PhaseDrift

Your evidence and costs.

See what stays local, which results sync to your account, and what can cost money.

The core tracker and code scans run on your machine. Optional connections have their own data paths and costs. Inspect the path before enabling it.

FeatureData goes toCost
Code scans and activity importYour local trackerLocal CPU and disk
Task reportYour local trackerWriting uses the active agent's existing tokens
Account history, when enabledPhaseDrift cloud: metric results and structured learningNo model call
Managed Jev analysis, with consentPhaseDrift gateway, then TypeSafeIncluded subscription allowance
BYO Jev analysis, with consentTypeSafe directlyYour provider billing
GitHub enrichmentLocal records read through your GitHub accessYour existing GitHub access

The website does not receive your project data. Model analysis is separate from browsing or saving a report.

What syncs to your account

Account history stores measured results: code-health totals, daily agent-work counters, qualified feedback timings and clean-repeat outcomes. It also stores structured improvements, trial definitions and outcomes, lessons, and learning observations. Missing or incomplete measurements stay labelled as such.

Source code, file and function inventories, diagnostic artifacts, transcripts, tool-event logs and original task reports stay in the local tracker. The cloud receives typed result rows and compact upload receipts, without keeping a raw upload archive.

Your setup agent enables account history after the computer is connected and the results selection is reviewed. Future results then sync automatically while collection runs. History sync is separate from model analysis and backups. You can pause it in Settings.

Keep private evidence private

Reports store the feedback you submit. Analyzer artifacts can include code fragments, diagnostics, and local paths. These local records are not automatically scrubbed public exports.

This website receives no tracker evidence or source code. Your coding host has its own data handling and billing.

The core runs locally

Your project tracker runs on loopback and stores its data in a local SQLite database. Code scans use your machine's CPU and disk. They don't call a model.

This website is documentation. It doesn't connect to your tracker, receive your project evidence, or upload your source code.

What gets stored

Native activity import keeps a restricted projection for recorded sessions, turns, usage, and tool activity. OpenCode, Grok Build, Claude Code, Gemini CLI and Cursor telemetry also retains only normalized metadata and counters. It doesn't import raw prompts, private reasoning, command arguments, or full transcript output into that projection.

Agent reports store the feedback you submit. Analyzer artifacts can contain diagnostic messages, code fragments, and local paths. Keep reports concise and free of secrets or customer information.

The database and analyzer artifacts are private local evidence, not automatically scrubbed public exports.

Costs you can expect
ActivityWhere the cost comes from
Browsing your local dashboardLocal application; no model call
Structural and full scansLocal CPU, disk, and possible dependency downloads
Agent work in Codex, Muse, or another hostYour existing host or provider billing
Managed report classificationIncluded subscription allowance
BYO report classificationYour TypeSafe account
Running a personal benchmark taskThe agent host you choose to execute it

Some scans take longer and use more storage than others. Sampled structural history is bounded; full historical analysis can require additional dependency environments.

Optional model analysis

Managed Jev analysis groups report outcomes, friction, causes and intervention types. Your coding agent supplies concrete suggestions. Connect a subscription, inspect the redacted payload, then accept separate cloud-analysis consent. Connecting the license alone does not authorize uploads.

Managed analysis sends minimized report prose through PhaseDrift's gateway to TypeSafe. BYO sends the same minimized prose directly to TypeSafe using your own key. Neither path adds code files, raw transcripts, screenshots or diagnostic traces. Reports themselves can contain private details, and pattern-based redaction is not guaranteed anonymization.

The managed gateway stores classifications for 30 days, usage counters, opaque fingerprints and account/device identifiers. Billing uses encrypted subscription credentials and short-lived encrypted recovery details. Report prose is not persisted at the gateway. Your saved local classifications remain available when analysis is off.

TypeSafe says requests are not used for training. Standard access does not promise zero retention. Review TypeSafe's data handling for its current terms.

Managed access includes 10,000 attempts and 10 million input tokens per UTC month across your projects and machines, with a 1,000-attempt daily safety limit. BYO uses your local policy and provider billing. Commercial builds still require a product license for either choice. Read plans and licensing before checkout.

Local tracking, reports and proposals work without model analysis. Setup · Limits and troubleshooting.

Optional GitHub enrichment

GitHub collection uses your local gh authentication to read PRs, reviews, findings, and checks. It doesn't comment on or merge PRs.

Retain and recover your evidence

Managed upgrades make a database backup before migration. Removing a project registration retains its data. Settings provides storage and retention controls; inspect those controls before pruning anything.

Use phasedrift stop to stop a manager-owned tracker. Never remove a lock belonging to a running instance.