Changelog

Lightcone 0.1 Fast

Lightcone 0.1 Fast is live and is now the default model in the composer. Your model selection is preserved across reloads.

Run agents from Slack

Lightcone for Slack is rolling out to selected workspaces, so teams can ask agents to work directly from a Slack conversation and continue the same task with follow-up replies.

Chat improvements

• Assistant replies and clarifying questions now appear together in one message.
• Opening a past chat restores the full initial prompt, including long and multi-line instructions.

SDKs

• lightcone-python 2.44.0 and lightcone-typescript 0.19.0 include the latest computer API types.

Both releases are backward-compatible. No action required.

New Coding Agent setup

New Coding Agent setup examples now use the desktop environment by default. Regenerate an older setup prompt to use the current examples.

Performance and reliability

• Agents respond sooner after computer actions.
• Drag actions now register motion correctly on sliders, canvases, and other motion-tracking widgets.
• Sign-in redirects are smoother and the signup modal no longer flashes while your session is loading.
• The Tzafon news list on mobile shows the collapsed preview and reveals every post when expanded.

Run agents from Slack

Lightcone agents can now be run from a connected Slack workspace. Replies are plain text and easy to follow. Slack connection is rolling out to selected workspaces; if you do not see the option yet, access is expanding.

Faster starts from saved environments

Agents reach a working state sooner when resuming saved environments. No action required.

Improvements

Longer prompts in the hero chat input. The main chat field gives more room as you type, making multi-line prompts easier to draft and review.

Fixed

• Sign-in redirects are smoother. Landing-page sign-in lands you in the app directly. No action required.
• Team member removal is more reliable. Removing a member from your workspace now completes as expected. Workspace admins: no action required.
• Privacy policy and DPA links resolve correctly from the navigation and shared links.

Chat continuity across reloads and follow-ups

Conversations now survive reloads, new tabs, browser restarts, and follow-up turns more reliably. Returning to an older chat restores the desktop that chat was using, including logins, files, and state.

If a session needs to reconnect, the next message can continue with the prior conversation as context instead of starting from scratch.

Live token streaming

Assistant replies now stream in token by token instead of appearing all at once when a turn completes. Replies settle into a single final message at the end of the turn.

Inline file artifacts

When an agent produces an .xlsx, .pptx, or similar file, the chat now shows a download card directly in the conversation. The download remains available after the desktop session ends.

Lightcone 0.1 model selection persists

The model picker now remembers your selection across reloads and navigation. Choosing Lightcone 0.1 stays Lightcone 0.1 until you change it.

Fixes

• Follow-up turns on a desktop task render consistently with the first turn.
• Clarification questions appear once in the feed.
• Desktop sessions on Linux type ~, _, `, and ^ correctly.
• Chat replies appear once in the feed.
• Sign-in and chat search input behavior is consistent across the app.

Developer notes

Agent task API additions. POST /agent/tasks accepts new optional fields for SDK users:

• computer_id - attach a task to an existing live computer.
• on_missing_computer: "create_new" - restore from the saved session if the referenced computer is unavailable.
• context - seed a new task with recent conversation turns.
• thread_id - group multiple 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.

SSE streams are resumable via Last-Event-ID against GET /api/agent/stream/{id} and replay missed events on reconnect.

No action required for existing integrations. All new request fields are optional and additive.

Output token ceiling for dashboard chat completions

Chat now applies a server-side maximum for requested output tokens. This protects workspace credit balances from unusually large single requests.

• Affected: callers requesting more than 4096 output tokens.
• Not affected: callers already requesting max_tokens <= 4096. No action required.
• Default ceiling: 4096.

Lightcone models in the model picker

The dashboard model dropdown now Lightcone 01 alongside the existing Northstar options.

Saved environments continue in place

Saving an environment now keeps your session in place after the save completes, so you can continue working without a reload.

Fixed

• Missing environments return 404, making API error handling more predictable.
• Session start cleanup is more reliable when a new session cannot be created.

