When Akon Labs released GitNexus this past summer, the startup made a straightforward pitch: coding agents work better when they stop guessing what the code around them actually does. The company, which went through Y Combinator earlier this year, claims its open-source knowledge graph lifted agent success rates by 85% on a recent benchmark and has since attracted 46,000 stars on GitHub as of late August.
The underlying problem is one familiar to anyone who has watched an AI assistant confidently rewrite a function without checking what else depends on it. Large language models, no matter how capable, don't natively understand the branching structure of a codebase. They infer relationships from text embeddings or pattern-match their way through grep results, often with uneven outcomes. GitNexus tries to replace that guesswork with something more deterministic: a graph built from static analysis that traces dependencies, call chains, and execution flows before the agent ever proposes a change.
The tool parses repositories using Tree-sitter, resolves imports, and applies Leiden community detection to cluster related symbols. The result is a queryable graph stored locally in a .gitnexus/ folder. Agents access it through a Model Context Protocol server that exposes five core operations—impact, context, detect_changes, rename, and cypher. Instead of scanning files blind, an agent can ask which code paths a proposed edit will touch, or what a given function calls downstream, before committing to a rewrite.
According to a feature in MarkTechPost published in April, the graph layer itself runs without embeddings or LLM calls. The company supports 13 programming languages and integrates with tools including Claude Code, Cursor, Antigravity, Codex, Windsurf, and OpenCode. A web interface runs in-browser via WebAssembly for visualization, and Docker images allow self-hosted deployments. Multi-repository graphs can link cross-project dependencies; one example on the company's homepage shows a tRPC monorepo with 1,816 edges connecting three separate repositories.
The company ran vendor-reported benchmarks using DeepSWE, a long-horizon software engineering test suite that appeared on arXiv in July. The company put the same GPT model through 3,471 trials across 113 tasks, comparing performance with and without the code graph. Pass rates climbed from 36.99% to 68.37% when agents had access to GitNexus, while cost per solved task fell 51% to 88 cents. On the most difficult tasks, agents with the graph passed 55.9% of the time versus 15.9% without it, according to results posted on the company's benchmarks page.
Those numbers are vendor-reported, though the methodology page includes model details and trial counts. "The bottleneck isn't the model. It's context," founder Subham Kundu wrote in the startup's Y Combinator launch post.
Kundu and co-founder Abhigyan Patwari launched publicly on August 17. The company says it has onboarded more than ten enterprise customers and reached 65,000 weekly npm downloads according to vendor-reported figures, though npm's public page does not display download counts for verification and the enterprise client list remains undisclosed. GitHub metrics showed 46,000 stars and 5,100 forks as of August 28.
The software is open-source under the PolyForm Noncommercial license. Akon charges $29 per seat for managed SaaS, with custom pricing for self-hosted enterprise deployments that can run air-gapped. "Security is the architecture, not an afterthought," the company notes on its security page, emphasizing that code never leaves customer networks.
The competitive landscape includes Sourcegraph Cody, which taps Sourcegraph's own code graph and search infrastructure to feed context to agents. Cursor offers agent workflows and cloud orchestration but does not maintain a dedicated graph layer. Other tools like LangGraph, Microsoft Agent Framework, CrewAI, and LlamaIndex provide orchestration primitives for multi-agent applications, though they focus more on control planes than codebase structure. Microsoft's GraphRAG research, which began circulating in 2024, popularized graph-augmented retrieval for LLMs across a range of domains.
Akon is now building Nexus Agent, a coding agent that runs terminal and web sessions on top of the GitNexus graph. The product is in private beta with cohort-based access. The company highlights blast-radius checks and verification loops as distinguishing features. Dealroom data indicates Y Combinator invested $125,000, valuing the company somewhere between $500,000 and $750,000, though funding details beyond the initial YC check have not been disclosed.
Whether structural graphs become standard infrastructure for AI coding tools or remain a niche optimization is still an open question. For now, Akon Labs is betting that agents will keep running into the limits of inference until someone hands them a map.
