Skip to content
Try out in chatDeveloper dashboardLogin
Guides

Changelog

API and SDK changes, breaking changes, and our deprecation policy.

API-, SDK-, and documentation-relevant changes, newest first. Breaking changes and deprecations are announced here before they take effect. For the full product changelog (dashboard, chat, Canvas, Slack, and platform updates), see tzafon.ai/changelog.

Run Northstar locally on a Mac. Northstar CUA Fast 1.6 is now available as an MLX 4-bit quantization (Hugging Face, Apache 2.0), pairing with the new MacControlKit Swift package (Apache 2.0) for on-device screen control. Together they let you build a computer-use agent with no API key, no network, and no cloud computer. See Run Northstar locally on a Mac.

Continuous learning (preview). Deploy a model that learns from your corrections. Because Lightcone is your inference endpoint, corrective turns in your traffic are captured automatically and folded into a nightly, eval-gated checkpoint by self-distillation, metered as compute units (forward 1 CU, backward 2 CU per 1k tokens). Behind a per-org feature flag; see Continuous learning. New: tzafon deploy, client.corrections.create, client.learning.status, client.learning.checkpoints (+ rollback), and the X-Lightcone-Training header.

Documentation overhaul.

Price drop (June 26). northstar-cua-fast input dropped to $0.50/M (was $1.00) and output to $1.50/M (was $5.00), a 50–70% reduction. See Pricing.

Tasks API additions (week of June 8). All new fields are optional and additive; no action required for existing integrations:

  • computer_id: attach a task to an existing live computer; on_missing_computer: "restore" | "create_new" | "fail" controls what happens if it’s gone.
  • context: seed a new task with recent conversation turns; thread_id: group tasks belonging to the same chat.
  • start_url, save_session, keep_alive, max_duration_seconds, metadata, idempotency_key: explicit task controls, plus task cancellation and deadline events.
  • Computer responses now include environment_id when a live computer was created from a saved environment.
  • Task SSE streams are resumable: reconnect with the Last-Event-ID header and missed events replay. See Tasks.

Chat Completions output-token ceiling (week of June 1). A server-side maximum of 4096 output tokens now applies. Callers already requesting max_tokens <= 4096 are unaffected.

Computer lifetime fields renamed. timeout_seconds and auto_kill on computer creation are deprecated in favor of max_lifetime_seconds and idle_timeout_enabled. The deprecated fields still work; requests using them receive Deprecation and Sunset response headers indicating when support ends.

# Before
computer = client.computer.create(kind="browser", timeout_seconds=3600, auto_kill=False)
# After
computer = client.computer.create(kind="browser", max_lifetime_seconds=3600, idle_timeout_enabled=False)
// Before
const computer = await client.computers.create({ kind: "browser", timeout_seconds: 3600, auto_kill: false });
// After
const computer = await client.computers.create({ kind: "browser", max_lifetime_seconds: 3600, idle_timeout_enabled: false });

April 27. tzafon.northstar-medium-alpha is available in the model picker and via GET /v1/models. Usage and billing records now group usage by stable model family while retaining the exact model_id for auditability.

April 20. TypeScript SDK 0.14.0 published, backward-compatible.

April 13. TypeScript SDK 0.13.0 with refreshed computers, computers.exec, computers.tabs, and responses resources. The OpenAPI/Swagger reference now matches the v2 API shapes; if you generate your own client from the spec, regenerate it.

April 8. Breaking: coordinates are no longer pre-scaled (Python SDK 2.35.0, TypeScript SDK 0.8). The API now returns coordinates in raw 0–999 model space. Previously, coordinates were scaled to pixels matching your display_width/display_height before being returned. Denormalize in your code:

def to_px(coord, dim):
return int(coord / 1000 * dim)
x = to_px(action.x, display_width)
y = to_px(action.y, display_height)

This affects all computer_call actions: click, double-click, right-click, scroll, drag, and any coordinate-based input. See Coordinates and the updated computer-use loop.

Also April 8: new model tzafon.northstar-cua-fast-1.2 at the same pricing as northstar-cua-fast.

March 12. Responses API.

  • Multi-turn conversations via previous_response_id, with no history management on your end.
  • Persistent responses: every response is stored and retrievable via GET /v1/responses/{id}; pass store: false to opt out (responses created with store: false can’t be referenced by previous_response_id).
  • Computer use via the API, with OpenAI drop-in compatibility: use the OpenAI SDK unchanged with a Lightcone base_url.

March 9. Faster session resume. Saved-environment snapshots restore faster, including across infrastructure moves.

  • Deprecated fields and endpoints keep working through the sunset date announced in the Deprecation and Sunset response headers, and are listed on this page.
  • Pin your model versions (tzafon.northstar-cua-fast-1.6, not an unversioned alias) and SDK versions in production; upgrade deliberately rather than tracking latest.
  • APIs marked beta can change with notice via this page.