When we hit, ~91% less context than a full doc read.
Compact is a versioned knowledge cache for AI coding agents — library procedures, LLM model cards (pricing, benchmarks, capabilities), and design-system style references. When the cache has the right answer, one query returns ~800 tokens instead of 8k–14k. Less context, better code.
How it works
Agent queries Compact
Via MCP (oracle_search) or the gate API. Natural-language query — optionally scoped with feed_type (oracle, llm-card, design-md) or git_repo.
Cache returns structured knowledge
~800 tokens. Procedures with code snippets for libraries; model cards with pricing and benchmarks; design references with style direction.
~91% fewer tokens when we hit
Compared to a single full doc read (8k–14k). Same search API for integration docs, LLM selection, and UI direction. Less context, better code.
Three caches. One search API.
Compact is not just library docs anymore. Pass feed_type to scope a query, or omit it to search everything. The MCP server oracle_search tool exposes all three domains.
Library & API docs
Versioned install steps, config patterns, code snippets, and common errors for 70+ libraries and services.
- “connect Prisma to Postgres”
- “add Clerk auth to Next.js App Router”
- “Supabase RLS in a server action”
LLM model cards
Structured pricing, context windows, benchmarks, capabilities, and task-fit for frontier models — updated weekly.
- “best coding model under $5 per 1M output”
- “anthropic/claude-sonnet-5 pricing and context”
- “compare Grok 4.20 vs GPT-5.5 for agents”
Design-system styles
Curated UI and brand direction from awesome-design-md — typography, color, layout patterns for product surfaces.
- “minimal SaaS dashboard style”
- “dark-mode developer tool aesthetic”
- “editorial landing page typography”
950+ records. Ready when your agent asks.
Library docs, LLM model cards, and design-system references — all searchable through the same MCP tool and gate API.
Libraries & APIs
LLM models
Design styles
Join the beta
Request access — we'll send your API key via email.
Every developer has debugged something that took hours and wasn't in any docs. That knowledge belongs in Compact so the next person doesn't lose the same time.
Submit a field trial — describe what you hit, what the fix was, and any relevant code. If it's a real sharp edge we'll add it to the cache within 48 hours.