Human-in-the-loop
Review and approve tool calls before execution. Per-agent HITL rules in YAML. Reject with reason, and the agent adapts.
Agent Orchestration
Stream responses, validate tool calls, and manage your fleet. Built for teams who treat AI agents as production infrastructure.
Every feature exists because real agent deployments need it. No filler, no toy examples.
Review and approve tool calls before execution. Per-agent HITL rules in YAML. Reject with reason, and the agent adapts.
SSE-powered streaming with typed events: thinking, content, structured responses. See the agent's reasoning as it happens.
Define agents declaratively: model, tools, middleware, MCP servers, subagents, HITL rules. Hot-reload without restart.
Extend agent capabilities with Model Context Protocol servers. Filesystem, git, browser, custom tools, all via stdio or SSE.
Browse MinIO folders, index with LightRAG or classical chunking, query across workspaces. Built-in file content extraction.
Agents can delegate to subagents with scoped tools and prompts. Complex workflows without monolithic prompts.
How it works
No black box. You see what the agent is doing, approve what matters, and ship the result.
Pick an agent from the sidebar, type your request. The agent starts working immediately. You see its thinking, tool selections, and partial responses as they stream in. No waiting for a final answer to know if it's on track.
When the agent wants to run a tool that could have consequences (a security scan, a file delete, a network call), it pauses. You get the full arguments, an approve or reject button, and a reason field. Nothing executes without your say-so.
The agent returns a structured response you can copy, share, or feed to the next step. Audit reports, code reviews, data analyses, all delivered as JSON alongside the natural-language summary. Your pipeline picks it up from there.
When an agent responds, you know exactly what it did, what it cost, and who approved it.
Every message, tool call, and approval is timestamped and stored. Replay any conversation step by step. See what the agent saw, what it decided, and why.
Per-message token counts and estimated spend, per agent and per thread. Know which agent burns your budget before the invoice arrives.
Who approved what, when, with what reason. Exportable as JSON for compliance. HITL decisions are first-class records, not log afterthoughts.
Streaming responses, thinking blocks, and structured outputs visible in real-time. You watch the agent work, with no black-box latency between prompt and result.
Launch the console and start managing your agent fleet.
Launch Console