Take control pauses agents cleanly

Taking control during a Northstar session now pauses the agent quickly and resumes cleanly when you hand control back. Manual intervention is more predictable during active runs.

Planner-driven agent runs in limited access

A new agent runtime is rolling out to selected workspaces. It adds planner-driven runs with follow-up clarification questions for GUI tasks.

Canvas continues to use the existing runtime.

Builder Weekend promo

Builder Weekend pricing is available to eligible new billing contracts from May 29, 2026 at 5:00 PM PDT through June 1, 2026 at 12:00 AM PDT.

Eligible workspaces see the promotion in the dashboard during the promo window.

Auto top-up in limited rollout

Auto top-up is rolling out gradually. Workspaces that already enabled auto top-up continue to work normally.

Developers page

A new /developers landing page is live with a product overview, terminal demo, and showcase carousel.

Reliability & Misc.

• Long-running sessions stay active more reliably.
• Sign-out and token-refresh flows recover more cleanly.
• Token verification is faster for repeated invalid tokens.
• Lightcone has a refreshed logo

Lightcone for Slack

Lightcone for Slack is rolling out to selected workspaces, so teams can ask agents to work directly from Slack. Replies are concise and easy to follow.

New Coding Agent setup wizard

The dashboard home now has a guided New Coding Agent flow that produces a ready-to-paste setup prompt for Claude, Codex, Cursor, and other coding agents.

Longer and faster agent sessions

• Auto-mode runs can now use up to 2000 steps.
• Saved environments restore faster and appear sooner after sign-in.
• Allocation under bursty load is more reliable.

Sign-in

Sign-in and session refresh are smoother for dashboard and SDK users.

Developer and billing updates

• Access tokens issued by /api/token now include billing identity claims for clients that need them.
• lightcone.ai/install now opens the CLI installer instructions.

Misc.

• Updated visuals across the app, including the home page and Canvas.

Canvas

Canvas is live at /canvas, giving you a spatial way to start, arrange, and revisit agent chats as tiles.

• Click empty space to start a chat at that location.
• Pause and resume running agents from the tile menu.
• Keep zoom and Create controls reachable while inspecting a tile.
• Use the workspace menu to reach Developer, Docs, and Terms.
• Scan selected tiles and run status more easily.

Smoother chat interactions

• Switching between Chat and Canvas feels smoother in supported browsers.
• The textarea keeps focus after you send a message.
• User messages render immediatel∂y, and failed sends restore your draft.
• Enabling Desktop and tools mid-conversation preserves the existing chat.
• Follow-up messages can reconnect to the same session when possible.

Take control pauses the agent

Taking control now pauses the agent while you work and resumes it when you exit. Manual intervention is clearer, and the run state stays intact.

Account, billing, and legal updates

• The app menu now includes service policy entries alongside Developer, Docs, and Terms.
• Free daily usage for all users, capped.

Reliability

• Agents stay reachable more reliably during infrastructure changes.
• Conversation history from streaming Responses API calls is associated with the correct response.

More control during runs

Lightcone now warns before a mode change resets chat and run state. Stopping a run also keeps the conversation intact: chat history, artifacts, and session state remain available for follow-up work on the same computer.

You can draft the next message while an agent is still working. Send stays unavailable until the current turn finishes, but the input remains editable.

Better run output

Agent-created files now appear directly under the relevant turn, including artifacts from later turns in multi-turn sessions. Messages also have rerun and copy actions, and ended sessions offer a clear path to start a new chat.

Dashboard and Northstar polish

• Northstar layout, markdown rendering, and footer links are more consistent across screen sizes.
• Organization switching keeps the selected workspace stable.
• The home prompt sends correctly.
• API key actions stay visible across viewport widths.

Reliability

Agents start more reliably when multiple computers are created for the same environment at the same time.

