OpenCode + GitNexus: Real Knowledge Graph & Memory for Your Agent

Hinzugefügt: 2026-05-12

[00:00:00]Most coding agents have one weakness that nobody really talks about. They read your code file by file. So, when you ask Anti-Gravity or Cursor or Claude Code to change one function, the agent looks at that function, edits it, and moves on. It doesn't know that 40 other places in the code base depend on what that function returns. It doesn't know which class extends which or which import chain breaks when you rename a method. The change looks clean in the diff. You approve it, and 3 days later, something silently fails in production.

[00:00:33]This happens because the agent has no map of your code base. It only sees what's in front of it at that moment.

[00:00:40]Git Nexus is the fix for that blindness.

[00:00:43]It's built by Abhigyan Paturi and the team at Aicon Labs. And the way they describe it is a nervous system for agent context. That's a fair way to put it. The tool scans your entire code base and builds a knowledge graph out of it.

[00:00:57]Every function call, every import, every inheritance chain, every execution flow from entry point all the way to the leaf nodes. All of that gets stored in a real graph database on your machine. Then it exposes that graph through MCP, so your agent can ask actual architectural questions before making any change.

[00:01:17]Instead of grepping and guessing, it queries the graph and gets the truth.

[00:01:21]The repo supports 14 languages, including TypeScript, JavaScript, Python, Java, Kotlin, C#, Go, Rust, PHP, Ruby, Swift, C, C++, and Dart. So, whatever stack you're on, there's a good chance it works out of the box. Okay, let me show you how to set this up in Anti-Gravity, because this is the part that trips people up. The Git Nexus repo has clear instructions for Cursor, Claude Code, Codex, and a few others, but Anti-Gravity is not on that list.

[00:01:52]So, you have to do a couple of extra steps. First thing, install Git Nexus globally instead of running it through NPX. Open your terminal and run npm install -g GitNexus. The global install matters because anti-gravity needs an absolute path to the binary, and that path has to stay the same across sessions. Once it's installed globally, go to the root of any project you want to index and run GitNexus analyze. Same command as before. If your project is not a Git repo, add {dash} {dash} skip Git. This one command walks the file tree, parses every file with tree-sitter, resolves imports across files, figures out which classes inherit from which, runs community detection to group related code into clusters, traces the execution flows from entry points, and builds a hybrid search index on top using BM25 plus semantic embeddings. All local, nothing leaves your machine. You can do this for as many repos as you want, and GitNexus keeps a global registry of all of them at {tilde}/.gitnexus/registry.json.

[00:03:01]When the analyze finishes, you will see a few new files in your project. There is an agents.md file, a cloud.md file, and a folder called.cloud/skills with four skill files inside. The skills are named exploring, debugging, impact analysis, and refactoring. Each one is a markdown file that tells your agent how to use the GitNexus tools for that specific task. This is the nudge that gets the agent to actually reach for the graph instead of falling back to grep.

[00:03:33]Now the anti-gravity part. Open anti-gravity, look at the top right corner, and click the three dots menu.

[00:03:40]You will see an MCP servers option.

[00:03:42]Click on it, then click manage MCP servers. This opens a config file called mcp_config.json.

[00:03:51]By default, it might be empty or just have curly braces. You need to add a GitNexus entry to it. But, before you paste anything, you need the absolute path to your global GitNexus binary, because Antigravity will not find it through path like a normal terminal would. To get the absolute path, open your terminal and run which GitNexus. On my machine, it returns the full path.

[00:04:15]Yours will be different depending on whether you use NVM, Brew, or a system node. Copy that full path, you will paste it in a second. Back in the mcp_config.json file, add an MCP servers block with a key called GitNexus. The command field gets the absolute path you just copied.

