Morph WarpGrep

Better, faster, cheaper search via subagents

A 12x faster code search subagent that searches your codebase in a separate context window. Up to 12 parallel tool calls across 3 turns—no embeddings required. Returns only relevant code—keeps your main agent clean.

Use cases

Code Search · Log Search · Package Search. Read the docs →

01
import { morph } from 'morph'
02
 
03
const result = await morph.warpGrep.execute({
04
  query: 'Find where billing invoices are emailed',
05
  repoRoot: './my-project',
06
})
07
 
08
// Clean results, no context pollution
09
console.log(result.contexts)  // Relevant code only
10
console.log(result.summary)   // Search findings
WarpGrep

Search codebases without blowing context

Isolated context window

WarpGrep is a specialized AI model that searches your codebase in a separate context window. Instead of your main agent running grep and accumulating context, WarpGrep does intelligent multi-step searching and returns clean results.

01
import { morph } from 'morph'
02
 
03
// WarpGrep searches in isolation
04
const result = await morph.warpGrep.execute({
05
  query: 'Find where billing invoices are generated',
06
  repoRoot: '.',
07
  excludes: ['node_modules', '.git'],
08
  includes: ['src/**/*.ts']
09
})
10
 
11
// Returns only relevant code sections
12
for (const ctx of result.contexts) {
13
  console.log(ctx.file, ctx.content)
14
}

Multi-turn reasoning

Your agent calls WarpGrep with a natural language query. WarpGrep reasons about what to search, runs up to 12 parallel tool calls (grep, read, list_dir) across 3 turns in under 6 seconds, and returns the relevant code sections. No embeddings required.

WarpGrep
Running 8 parallel searches…
├─grep -r "webhook.*handler"2.1k
Searching…
├─grep -E "async.*subscribe"1.8k
Searching…
├─read src/api/billing.ts:1423.2k
Reading…
├─grep "customer\.create"1.4k
Searching…
├─list_dir src/services/0.8k
Listing…
├─read src/lib/stripe.ts:1-802.6k
Reading…
├─grep -r "invoiceId"1.9k
Searching…
└─read src/api/checkout.ts:552.3k
Reading…
24 tool calls · 4 turns~5.8s

Best-in-class F1 score

WarpGrep achieves 0.73 F1 in just 3.8 steps—3x fewer than comparable agentic approaches. Deep logic queries show 2x-6x better performance.

1Warp Grep
0.733.8 steps
2SWE Grep
0.723.7 steps
3Claude Haiku
0.7212.4 steps
4Gemini Flash
0.6610.8 steps
5GLM 4.5
0.3414.5 steps
6mgrep
0.181 steps

WarpGrep achieves 0.73 F1 in 3.8 steps — 3x fewer than comparable agentic approaches

How WarpGrep works




Parallel breadth-first search

Parallel breadth-first search


Ship your first search workflow in minutes.

Try it now

$10/month in free compute