Northstar medium model

tzafon.northstar-medium-alpha is now available in the Lightcone model picker. API and SDK clients can also discover it through GET /v1/models.

curl https://api.lightcone.ai/v1/models \
  -H "Authorization: Bearer $TZAFON_API_KEY"

Usage records group models by family

Usage and billing records now group model usage by stable model family while still retaining the exact model_id for auditability. Routine model version updates should be easier to read in usage records.

Reliability and compatibility

• Saved environments resume more consistently, including sessions that depend on browser auth, TLS, and screenshots.
• Sign-in and onboarding redirects are smoother across the dashboard, get-started, and workspace setup flows.
• The Claude Code button on the dashboard home works again.
• OpenAI-compatible tool calls are more reliable for conversational prompts, and streaming tool-call IDs use the standard call_... prefix expected by OpenAI SDK consumers.

Sign in to the Mac app with your workspace account

The Lightcone Mac app now signs in through your existing workspace session. If you are already signed in on the web, the app can finish setup and return you to the desktop app without manually copying an API key.

If your workspace still needs setup, the same flow takes you through setup first.

Developer onboarding in the dashboard

New developer accounts now get a guided dashboard setup flow that creates an API key and prepares the account for usage. The first-run path is shorter: sign up, complete setup, and leave with a working key.

Quick-start prompts start runs directly

Dashboard quick-start cards and the main dashboard prompt now start a run directly instead of only filling the input. Shared /run?prompt=... links keep their existing pre-fill behavior unless the link explicitly starts a run.

TypeScript SDK 0.14.0

TypeScript SDK 0.14.0 is available with the latest API surface. This update is backward-compatible for existing SDK users.

npm install [email protected]

Product polish and fixes

• The platform page, homepage model carousel, and refreshed navigation are live.
• The live computer view opens sooner after you submit a task, so agent work is visible earlier.
• Credits, dashboard links, run-page scrolling, streaming runs, and image inputs are more reliable.

TypeScript SDK 0.13.0

The TypeScript SDK is published at 0.13.0 with refreshed computers, computers.exec, computers.tabs, and responses resources.

npm install [email protected]

Compatibility: backward-compatible for existing SDK users.

Cleaner API reference

The OpenAPI and Swagger references now match the request and response shapes used by the v2 API. If you use the official SDKs, no action is required. If you generate your own client from the OpenAPI spec, regenerate it to pick up the corrected shapes.

Run view polish

Live agent feeds now stay focused on the active screencast when a run emits multiple start events, so the run view follows the session you actually need to watch.

Breaking: Coordinates are no longer pre-scaled

The API now returns coordinates in raw 0–999 model space. Previously, coordinates were scaled to pixel values matching your display_width and display_height before being returned. You must now denormalize in your code:

def _px(coord, dim):
    return int(coord / 1000 * dim)

x = _px(action.x, display_width)
y = _px(action.y, display_height)

This affects all computer_call actions — click, double-click, right-click, scroll, drag, and any coordinate-based input.

New model: tzafon.northstar-cua-fast-1.2

Available at the same pricing as northstar-cua-fast.

Pricing

tzafon.northstar-cua-fast and tzafon.northstar-cua-fast-1.2: $1.00 / 1M input tokens, $5.00 / 1M output tokens. Each request using the computer_use tool incurs approximately 1,974 additional input tokens of fixed overhead.

Multi-turn conversations

Chain requests using previous_response_id — pass the ID from any previous response and the model picks up exactly where it left off. No history management on your end.

Persistent responses

Every response is stored and retrievable by ID via GET /v1/responses/{id}. Pass store: false to opt out. Responses created with store: false cannot be referenced via previous_response_id.

Computer use

Full browser and desktop automation via the API. Use the computer_use_preview tool — the same interface as OpenAI.

OpenAI drop-in compatibility

Use the OpenAI Python or TypeScript SDK unchanged — point base_url at Lightcone. 37 Responses API fields supported.

VM snapshots are now compressed, stored on S3, and restored cross-node — sessions resume faster even when infrastructure moves.