SOTA · Terminal-Bench800+codedb

The terminal coding agent. Smarter every release.

graff is free, open, and SOTA on Terminal-Bench — powered by codedb. Bring your own API key, or use the codegraff gateway for pay-as-you-go access to Claude, DeepSeek, GPT-5.5, Grok, Kimi, and MiMo — no key management needed.

curl -fsSL https://codegraff.com/install-graff.sh | sh
See Pro pricingview source

macOS arm64 + Linux x86_64 · zero runtime deps

~ codegraffgraff8.4s
$graff -p "add error handling to handleAuth"
→ read src/auth.zig (full file · 1,240 lines · 2,103t)
→ read src/errors.zig (full file · 312 lines · 487t)
→ grep 'handleAuth' (5 matches · whole-file context)
→ edit src/auth.zig (+12 -3)
→ shell cargo test (ok · 0.7s)
// 5 tool calls · 18.4k tokens · stock file tools
task wall-clockstock

real graff session · same SOTA, fewer tokens with Pro

tokens saved · last 30d

ops · last 7d

per op (codedb p50)

muonry outline / codedb lookup

counters from activated installs · per-op & speed numbers from bench/ on github · hyperfine, n=10, macOS arm64

Engineering teams shipping with Codegraff

PlasmaBot MDtrilok.ai

Using Codegraff in your team? Send Rach a line - happy to credit your team next to a quote here.

Why it's SOTA · codedb

Retrieval that picks the right file.

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. On a 4-task SWE-bench Lite slice graded by a deterministic merged-PR oracle, codedb hit 4/4 recall, 3/4 top-1 with the smallest token footprint of six backends tested. Ships free with graff.

SWE-bench Lite writeup →justrach/codedb800+
Codegraff Pro/optional · SOTA per token

Same SOTA. Fewer tokens to get there.

Free graff is already SOTA on Terminal-Bench. Pro tools hit the same eval scores in fewer tokens per task — a faster search engine and lower-overhead file primitives, swapped in over the same install. Your agent stops re-reading whole files, finishes more in the same context window. Every release is benchmarked — see the changelog.

The problem

Eight engineers × $200/mo agents all reading whole files.

Default file tools shipped with Claude Code, Codex, Cursor, and Windsurf were built for humans, not agents. Every read pulls 2,000+ tokens. Every search returns lines without context. Multiply that across a team and you're burning weekly caps on plumbing - for every dev, every week.

Token usage - last 14 days

hover a bar

tokens consumed per coding session

What Pro adds

Four Pro tools. // muonry · zigrep · zigread · zigpatch.

Scope mode - context, not noise

Every match returns its enclosing function. Not a line. Not a file. The exact block your agent needs to act on.

zigrep -S 'validateInput' src/
ripgrep can't do this

Structural read - 47 tokens vs 2,103

outline gives the full symbol map of a file. Then symbol pulls exactly the function - never the whole file.

zigread outline src/main.zig
outline first

Symbol-safe patching - no line drift

Edit by function name, not line numbers. Survives upstream diffs, concurrent agents, and mid-session file growth.

zigpatch --symbol parseToken src/auth.zig
edit by name

Batch ops - 10 reads in 7ms

Persistent daemon. No fork/exec per call. Reads, searches, diffs, and edits run in parallel over one stdio pipe.

batch [read, search, diff, edit]
0.7ms per op

Benchmark - same query, two tools

zigrep5.2ms
ripgrep33.1ms

hyperfine · 4.7 GB repo · macOS arm64 · n=10

Install · 30 seconds

curl -fsSL https://codegraff.com/install.sh | sh
 activate CGPRO-XXXX-XXXX engineer@company.com
 mcp register --client claude-code

Runs locally — your code, prompts, and repo state never leave the machine. MCP-standard, no lock-in: uninstall and your agent falls back to its built-in file tools.

Built by Rach Pradhan · hi@codegraff.com

Codegraff Gateway/pay-as-you-go · no API keys to manage

One login. Six models. No keys.

Run graff provider login codegraff and you get instant access to Claude Opus 4.8, DeepSeek V4 Pro, GPT-5.5, Grok Build, Kimi K2.6, and MiMo V2.5 Pro. No API keys to manage, no provider accounts to create. Credits start at $10 — top up from the dashboard.

Claude Opus 4.8
$5/Mtok in
Claude Sonnet 4.6
$3/Mtok in
DeepSeek V4 Pro
$1.74/Mtok in
GPT-5.5
$5/Mtok in
Grok Build
$1/Mtok in
Kimi K2.6
$0.95/Mtok in
MiMo V2.5 Pro
$0.44/Mtok in
MiMo V2.5
$0.14/Mtok in

Live network usage →

Pricing

Per-seat math. // no contact-sales wall.

Individual

$99/year

1 seat · all tools · key by email

most teams

Team

$160/year

2 seats · all tools · keys by email

Team Plus

$349/year

5 seats · $69.80/seat · all tools

Stripe checkout link by email · same-day

Buy 5 seats →

Engineering org · 10+ seats

Volume pricing, invoiced annually, MSA on request. One-day turnaround on quotes.

Talk to Rach →

Refund within 14 days · License keys by email · No subscription auto-renew lock-in

Stop your team's agents from burning the cap.

One install per engineer. Works with Claude Code, Codex, Cursor, Windsurf. Per-seat licenses, by email.

See team pricing →Talk to Rach

Refund within 14 days · License keys by email · No subscription auto-renew lock-in

Questions? hi@codegraff.com · Terms · Privacy