Explore
Fast context gathering combining codebase search, documentation lookup, and web research.
Query Modes
Detect mode from query:
| Pattern | Mode | Primary Tools |
|---|---|---|
| "how does X work", code paths, files | Codebase | Grep, Glob, Read |
| library names, APIs, frameworks | Docs | Context7, Exa |
| current events, external services | Web | WebSearch, Exa |
| mixed/unclear | Hybrid | All tools |
Execution
Codebase Mode
code
1. Grep with broad keywords → find entry points 2. Glob for related files → patterns like **/*auth*.ts 3. Read key files → understand context 4. Follow imports → map dependencies
Optimized queries:
- •Function:
functionName\s*[=:(] - •Component:
(function|const)\s+ComponentName - •Import chain:
from ['"].*moduleName - •Type/interface:
(type|interface)\s+TypeName
Docs Mode
code
1. mcp__context7__resolve-library-id → get library ID 2. mcp__context7__query-docs → fetch relevant docs 3. Extract: setup, APIs, examples, patterns
If Context7 lacks info, use Exa:
- •
mcp__exa__get_code_context_exafor code examples (max 2-3 calls, 0.05$/call)
Web Mode
code
1. WebSearch with specific keywords 2. WebFetch on authoritative sources 3. Summarize key findings
Or use Exa:
- •
mcp__exa__web_search_exafor comprehensive search
Hybrid Mode
Run searches in parallel:
- •Grep/Glob for codebase matches
- •Context7 for library docs
- •WebSearch for external context
Output Format
CRITICAL: Output all findings directly. NEVER create files.
Findings
<codebase> **[File Path]** (lines X-Y) - Purpose: [what it does] - Key code: [relevant snippet or description] - Connects to: [related files/features] </codebase> <documentation> **[Library/API]** - Setup: [installation/config] - Key APIs: [functions/methods] - Example: ``` [code snippet] ``` </documentation> <web> **[Topic]** - Summary: [key information] - Source: [URL] </web>Connections
- •[How pieces fit together]
- •[Patterns discovered]
- •[Dependencies/relationships]
Gaps
- •[Missing information needing follow-up]
Rules
- •Speed first: Use parallel searches when possible
- •Cost aware: Context7 before Exa, limit Exa to 2-3 calls
- •Targeted: Extract only what's relevant to the query
- •No files: Output everything in response, never create markdown files
- •Source everything: Include file paths, URLs, library IDs
Query
$ARGUMENTS