compact

Compact

Versioned, procedural docs for 70+ APIs and libraries. Less context, better code.

usecompact.dev · Source on GitHub

Compact is a knowledge cache of actionable procedures — installation steps, configuration patterns, code snippets — extracted from official docs and from developer field trials. Agents and IDEs query it before generating integration code so they get current, versioned answers instead of outdated training data.

Compact also covers two other domains via feed_type:

• llm-card — structured LLM model cards: pricing, context windows, benchmarks, capabilities, privacy tier. Query it before recommending a model ("best coding model under $5 per 1M output tokens", "which model has a 1M context window and vision").
• design-md — curated design-system and UI style references for visual/brand direction.

See SEARCH-API.md § Content types for the full list and query examples.

---

Why Compact?

AI coding assistants hallucinate integration code. They confidently generate API calls for library versions they were never trained on, skip required config steps, and miss the one flag that makes everything work. You find out at runtime.

Compact fixes the retrieval problem, not the model:

• Fewer tokens, lower cost — one targeted query returns the exact procedure instead of feeding entire docs or repository files into the context window
• Less hallucination — versioned, sourced procedures give the model ground truth to work from instead of interpolating from training data
• Local and open-source models punch above their weight — a model with no web access and a smaller context window can produce correct integration code when Compact hands it the right steps up front; no browsing, no RAG pipeline to build
• The community compounds the value — every field trial contribution adds an edge case, an undocumented behavior, or a version-specific fix that benefits every agent querying Compact; the more developers use and contribute, the sharper the cache gets

One MCP tool call before the code gets written. No new model, no fine-tuning, no prompt engineering.

---

Query the API

Endpoint: https://gate.usecompact.dev/search
Method: POST
Auth: Authorization: Bearer <your-api-key>
Body: { "query": "how do I connect Prisma to Postgres", "limit": 3, "git_repo": "prisma/prisma" } — default 3; broader queries often use 5–7.

curl -s -X POST https://gate.usecompact.dev/search \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ORACLE_API_KEY" \
  -d '{"query": "add Clerk auth to Next.js", "limit": 3}'

Try it now: shared beta key (no signup) in BETA.md. Personal keys: usecompact.dev. Beta terms: BETA.md. API contract: SEARCH-API.md.

---

Use in your IDE (MCP)

The Compact MCP server exposes oracle_search (query the cache) so Cursor, Claude Code, Gemini, and other MCP clients can use Compact before writing integration code. Cache misses are tracked internally — agents do not get a write tool (oracle_log was removed; untrusted agents must not write to the learning loop).

From this repo:

cd mcp-server
npm install
npm run build

Then point your IDE’s MCP config at mcp-server/dist/index.js and set ORACLE_API_KEY (shared key in BETA.md). See SETUP.md for all clients (Cursor, Claude, Gemini, VS Code) and CLIENTS.md.

From npm (when published):

npx oracle-knowledge-mcp

---

Docs in this repo

VS Code: Opening this repo in VS Code uses the included .vscode/mcp.json so the Compact MCP is ready after npm run build in mcp-server/ (requires Copilot agent mode).

---

Contribute

Every developer has debugged something that took hours and wasn’t in any docs. That knowledge belongs in Compact.

• Email: [email protected]
• Twitter: tag @usecompact with #compactfieldtrial

See CONTRIBUTION-SYSTEM.md for what to send and how we review.