Krawl

Krawl is an autonomous deep research API that performs recursive breadth×depth tree search, synthesizes findings into a cited report with source verification, and streams progress via SSE events. It includes specialized crypto research tools, cross-session memory, document uploads, and scheduled research lookouts.

Built with FastAPI and litellm for multi-provider LLM routing. No agent framework — raw tool-calling with a 4-tier model strategy across AWS Bedrock and Anthropic.

Key Capabilities

• Recursive research tree — parallel queries at each depth level, learning extraction, gap detection, automatic recursion
• STORM perspective discovery — multi-perspective research inspired by Stanford's STORM methodology
• Multi-pass synthesis — gap check → revise cycles with quality scoring (Opus 4.7 for final synthesis)
• 4-tier model strategy — Haiku for high-volume tasks, Sonnet for research/planning, Opus for synthesis
• Cross-session memory — knowledge base that persists learnings across research sessions
• 9 research modes — deep, quick, crypto, token-analysis, protocol-research, whale-tracking, narrative, risk-assessment, yield-strategy
• Document upload — PDF, TXT, MD, CSV, JSON — up to 5 files, 10MB each
• 8 research templates — pre-built query patterns for common research tasks
• Research steering — inject instructions mid-stream to adjust research direction
• Session continuity — follow-up research that builds on prior context
• Source policy — domain allow/deny lists and freshness filters
• Structured output — JSON Schema-driven extraction alongside the markdown report
• Export — PDF, HTML, Markdown download
• Lookouts — scheduled recurring research with change detection and webhooks
• Audit trail — full log of every search, extraction, and synthesis step
• 25+ SSE event types — real-time streaming of every research phase

Architecture at a Glance

Client POST /research { query, mode, breadth, depth_levels }
  │
  ├─ Phase 1: Planning (Sonnet 4.6)
  │    └─ Structured research plan with topics and tasks
  │
  ├─ Phase 1.5: STORM Perspective Discovery (Haiku 4.5)
  │    └─ 3-5 research perspectives with key questions
  │
  ├─ Phase 2: Recursive Tree Search
  │    ├─ Generate N queries (Haiku 4.5) × depth levels
  │    ├─ Execute searches in parallel (Exa, X, GitHub, crypto, etc.)
  │    ├─ Extract learnings (Sonnet 4.6)
  │    ├─ Detect knowledge gaps (Sonnet 4.6)
  │    └─ Recurse with depth-1, breadth/2 if gaps remain
  │
  ├─ Phase 3: Verified Synthesis (Opus 4.7, 1 call)
  │    ├─ Gap check → optional re-synthesis
  │    └─ Multi-pass review → revise loop (up to 2 passes)
  │
  └─ SSE stream: plan → perspectives → queries → sources → learnings → report_chunk → result

Quick Start

curl -X POST https://api.krawl.sh/research \
  -H "Content-Type: application/json" \
  -H "X-API-Key: your-key" \
  -d '{"query": "What is the current state of AI agents?", "mode": "deep"}'

The response is an SSE stream. See SSE Events for all event types.

Pages

• Getting Started — Setup, env vars, first request
• Architecture — Deep dive into the research pipeline
• Models — 4-tier model strategy, provider routing, circuit breaker
• Research Modes — All 9 modes explained
• API Reference — All 22 endpoints
• SSE Events — All event types with schemas
• Crypto Tools — Crypto-specific tools and modes
• Templates — 8 pre-built research templates
• Memory — Cross-session knowledge base
• Export — PDF, HTML, Markdown export
• Lookouts — Scheduled recurring research
• Deployment — Fly.io production deployment
• Testing — Test suite guide