1. Capture
Keep turns, finished tools and evidence with their source.
Your workspace
Overview
Loading your captured activity…
Your first steps
Progress comes from received records. A completed step does not mean every source is connected or healthy.
A little clarity goes a long way
Your AI, a little clearer.
The work, at a glance
Received records, not an all-clear. Open to learn more.All received historySDK operations
Loading SDK monitoring...
The latest chapter
Latest 100 captured turns, grouped by UTC date.
Source: captured turn records. Missing dates do not establish inactivity.
Worth a closer look
View all →Rule matches are findings to review, not proven failures.
Who is doing the work?
Manage →Received history does not prove a source is online.
Projects
All received turnsProject labels fall back to captured working directories.
Captured turns
Follow the work, wherever it runs
A connection you can check.
See what arrived, what is waiting, and which computer needs a closer look.
Are your collectors reporting?
Recent means a report reached this service within two minutes. Counts describe the last report; they do not prove every AI action was captured. Refresh to fetch new reports.
Received activity by agent
These agent cards show received history. Use the collector reports above for the last reported local queue and delivery checks.
Capture where the work happens
Link the computer running Claude Code or Codex. A collector on that machine can record work across projects and retry delivery after a connection failure.
When you dispatch from your phone, capture must run on the machine executing the agent. Signing into a subscription does not expose its browser or mobile conversation history.
SDK event journal
Latest 100 received SDK events. Local queued events appear after delivery.
Choose a project to see its captured work across your agents. Projects come from the source's project label or working directory.
Your attention, your rules
A good signal helps you notice.
Start with our built-in checks. Add your own for the way you work.
Decide what deserves attention
A signal is a rule that points you to a moment worth reviewing. A match is a clue, not a verdict.
Advanced: where rules run
Built-in rules run in the collector before text is removed. Custom rules run here when a turn arrives, using only uploaded, scrubbed information. Phrase rules cannot inspect text that stayed on your computer.
Saving a rule does not backfill history. A re-sent turn is evaluated with the current custom rules. Pausing keeps earlier findings and saved decisions until a turn is re-sent. Saved definitions remain available for interpreting evidence; create a new signal to change a rule.
Private examples, with your permission.
Collect selected session or SDK metrics and your explicit signal reviews automatically. Prompt text, tool contents and model training are not included in this format.
Your dataset versions
Each version is a fixed snapshot. Examples need further qualification and separate evaluation groups before they may be used for training.
Learning is being built. Saved signal verdicts and configured metadata dataset collection form the foundation. Training, clustering and promotion are not active.
2. Review
Save whether a finding matches what happened against an exact signal and evidence revision.
3. Collect examples
Enable private metadata datasets where the service supports them. Qualification and training remain separate.
Open datasets4. Evaluate & improve
Test candidates on held-out evidence, promote reversibly, then monitor the result.
Your saved decisions
Review signals →Latest 1,000 reviews. A review does not start training or change a rule.
Appearance
Saved on this browser.
Your privacy
Content is opt-in and scrubbed by the collector before upload. Metrics do not require prompt or response text.
Manage collection permission in Datasets. Redaction editing and plan management are not available here yet. Pricing and subscription limits are not inferred from token counts.
Your cloud allowance
Loading your allowance...
The research behind RROFLOW, including all eight Raindrop documents. Historical observations and proposals are not proof that a feature is live.