RetrySight Cloud staging is live.

Open Cloud
RetrySight
Download Lite Start trial
Documentation menu
Docs home

Docs concepts

Architecture

Collector, manager, and dashboard — how RetrySight moves coding-agent telemetry in Lite, Cloud, and Enterprise.

On this page
  1. Collector
  2. Manager
  3. Dashboard
  4. Data path

Cloud and Enterprise share a collector → manager → dashboard pattern. Lite collapses collector and manager into one local process.

Product diagrams (system context, ingest flow, deployment, tiers) stay on the Product page. This page is the short operational model.

Collector

Sits on each developer machine and watches agent activity.

  • Lite: in-process collectors inside the desktop app
  • Cloud: laptop collector with bundled JRE; enroll from the dashboard
  • Enterprise: background fleet agent
  • Reads sessions from Cursor, Claude Code, Codex, Claude Desktop, and other registry sources
  • Queues events locally in SQLite if the manager link drops (~/.agent-retry/buffer.db on Cloud)
  • Sends retry and token data over authenticated REST (X-Agent-Api-Key)

Claude Cowork is not a log-file collector — it POSTs OTLP logs. Application traces are a separate OTLP spans path. See Cowork and Cloud traces.

Manager

Fleet brain — stores telemetry and serves the API.

  • Lite: embedded backend on 127.0.0.1:18081
  • Cloud: https://cloud-staging.retrysight.com (dashboard and API share this staging origin)
  • Enterprise: Docker/Helm stack on your network (Core dashboard 8080, API 28471)
  • GraphQL for the dashboard, including live subscriptions (taskUpdated, agentHeartbeat)

Cloud UI requests carry a JWT with account_id. Agent ingest looks up api_keys.org_id from the rk_live_* key. Requests without an account are rejected in strict tenancy mode.

Dashboard

  • Lite: Flutter desktop UI
  • Cloud: web console — Collectors, Live, Tasks, Cost, Traces (pages)
  • Enterprise: web console on your network
  • Fleet KPIs, task drill-down, live feed, and cost views (Cloud; Enterprise Analytics+)

Data path

  1. Coding agent writes local logs (or Cowork exports OTLP logs).
  2. Collector tails those logs and batches to the manager (default 10s on Cloud).
  3. Dashboard queries GraphQL. Optional ClickHouse dual-write on Enterprise Integration+.
  4. Optional app traces POST to /api/v1/otel/v1/traces bound by Task ID.

Nothing in this path requires an SDK in your application. See data residency.