Agent initialized. Upload files and type a task to begin.
TOOLS▼
WEB SEARCHno key▼
TODOS▼
SUB-AGENTS▼
HOOKS▼
CONTEXT▼
0
Messages
0
Tokens
Install Skill
Installing...
Drop .skill, .zip, .json, or SKILL.md here
Add Custom MCP Tool
Most MCP servers don't send CORS headers for browsers. If a direct connection fails with a CORS error, set a proxy URL here — it's prepended to the target URL (use {url} for query-param style). A 20-line Cloudflare Worker is enough.
Or add a single server manually
None
None
Settings
No per-model limits set. Click "Import Selected" or add a row below.
No per-model reasoning settings. Auto keeps DeepSeek and MiMo on, others off.
When remote sandbox is active, only Bash is exposed and it runs inside a Daytona Linux container. PythonExec / JSExec / NodeExec / Pyodide tools are hidden — invoke python3, node, pip, etc. via Bash. Read/Write/Edit target Daytona first; /foo.js and /home/daytona/foo.js refer to the same sandbox file. Only files under /home/daytona/outputs are pulled back to VFS /outputs after Bash. Put final results in /home/daytona/outputs.
Uses browser notifications for completed or failed chat and media runs. Click a notification to reopen that conversation.
Unattended mode stops when the assistant outputs the completion marker. It will not approve Plan Mode or destructive human prompts.
Built-in roles: researcher, critic, writer, coder (all read-only). The lead agent can issue several SwarmSpawn calls in one turn — they run in parallel up to the concurrency cap. Workers can also chain via SwarmHandoff (researcher → critic → writer; coder → critic → writer). Cycles and depth-overflow are rejected. Workers cannot ask the user, use MCP, or spawn more workers. Plan Mode and Ralph Loop are honored as guards.
Agent-created roles appear in the Swarms sidebar with a blue [agent] badge. You can delete them like any custom role; built-in and user-authored roles are protected from RoleManager.
Triggered tasks consume LLM tokens automatically. Tasks run only while this browser tab is open; closed tabs miss fires.
5-field local-time cron: minute hour day-of-month month day-of-week. Each fire opens a fresh conversation. Ralph Loop has priority — fires that overlap a Ralph run are skipped and rolled forward.
Memory = Information + Time Label + Searchable Relationships. Records live in IndexedDB and sync via the cloud-sync layer below when configured.
Delete all local settings, conversations, memories, skills, files, sync config, and cached data on this device.
Settings (including S3 credentials and encryption passphrase) are stored locally in your browser (localStorage). Keys never leave your machine except when sent directly to the configured API endpoints or S3 bucket.
Review Plan
Approve turns Plan Mode off and lets the agent execute write tools. Reject keeps Plan Mode on and returns your feedback so the agent can revise.
Input Requested
Add Scheduled Task
Add Hook
Hooks run with the same power as JSExec. They can read and modify agent state. Errors in a hook are swallowed and logged — they never break the agent loop.
Add Custom Role
Custom roles are visible to the lead agent only when Agent Swarm is enabled in Settings. Plan Mode and the global allowWriteRoles switch still apply on top of any tool you grant here.
MemoryInformation + Time + Relationships
Memories carry a time label and tags. They are injected into the system prompt at the start of each turn ranked by recency, tag overlap, and keyword match. Pinned memories are never auto-pruned.