ROI & Benchmarks
Executive intro, savings math, and benchmark data for Coregit vs GitHub.
Executive intro
Coregit is a serverless Git hosting platform built for AI agents and automated workflows. One API call commits any number of files — GitHub needs 13+ calls for the same operation. This translates directly to faster agent execution and lower compute costs.
The key numbers
Commit 100 files: Coregit 19.8s vs GitHub 72.1s — 3.6x faster. One API call vs one hundred five.
Write throughput: 15,000 commits/hour vs GitHub's 38. That's 394x more capacity. GitHub's 500 writes/hour rate limit means an AI agent hits the wall after 38 ten-file commits per hour. Coregit handles 15,000 — your agent never waits.
Benchmark summary
Measured April 2026, Kazakhstan → US-East, private repos with authentication.
| Operation | GitHub | Coregit | Faster | API Calls |
|---|---|---|---|---|
| Commit 1 file | 2,217 ms | 2,148 ms | ~Parity | 4 vs 1 |
| Commit 5 files | 4,829 ms | 3,456 ms | Coregit 1.4x | 8 vs 1 |
| Commit 10 files | 8,387 ms | 4,183 ms | Coregit 2.0x | 13 vs 1 |
| Commit 100 files | 72,064 ms | 19,769 ms | Coregit 3.6x | 105 vs 1 |
| Read file (warm) | 735 ms | 800 ms | GitHub 1.1x | 1 vs 1 |
| List tree | 797 ms | 752 ms | Coregit 1.1x | 1 vs 1 |
| List commits (warm) | 829 ms | 474 ms | Coregit 1.7x | 1 vs 1 |
| Diff branches (warm) | 738 ms | 752 ms | ~Parity | 1 vs 1 |
Full benchmark details: Scalability Benchmarks
Savings model
Single-call commits eliminate HTTP chatter
GitHub needs N+3 sequential API calls for an N-file commit (N blob creates + get HEAD + create tree + create commit + update ref). Coregit does it in 1 call.
| Workload (10-file commits/day) | GitHub API calls/day | Coregit API calls/day | Time saved/day |
|---|---|---|---|
| 100 | 1,300 | 100 | ~7 minutes |
| 1,000 | 13,000 | 1,000 | ~70 minutes |
| 10,000 | 130,000 | 10,000 | ~12 hours |
| 50,000 | 650,000 | 50,000 | ~58 hours |
Agent compute savings
AI coding agents (Cursor, Claude Code, Codex) spend compute time waiting for git operations. Faster commits = less billable agent time.
| Agent workload | GitHub wait time | Coregit wait time | Compute saved/month | $ saved (@$1.50/h) |
|---|---|---|---|---|
| 5K commits/day (10 files each) | 11.7 h/day | 5.8 h/day | 177 hours | $266 |
| 15K commits/day | 35 h/day | 17.5 h/day | 525 hours | $788 |
| 50K commits/day | 117 h/day | 58.3 h/day | 1,761 hours | $2,642 |
Seatless usage pricing
Usage tier is pay-as-you-go with no seat charges. Automation-heavy orgs with dozens of service accounts pay only for actual git operations.
| Org footprint | GitHub Team | Coregit usage | Monthly savings |
|---|---|---|---|
| 40 repos, 20 humans + 20 agents | 40 seats × $4 = $160 | ~$65 | ~$95 |
| 100 repos, 8 engineers + 52 bots | 60 seats × $4 = $240 | ~$65 | ~$175 |
Zero egress
Coregit runs on Cloudflare R2 with zero egress costs. A team mirroring 3 TB/month of clones pays ~$270 on AWS S3 egress but $0 on Coregit.
Rate limits
GitHub: 5,000 requests/hour, 500 writes/hour (content creation). At 13 calls per 10-file commit, you hit the write limit after just 38 commits/hour.
Coregit: 15,000 requests/hour per key, 50,000/hour per org. At 1 call per commit, that's 15,000 commits/hour — 394x more throughput than GitHub's write limit.
Why GitHub reads are faster (for now)
GitHub is a monolith with in-memory caches on one server — auth, repo lookup, and file read happen in the same process with 0 network hops between components. Coregit is a distributed system on Cloudflare Workers where each component (KV cache, R2 storage, Durable Objects, Hyperdrive DB) is a separate network call.
For single-file reads, GitHub's architecture wins by ~1.1x. For multi-file writes, Coregit's atomic API wins by 2-3.6x — the overhead of distributed components is amortized across one call instead of 105. Diff and list operations are now at parity or faster thanks to KV caching and shared tree resolution.
Roadmap: The Session API (Zero-Wait Protocol) eliminates per-request auth overhead for agents using sessions. With the session hot layer, reads from recently-written data serve from Durable Object storage (~2ms) instead of R2 (~200ms). Warm list tree and list commits already beat GitHub today.
How to use these numbers
- For agent builders: multiply commit count × time saved to estimate compute savings
- For platform teams: compare rate limits (38 commits/hr on GitHub vs 15,000/hr on Coregit)
- For cost modeling: combine seat savings + compute savings + egress savings