilities and your team's permissions.

Install

Linux and Python 3.12 or later are required. The initial release targets Linux x86-64 and ARM64; macOS and Windows are not qualified. No CUDA, Torch or training environment is required for the ordinary client.

For the CLI, install an isolated tool environment:

uv tool install --python 3.12 'cursive-blink==0.2.0'
blink --help

If your shell cannot find the command, run uv tool update-shell and start a new shell. For Python integration, install into your project's environment instead:

uv add 'cursive-blink==0.2.0'

uv tool install does not expose the SDK to an unrelated Python environment. For offline clusters, ask your operator for the matching wheel and a dependency wheelhouse for the correct Python version and architecture.

Configure and check access

Replace the example URL with the service URL supplied by your operator. This command prompts for the customer API key without echoing it:

blink configure --base-url https://blink.example.com --output client.json
blink check --config client.json

For noninteractive setup, add --key-file /private/blink/team-key.txt; the file must be owned by you and have 0600 permissions. Never put the key in a command-line argument, a checked-in script or a chat message. Configuration and transfer journals contain credentials and must remain private. The client checks the key with the configured service before saving a new configuration.

configure authenticates an existing key; it does not sign you up, create a team or bootstrap operator credentials. It will not overwrite an existing config. check performs read-only authentication and capability checks and reports separately whether campaign discovery is available. Run either command with --help for all options.

Common commands

In these examples, client.json is the private configuration produced by setup. Replace uppercase placeholders with IDs/settings returned by your service.

blink capabilities --config client.json
blink upload --config client.json /absolute/checkpoint --request-key MY_UPLOAD_KEY
blink resume --config client.json job_RETURNED_JOB_ID
blink campaigns --config client.json recipes
blink campaigns --config client.json submit RECIPE_ID \
  --recipe-digest RECIPE_DIGEST --request-key MY_CAMPAIGN_KEY --model-id MODEL_ID
blink campaigns --config client.json watch cam_CAMPAIGN_ID
blink campaigns --config client.json artifacts cam_CAMPAIGN_ID
blink campaigns --config client.json download cam_CAMPAIGN_ID art_ARTIFACT_ID \
  --output /private/results/result.json

Uploaded checkpoint preparation/model registration follows the workflow exposed by your selected deployment profile; an upload receipt alone is not a ready model. blink deploy, connect, heartbeat and stop support customer-owned harnesses. Use blink COMMAND --help for their required options.

Keep each request key stable when retrying the same submission. Uploads run in a detached local worker and maintain durable resume journals: that machine still needs to stay running until transfer completes. Exiting campaign watch does not cancel the campaign. Use explicit cancellation, and distinguish accepted cancellation from confirmed resource cleanup. Missing results or measurements are not successful results or zero usage.

Supported task archives can be uploaded with campaigns upload-task and submitted with campaigns submit-tasks only when the service enables a qualified runtime. Upload acceptance is not execution approval. Team key replacement/revocation and explicit retention previews are available through the same authenticated client; revoking a management key does not cancel existing deployments or inference keys.

Python SDK

from cursive_blink import BlinkClient

client = BlinkClient.from_config("/private/blink/client.json")
print(client.capabilities())
print(client.campaigns())

The SDK supports checkpoint upload/resume, model/deployment management, campaign submission/status/cancellation, scoped connection credentials and checksummed artifact download. Optional cursive_blink.harness.ManagedHarness manages a deployment lease and cleanup around your own harness; it does not supply or run the harness. Its completion adapter currently returns non-streaming responses.

Optional agent tools

uv tool install --python 3.12 'cursive-blink[mcp]==0.2.0'
blink mcp --config /private/blink/client.json

This starts a local stdio MCP server using the same customer credentials. MCP exposes its own supported deployment-management tools; it is not a promise of CLI/campaign feature parity. Run it where checkpoint files are readable. Core CLI/SDK use does not need the MCP dependency.

Compatibility and licence

Use the release recommended by your operator. Server-only updates do not require automatic client upgrades; new client features still depend on server support. This public package uses the cursive_blink Python namespace. Earlier private monolithic builds used blink; migrate example imports accordingly. Do not upgrade an existing Blink server environment using this customer distribution.

The included customer client is MIT licensed. This licence does not cover Blink's private backend, third-party dependencies, customer data, credentials, model weights or service access. Dependencies retain their own licences.