Codegraff: 40× leaner file tools for your coding agent
SOTA · Terminal-Bench800+codedb<br>Your agent re-reads the whole file. You pay for every line.<br>CodeDB Pro cuts that token bill ~97%.<br>Free graff is already SOTA on Terminal-Bench. CodeDB Pro keeps the exact same eval scores and gets there in a fraction of the tokens. Your agents just stop dumping whole files into context.<br>curl -fsSL https://codegraff.com/install.sh | bashcopy<br>Cut your token bill · $99/yrsee the math →<br>macOS arm64 + Linux x86_64 · zero runtime deps
~ codegraffstock file tools<br>tokens this task<br>18,400<br>5 tool calls · 8.4s wall-clock · Claude / Codex native
baseline
$graff -p "add error handling to handleAuth"<br>read src/auth.zig 1,240 lines 2,103 tok<br>read src/errors.zig 312 lines 487 tok<br>grep "handleAuth" whole-file context<br>edit src/auth.zig rewrites 169 KB<br>shell cargo test ok · 0.7s
same task, same result · hover to pause
Engineering teams shipping with Codegraff<br>trilok.ai
2.5B<br>tokens saved · last 30d
145k<br>ops · last 7d
63µs<br>per op · codedb p50
47t / 14t<br>outline / lookup
counters from activated installs · speed numbers from bench/ on github · hyperfine, n=10, macOS arm64
The token ledger<br>The same task. A fraction of the tokens.<br>One real file, measured three ways. Default tools dump the raw text. codedb, the free engine, reads it structurally. CodeDB Pro lands the exact symbol and edits in kilobytes. The red bar is what you pay for today.
Understand a 2,300-line file<br>29× leaner than the raw dump
865 tok<br>with CodeDB Pro
codedb 6,305 tok25,324 toktoday
Search the whole codebase<br>codedb’s search engine ships inside Pro
built in<br>with CodeDB Pro
codedb 2,160 tok5,938 toktoday
Edit one function<br>Pro only · codedb reads, it doesn’t edit
1.9 KB<br>with CodeDB Pro
169 KBtoday
measured · codedb-pro 0.2.9 · same 2,318-line file · "built in" = Pro ships codedb's search engine
How CodeDB Pro works<br>It finds the right file. Then it barely reads it.<br>01codedb picks the right file<br>Most agent failures aren't bad reasoning, they're the agent looking at the wrong file. codedb is graff's retrieval layer: a local index of symbols, outlines, and call paths the agent queries before it acts. SOTA on Terminal-Bench, free, open-source.
02CodeDB Pro gets there in ~97% fewer tokens<br>Same eval scores, a fraction of the footprint. outline returns 47 tokens where a raw read pulls 2,103. symbol lookups land the exact function, never the whole file.
03Then Pro edits, symbol-safe<br>The one thing the free engine can't do. Symbol-safe patches write 1.9 KB where default tools rewrite 169 KB. Immune to line drift, concurrent agents, and mid-session file growth.
SWE-bench Lite writeup: 4/4 recall, 3/4 top-1 →justrach/codedb →
What CodeDB Pro adds<br>Compact reads. Symbol edits. Batched ops.<br>Scope mode context, not noise<br>ripgrep can't do this<br>Every match returns its enclosing function. Not a line. Not a file. The exact block your agent needs to act on.<br>zigrep -S 'validateInput' src/<br>Structural read 47 tokens vs 2,103<br>outline first<br>outline gives the full symbol map of a file. Then symbol pulls exactly the function, never the whole file.<br>zigread outline src/main.zig<br>Symbol-safe patching no line drift<br>edit by name<br>Edit by function name, not line numbers. Survives upstream diffs, concurrent agents, and mid-session file growth.<br>zigpatch --symbol parseToken src/auth.zig<br>Batch ops 10 reads in 7ms<br>0.7ms per op<br>Persistent daemon. No fork/exec per call. Reads, searches, diffs, and edits run in parallel over one stdio pipe.<br>codedb-pro batch [read, search, diff, edit]
Benchmark · same query<br>6.4× faster
zigrep5.2ms
ripgrep33.1ms
hyperfine · 4.7 GB repo · macOS arm64 · n=10
Pricing<br>Per-seat math. No contact-sales wall.<br>$99/year is one seat. The ledger above is one real file: 25,324 tokens of raw dump down to 865, edits in kilobytes not hundreds. Every week you wait, that gap is tokens you already pay for, spent re-reading files.<br>14-day refund · no auto-renew lock-in · key by email, instantly<br>Individual<br>$99/year<br>1 seat · $99/seat · all tools, all editors<br>Get Pro · 1 seat<br>most teamsTeam<br>$160/year<br>2 seats · $80/seat · save $38 vs 2 solo seats<br>Get Pro · 2 seats<br>Team Plus<br>$349/year<br>5 seats · $69.80/seat · save $146 vs 5 solo seats<br>Stripe checkout link by email · same-day<br>Get Pro · 5 seats →
Engineering org · 10+ seats<br>Volume pricing, invoiced annually, MSA on request. One-day turnaround on quotes.
Talk to Rach →<br>Refund within 14 days · License keys by email · No subscription auto-renew lock-in
Codegraff Gateway/pay-as-you-go · no API keys to manage<br>One login. Twelve models. No keys.<br>Run graff provider login codegraff and you get instant access to the frontier set. No API keys to manage, no provider accounts to create. Credits start at $10, top up from the dashboard.<br>Claude Opus 4.8<br>$5/Mtok in<br>Claude Sonnet 4.6<br>$3/Mtok in<br>DeepSeek V4 Pro<br>$0.52/Mtok in<br>GPT-5.5<br>$5/Mtok in<br>Grok Build<br>$1/Mtok in<br>Kimi K2.7 Code<br>$1.14/Mtok in<br>Kimi K2.7 Code Highspeed<br>$2.28/Mtok in<br>GLM-5.2<br>$1.68/Mtok...