ReAct agent runtime
Multi-step reasoning with streaming tool calls over WebSocket, with full event replay on reconnect.
From a browser-based coding agent to webhook automation and multi-tenant governance — accurate to what ships today, with beta work clearly marked.
Multi-step reasoning with streaming tool calls over WebSocket, with full event replay on reconnect.
Agents request approval for risky tools, ask to clarify when unsure, and pause for your decision.
Background execution for signals and schedules with planning, working memory, and budget guardrails.
Clone a GitHub repo, browse files, and chat with an agent that proposes diffs you review and export as a PR.
Per-file and per-hunk accept/revert against a clone — focused verify runs before anything ships.
Symbol and call-graph indexing (CRS) grounds edits in callers, callees, imports, and related tests.
Upload PDFs, text, and markdown — chunked and embedded for per-agent semantic retrieval.
Index sites via sitemap or crawl, grouped as one source with live progress.
Agents distil knowledge cards from their own history and inject them into future runs.
Builtin + MCP + OpenAPI/Postman services, lazily loaded and authorized per agent.
Agents find the right tool via embedding search instead of a bloated prompt.
Connect stdio/HTTP/SSE MCP servers with credential injection and circuit breakers.
External systems fire agents with idempotency, HMAC auth, retries, and a dead-letter queue.
Run agents periodically with per-schedule budgets, max runs, and auto-pause on failure.
Compose multi-step automations on a node canvas with retries, approvals, and sub-workflows.
Organizations, workspaces, teams, and RBAC with workspace-scoped resources and audit trails.
Fernet-encrypted vaults and OAuth connections — secrets are never exposed to agents or logs.
Every LLM call is logged with tokens, latency, and frozen cost — per user, agent, and source.
Start free, bring your own model keys, and self-host in minutes.