Xagent v0.6.1
Seedance video generation, the ElevenLabs voice suite, and A2A agent protocol support all land
📅 Released July 12, 2026
🔄 How to update
cd xagent && git pull
docker compose up -d
pip install --upgrade xagent-ai
✅ Highlights
Seedance video generation
Generate video content directly inside an agent — no separate tool needed.
ElevenLabs voice suite
TTS, STT, sound effects, and music generation all land in one release.
A2A agent protocol
Adds Agent-to-Agent protocol support, extending multi-agent interoperability.
🚀 What's New
🎬 Seedance video generation
Adds Seedance video generation, letting agents generate video content directly.
🗺️ Key-based MCP catalog apps (Google Maps)
Supports connecting MCP catalog apps that require API-key auth (e.g. Google Maps), plus a new shared-env layer.
🖼 iframe embed auth parameter
Adds an embed_origin parameter for origin authentication in iframe-embedded scenarios.
🔊 ElevenLabs TTS
Adds ElevenLabs text-to-speech support, extending the voice capability ecosystem.
🔑 Multiple API keys per agent
Supports multiple API keys per agent with a centralized management page.
🛡️ Admin agent-config visibility & log rendering
Admins can view agent config details, with conversation-log rendering on the frontend.
📦 Sandbox management improvements (idle TTL)
Adds automatic sandbox reclamation on idle TTL, improving resource efficiency.
📈 Context-window-based compaction threshold
Adds an automatic compaction threshold based on context-window usage, with usage indicators.
🧭 Execution scope context
Adds execution-scope context, enabling scope-based sandbox isolation.
🗂 Decoupled sandbox mount root
Decouples the sandbox mount root from full workspace paths, supporting scope-prefix mounts.
🎙 ElevenLabs STT
Adds ElevenLabs speech-to-text support.
🔐 OAuth token resolver hook for MCP
Adds an OAuth token resolver hook, supporting custom auth-resolution logic for MCP configs.
📊 Optional quota enforcement hook
Adds an optional quota-enforcement hook interface for extending usage-limit logic.
🤝 A2A agent protocol support
Adds Agent-to-Agent (A2A) protocol support, extending multi-agent interoperability.
🧮 Enhanced model usage details
Usage details now include model_id and prompt-cache-hit token counts.
🛠 Tool allowlist & layered quota error codes
Adds a positive tool-allowlist hook mechanism and layered quota error codes.
🎵 ElevenLabs sound effects & music
Adds ElevenLabs sound-effect and music-generation capabilities.
🔁 Serialized task execution
Task execution control is now serialized, avoiding concurrency conflicts.
⏸ Per-step interrupt seam
Adds a per-step interrupt mechanism, supporting mid-run quota enforcement.
⚡ Trigger pipeline optimization
Optimizes trigger provisioning, auditing, and execution performance.
🗄 Checkpoint trace dedup
Deduplicates checkpoint trace storage and prunes history (20 GB → 2.2 GB on a real DB).
💬 AI Chatbot template execution mode change
Changes the execution mode of the AI Chatbot Agent template.
🔑 Unified catalog-app auth classification
Establishes a single source of truth for catalog-app auth classification.
📚 Slimmer KB compatibility adapters
Shrinks the knowledge-base compatibility adapter layer.
🛠 Stability & Polish
- Strengthen object-storage key validation
- Gate App Widget access behind a per-agent widget key
- Merge duplicate alembic migration heads left by an earlier fix
- Shrink KBCoordinator down to a pure router
- Decouple durable file previews from storage paths so they stay independently available
- Persist connector runtime snapshots for web tasks
- Scope model providers by category
- Enforce scope on by-id get / update / delete operations
- Scan due scheduled triggers in-process so they fire without Celery beat
- Confine file-id resolution to the active execution scope
- Scoped executions fail closed instead of falling back to local execution
- Fix data loss caused by LanceDB schema mismatches
- Fix MCP runtime translation-key resolution errors
- Bind durable-storage handles to the active execution scope
- Point the homepage docs link to the new hosted knowledge base
- Re-check containment when resolving relative workspace paths
- Make task message delivery interrupt-safe
- Push memory-dimension filters down into the LanceDB where clause
- Filter TTS voice metadata by configured models
- Preserve OAuth resolver actor diagnostics
- Deduplicate replayed user messages
- Add a PyPI badge and pip / curl install instructions to the README
- Remove a stray superpowers spec file and CHANGELOG.md