How It Works

From session reflection to eval-gated workspace knowledge in three acts.

Act 1

Dream

OpenCore's dream and core-delta skills already capture session reflections and delta-batches per project. The Q* plugin watches those outputs and uploads them to Q* cloud automatically — zero new workflow.

OpenCore-native
Consumes dreams and delta-batches your agents already produce
Automatic upload
Q* plugin watches docs/dreams/ and uploads seamlessly
No code captured
Only patterns and observations
PII stripped client-side
Identifiers removed before upload
Claude Code
Session ending...
── qstarlabs dream ──
Pattern: Async state mutations during React renders cause subtle hydration mismatches in SSR contexts.
Anti-pattern: Mutating shared state in useEffect during concurrent render passes.
✓ Uploading to Qstar Labs...
Act 2

Distill

Raw dreams pass through five stages to become eval-gated workspace knowledge.

01

Strip

PII and project specifics removed. Company names, URLs, internal tools replaced with abstract equivalents.

02

Normalize

Pattern rewritten into canonical form. Embedding generated for semantic clustering.

03

Cluster

HDBSCAN clustering finds natural groupings across workspace dreams. Stable fingerprints track clusters across reruns.

04

Eval Gate

Paired-replay on held-out projects. Only patterns that improve workspace-wide performance without regression pass.

05

Promote

Passing candidates become immutable workspace gradients with full lineage. Rollback available to workspace admins.

"One person's insight is an opinion.
Independent rediscovery by many is knowledge."

Act 3

Distribute

The loop closes. Workspace gradients flow back to every connected agent through three channels.

MCP

Your agent queries gradients in real-time during sessions. Native integration with Claude Code, Cursor, Copilot.

🔌

REST API

Integrate workspace knowledge into any tool, pipeline, or custom workflow.

📄

File Export

Pull curated gradients into your CLAUDE.md or AGENTS.md. Works with any tool today.

The loop closes. Every project teaches. Every agent learns.

See what gradients look like

Explore Features