New features, improvements, and fixes shipped to PixTwins.
Migrated from our custom orchestrator to the Claude Agent SDK, enabling Claude Code-level autonomous workflows for all users. Agents now have true agentic loops with tool use, error recovery, and multi-step planning built in.
Always-on agents can now define pre-filters on their triggers so they only process events that match specific criteria. This reduces noise and saves tokens by skipping irrelevant webhook payloads before they reach the agent.
Fixed a bug where OAuth callbacks would redirect to the wrong origin when behind a reverse proxy. The redirect URL now correctly uses forwarded headers instead of the internal nextUrl.origin.
All OAuth tokens stored in the database are now encrypted at rest using AES-256-GCM. This adds a critical layer of security so that even if the database is compromised, integration tokens remain protected.
Code execution results now stream directly into the chat interface. Both stdout and stderr from Docker sandbox runs are displayed inline so you can see exactly what your scripts produce without leaving the conversation.
Agents can now modify their own automation configurations through a new tool call. This means you can ask an agent to change its schedule, update trigger filters, or adjust its behavior without visiting the settings page.
Workspaces now support multiple independent chat threads. Create task chats for one-off work or always-on chats for continuous monitoring. Each chat maintains its own conversation history while sharing workspace-level agents and memories.
When a conversation exceeds approximately 150K tokens, the system automatically summarizes earlier context using Sonnet. This keeps conversations functional indefinitely without losing important details from earlier in the thread.
Introducing 24/7 monitoring agents that stay active and process real-time events from your connected apps. Set up a Slack monitor, Gmail watcher, or analytics alerter that works around the clock without manual prompting.
Long-running agent tasks that hit the 30-minute timeout now automatically continue where they left off. Sonnet generates a continuation summary, and the frontend fires the next request seamlessly so complex tasks run to completion.
Agents can now write and execute Python and Node.js code in isolated Docker containers on the VPS. Each execution runs in a fresh container with no network access, memory limits, and a 2-minute timeout for safety.
A multi-agent workflow for end-to-end video creation. From research and brief writing through storyboarding, asset generation, and final editing -- your agent team collaborates to produce complete video content.
Each agent now has its own set of allowed Composio toolkits defined in the composio_toolkits field. This means your CRO agent only sees analytics tools while your email agent only sees Gmail -- reducing confusion and improving accuracy.
Use your own Claude, Gemini, or OpenAI API keys for unlimited usage beyond your plan's token allocation. Keys are stored encrypted and used transparently -- you just pay your plan fee plus your own API costs with zero markup.
A persistent memory system that lets your agents remember brand guidelines, preferences, and past decisions. Memories are stored with vector embeddings and BM25 indexing for fast semantic retrieval during agent execution.