Coregit

Coregit is a serverless Git API built for AI agent infrastructure, offering atomic multi-file commits, semantic code search, and Git hosting through a REST API that runs on Cloudflare Workers.¹ The company launched on April 10, 2026² and open-sourced the codebase under AGPL-3.0 four days later.⁹ Coregit is part of the Founders Inc. Canopy accelerator batch.⁸

Founding

Alemzhan Jakipov, 18 years old and based in Semey, Kazakhstan, co-founded Coregit with a developer known publicly as "auraml." The founding blog post describes the team as "two developers from Kazakhstan."² Jakipov previously operated under the entity name Strayl AI; the Coregit blog footer reads "© 2026 Coregit, Inc."²

The founding blog post identifies GitHub's API rate limits as the impetus for building a new Git hosting product.² GitHub's API allows 500 write operations per hour, and committing 10 files requires 13 sequential API calls — capping any AI agent at 38 ten-file commits per hour.⁵

Architecture

Coregit reimplemented Git's entire object model (blobs, trees, commits, tags, packfiles, and refs) from scratch in TypeScript to run on Cloudflare Workers.⁷ Cloudflare Workers are V8 isolates with no filesystem access, no native bindings, and no subprocess calls, making it impossible to wrap the standard git binary or libgit2.²

The storage stack uses four Cloudflare services and two external databases. Cloudflare R2 stores git objects in a content-addressed layout mirroring Git's on-disk format, with zero egress fees. Cloudflare KV caches trees, auth tokens, metadata, and commit lists at 5–15ms latency. Cloudflare Durable Objects handle per-repo hot state and rate limiting at approximately 2ms. Neon PostgreSQL, connected via Cloudflare Hyperdrive, stores repository metadata, the code graph, and usage data.⁷ Pinecone provides the vector store for semantic search, using Voyage AI's voyage-code-3 model at 1,024 dimensions and rerank-2.5 for result reranking.⁷

The commit endpoint accepts 1 to 1,000 file changes in a single POST request. It resolves the branch ref, flattens the current tree to an in-memory map keyed by file path, applies all changes, builds new tree objects bottom-up with parallel subtree construction, creates the commit, and updates the ref atomically via compare-and-swap.² An "edit" action supports surgical line-range replacement or old/new string substitution so agents do not retransmit entire files.⁷

A six-layer cache hierarchy minimizes read latency: in-memory Map (0ms, 32MB cap), Durable Object (approximately 2ms), KV (5–15ms), Edge Cache API (under 5ms for immutable SHA objects with a 1-year TTL), Hyperdrive (0–5ms), and R2 (50–200ms).⁷ Commit endpoints return a 201 response with the new SHA in approximately 2ms after a Durable Object acknowledgment; the R2 durability write fires in parallel and completes 200–500ms later.² The web framework is Hono v4.7, handling 40+ API routes across Git Smart HTTP and REST endpoints.⁷

Benchmarks

Benchmarks run in April 2026 between Kazakhstan and the US-East region show a direct tradeoff: GitHub wins on single-file operations and Coregit wins as file count grows.⁵ Committing a single file takes 1,195ms on Coregit versus 610ms on GitHub. Committing 100 files takes 19,769ms on Coregit versus 72,064ms on GitHub — a 3.6x speed advantage — using one API call versus GitHub's 105 sequential calls.⁵

The write throughput gap is larger than the latency gap. GitHub's effective write ceiling for a 10-file commit is 38 commits per hour (500 write ops ÷ 13 calls per commit). Coregit's limit is 15,000 requests per hour per API key, and each commit of any size uses one request, giving a ceiling of 15,000 commits per hour — 394 times GitHub's limit.⁵ GitHub also beats Coregit on branch creation (832ms vs. 1,562ms) and branch listing (420ms vs. 758ms), because GitHub's servers colocate operations that Coregit distributes across Workers.⁵ The benchmark documentation presents these tradeoffs explicitly.⁵

The ROI documentation calculates that a team doing 5,000 ten-file commits per day would wait 11.7 compute hours on GitHub versus 5.8 hours on Coregit, saving approximately $266 per month at $1.50/hr compute cost.⁴ Cloudflare R2's zero egress pricing means a team mirroring 3TB per month would pay approximately $270 in AWS S3 egress fees and $0 on Coregit.⁴

LLM Wiki

On April 11, 2026, one day after the main launch, Coregit shipped LLM Wiki, a Git-backed knowledge base system for AI agents.³ Each wiki is a Git repository with a raw/ directory for immutable source documents, a wiki/ directory for LLM-generated pages with YAML frontmatter including related, sources, and confidence fields, and an auto-generated [[llms-txt|llms.txt]] file — a format for machine-readable site content.³ The knowledge graph is built from related frontmatter parsed into PostgreSQL. Semantic search works across both directories via the same Voyage AI and Pinecone pipeline used for code search.⁷

Security

API keys are stored only as SHA-256 hashes; plaintext is never stored or logged, and timing-safe comparison is used on each request.⁶ Two credential types exist: master API keys (org-scoped, prefixed cgk_live_) and scoped tokens (repo-scoped, prefixed cgt_, valid for 1 hour to 30 days).⁶ All writes produce an immutable audit log entry including actor ID, action, resource, IP address, request ID, and timestamp.⁶ Per-commit payload limits are 1,000 files, 10MB per file, and a 5MB REST body; pushes via Git Smart HTTP are capped at 32MB.⁶

Pricing

The free tier includes 5 repositories, 1GB storage, 10,000 API calls per month, and 5GB Git transfer per month.¹ The pay-as-you-go tier charges $0.10 per GB of storage per month, $0.50 per 1,000 API calls, and $0.10 per GB of Git transfer, with a $49 per month custom domain add-on.¹ Enterprise pricing is custom, with volume discounts, SSO/SAML, and dedicated SLAs.¹

Traction

Prettiflow, an AI code formatting and workflow company, was Coregit's first paying customer and hit 30,000 API requests within days of launch.⁹ Coregit's homepage features testimonials from three Prettiflow employees, including CEO Amit Dey.¹ The Hacker News submission, posted under the username AlenzhanJ, had 1 point at the time of research.¹⁰ The LinkedIn post announcing the open-source release received 216 reactions and 32 comments.⁹

References

1. Coregit — The Git layer for AI agent infrastructure(accessed Apr 18, 2026)
2. Introducing Coregit — Coregit Blog(accessed Apr 18, 2026)
3. LLM Wiki launch — Coregit Blog(accessed Apr 18, 2026)
4. ROI Benchmarks — Coregit Docs(accessed Apr 18, 2026)
5. Scalability Benchmarks — Coregit Docs