Try for free
Changelog

What shipped, when.

A running list of what we’ve built, grouped by the part of the product it touches.

Ultron Shell

  • Brought the shell worker in-repo with live streaming and made it the native sandbox backend and the Computer tab, retiring the previous external one.
  • Pod-truth /work: the workspace never comes up empty and never silently re-clones, with a cold-start "resuming workspace" status and a gate backstop.
  • Native file search in the sandbox (shell_glob, shell_grep) and a general monitor tool that streams any command's output live into chat.
  • Hard VERIFY, COMMIT and DEPLOY gates, and execution now requires a selected repo and branch.

Tool calling & capability discovery

  • A deferred tool registry with tool_search, later merged into discover_capability: one door, one registry, one gate.
  • discover_capability wired into chat with 35 integration actions; engineering shell tools registered as a first-class tool kind.
  • Trimmed roughly 34 rare tools out of the always-loaded core set to keep the prompt lean.
  • todo_write plus enter and exit plan mode for in-chat planning.

Chat engine & model routing

  • One model registry as the single source of truth for every tier (Minuet, Allegro, Forte, Plinth), plus a new Plinth tier (free, 262k context) and a fast no-tools lane.
  • A unified Completion Gate replacing scattered stop-guards, with per-turn cost accounting and a consecutive-failure breaker.
  • delegate_subagents routed by tier, a floating text-selection picker (Copy, Markdown, Attach as context), and cleaner tool and reasoning rendering.
  • Shared chats now carry the full tool tree and interleaved stream.

Browser & computer use

  • A stateful browser_session worker wired into chat, with login and context persistence and a live Computer-tab view to watch Ultron browse, including screencast and take-over.
  • Anti-bot resilience: stealth input, detect-and-degrade on bot walls, a residential-proxy fallback, and browse-as-country routing.
  • Structured extraction with PDF support, a vision fallback for actions, network-response capture, research and authenticated-crawl composites, and session recording.

Memory & Brain

  • Hybrid recall (vector plus full-text) and embedding-based memory recall on shell turns, with a nightly consolidation cron.
  • Brain wired end to end: Memories, Files (in-app preview, locked bucket), Agreements, and a full Notes and Pages system with editor, database views, templates, trash and revisions.
  • Projects group your conversations so threads, deliverables and memories share one workspace.

Lead sources & signals

  • Lead-scraping actors that stream straight into your CRM: Google Maps brick-and-mortar, LinkedIn job-listings hiring signals, and tracked-profile job and title change detection.
  • Company enrichment across LinkedIn and web sources feeding the prospecting and outreach pipeline.

Skills & the engineering unit

  • A shell-proof skills and techniques router with a deterministic turn-mode gate and tier-aware routing, plus library_search and a research skill.
  • A lifecycle hook system, HITL pause and resume with a wait_for_approval tool and an Execute or Reject pill, and 12 engineering skills bundled into the sandbox template.

Credentials, vault & security

  • A credential vault with encryption and login injection, auto-vaulting of secrets pasted into chat, named API-key secrets, and 13 bring-your-own-key validators.
  • Credential injection at the egress edge (the secret-exfil fix), per-app limits, and removal of the global key from the container.

App platform & deploys

  • deploy_wfp ships built /work apps live, with an executable backend, static assets and server-side secrets.
  • Provisioning tools for a database, object storage, a key-value store and an AI gateway, per-user resource ownership, an outbound egress guard, and app lifecycle (list and delete).

Background jobs & missions

  • Generic background-job inspect, control and create tools with every kind wired, plus missions with auto-attribution and notifications.
  • Warm in-turn await for background jobs and a surfaced kill reason on dead jobs.

Milestones & credits

  • Milestones: all-user tracking, an activity event stream, and credit-grant observability.
  • A rebuilt credit system (1 USD = 1000 credits, tier-aware deduction, a daily free allowance), a unified wallet and a per-model rate catalog, plus a referrals invite loop.
  • Native in-app checkout with an inline subscription manager and invoice downloads.

Calls & Sales Copilot

  • A meeting-bot wired end to end, and a Sales Copilot with a live transcript, stage detection, a moment rail, pre-call brief and post-call review, with transcripts materialized into Brain.

Carousels, images & video

  • A self-contained carousel render service, carousel-extract (image to editable template), inline image generation, and iterative single-slide editing.
  • A reference library of brand design systems and infographics wired into the planner, so generated slides match a real visual language.
  • Drop-a-video-in-chat ingest end to end with video insights, plus lightweight image description.

Voice, connectors & MCP

  • Voice dictation (mic to transcription to composer) with voice-activity auto-stop.
  • A 42-app connector registry, custom user-added remote MCP servers, a one-tap key sheet, and the Ultron MCP server with an external bridge.

GitHub & import

  • Per-conversation repo and branch lock, an app-token broker in the shell, and git-backed /work persistence.
  • Import of existing ChatGPT and Claude chat history.