[00:04:34]The args field is an array with one string, just the word MCP. Save the file. The whole entry is maybe seven lines of JSON. After saving, Antigravity picks it up. Go back to manage MCP servers, and you should now see GitNexus listed with a green status. Click on it to expand, and you will see all the supported tools, things like list repos, query, context, impact, detect changes, rename, and cipher. If you see those tools, the connection is working. To verify it actually responds, start a new chat with the Antigravity agent and just ask it to run GitNexus list repos. It should come back with all the repos you have indexed, the file paths, node counts, edge counts, execution flows, and last commit hashes. When I tested it on my setup, it returned two repos, a chat app with 268 nodes and 294 edges, and an analytic dashboard with 366 nodes and 491 edges. It even flagged which one was my active workspace and offered to jump straight into that code. That is the signal that everything is wired up correctly. Now, the tools that matter day-to-day. Query does process grouped hybrid search using BM25 plus semantic embeddings. So, when you ask the agent to find the authentication flow, it returns the actual cluster of related functions instead of random text matches. Context gives you the 360° view of any symbol, every caller, every callee, every process that symbol participates in.

[00:06:06]Impact is the blast radius analysis. So, you can ask what breaks if I change this function and get a real answer with confidence scores per dependency. Detect changes maps a Git diff to all the affected processes. Rename does multi-file coordinated renames using the graph plus text search with a dry run option. And Cypher lets the agent run raw graph queries against Ladybug DB when it needs something custom. This is genuinely the part that changes how Antigravity behaves. Most agents fail on big refactors because they only see the file they are editing. With these tools, the agent asks the graph first, plans the change, and then executes. Even smaller models start performing like the big ones because the heavy lifting is done by the precomputed index, not the model's reasoning. There is also the web GUI, which is worth knowing about. Open gitnexus.versal.app in your browser.

[00:07:03]By default, the page shows a small instruction screen telling you to run GitNexus serve in your terminal. Run that command, it starts a local server on port 4747. The browser auto connects.

[00:07:15]You get the full interactive graph of your code base, files and folders in the left sidebar, a force-directed graph in the middle with every node colored by type, a search bar at the top, and a query button at the bottom left for raw Cypher. Top right has a settings icon and a Nexus AI button. Click settings, paste in an API key from OpenAI, Anthropic, Open Router, or Grok. And now you You chat with the code base directly in the browser. The chat uses a LangChain React agent that has access to the same MCP tools. So, answers are grounded in the graph rather than hallucinated. If you don't want to use the hosted page, the repo also has instructions to run the whole thing locally. You clone the repo, install dependencies in the GitNexus shared folder, build it, then go into the GitNexus web folder, install and run npm run dev. Then in another terminal, you run GitNexus serve as the back end. Same UI, fully offline, useful if you're on a corporate network or working with private code that you want extra layers around. There's also a Docker setup if that's more your thing. Two signed images on GHCR and Docker Hub. One for the CLI server and one for the web UI.

[00:08:28]You run Docker compose up -d and both come up. The images are signed with Cosign and tied to specific Git tags, which is honestly a level of supply chain security you don't see in most open source tools.

[00:08:42]They even ship a Kubernetes cluster image policy if you want to enforce signature verification at admission time. For pricing context, the open source version is licensed under PolyForm non-commercial, so it's free for personal and non-commercial use.

[00:08:58]They have an enterprise SaaS at akonlabs.com if you want PR review automation, auto re-indexing, multi-repo support, and OCaml language coverage.

[00:09:08]But for solo devs and most teams, the open source version covers everything you actually need. One last tip, after you make commits, the index can go stale. If you're using Claude code, the post-tool use hook will detect this and prompt the agent to re-index. For other editors, you just run GitNexus analyze again or GitNexus analyze {dash} {dash} force if you want a clean rebuild.

[00:09:31]Re-indexing on a medium repo takes maybe 30 to 60 seconds. That's GitNexus. Index your repo once, plug it into Open Code through MCP, and your agent stops shipping blind edits. But, the real value is the MCP server feeding pre-computed structure to whatever agent you use. Try it on a project you know well, and see if the graph surfaces relationships you forgot existed. All right, so that's it from the video, and I hope you enjoyed it. If you did, please like this video and subscribe to the channel, and I'll see you in the next video.