Skip to main content

v0.3 — March 2026

Kimi K2 model upgrade
  • Upgraded the primary execution model to Kimi K2 via Moonshot AI
  • Significant improvements to cold email personalization, competitive analysis depth, and content quality
  • Model routing updated: Kimi K2 for complex reasoning, fast model for lightweight tasks
Compression engine v2
  • 3-layer cascade compression: MicroCompact → Session Memory Compact → API Digest
  • Sessions now run indefinitely without manual restarts
  • CompactBoundary mechanism prevents re-compressing already-compressed content
Brain Graph
  • Visual semantic memory map launched
  • Navigate, audit, and delete memory entries from a connected graph view
  • Edge strength reflects semantic similarity between memory nodes
50+ MCP servers
  • Expanded from 20 to 50+ pre-loaded MCP servers per sandbox session
  • Added: Sentry, PostHog, Webflow, Monday, Make, Cloudflare, Atlassian
  • Per-user dynamic loading: servers load based on your connected integrations

v0.2 — January 2026

Canvas system
  • Launched 30+ visual canvas templates across Sales, Research, Content, Ops, and Data categories
  • All canvases pre-research with 3 live Brave searches before generation
  • PDF and PNG export from every canvas
  • Inline rendering — no separate builder or page switch
Composio integration
  • 250+ integrations available via Composio gateway
  • Auto-routed in sandbox sessions based on connected services
Parallel execution
  • Increased to 75 concurrent task slots (5 agents × 15 workspaces)
  • Within-skill parallel tool execution via Promise.all
  • Semaphore control: max 2 concurrent skills per session
Skills expanded to 35+
  • Added: ICP Deep Match, Trigger Events, Pricing Research, Post-Call Blueprint
  • All skills now auto-save memory and auto-log activity after completion

v0.1 — November 2025

Initial release
  • Five specialized agents: Cortex, Specter, Striker, Pulse, Sentinel
  • 15 skills across Research, Lead Gen, Sales, Content, and Ops
  • Core chat interface with SSE streaming and tool activity feed
  • VIRALS scoring for content (0–12 scale)
  • Cold email 6-point quality gate
  • Memory system with 4 types and Sonnet-powered recall
  • Lessons system: errors stored and applied on next run