@VincentLogic: This open-source project cuts Claude Code's costs by 25%. It doesn't build new models or a new IDE. It just draws a "code map" for the AI coding agent. Traditional approach: the model reads the entire repo → token explosion. Its approach: first parse the code with Tree-si…

This open-source project slashes the cost of Claude Code by 25%.

It doesn’t build new models or new IDEs.
It simply draws a “code map” for the AI coding agent.

Traditional approach: The model reads the entire repo → token explosion.
Its approach: First parse the code into a graph (functions, calls, references) using Tree-sitter, store it in local SQLite.
When the AI wants to know who calls a function, it doesn’t read the whole file—it just queries the graph.

In short: Others give the AI the whole book to read. It gives the AI a table of contents.

Official benchmarks ran on 7 large repos (VS Code, Django, Tokio, Gin, OkHttp…):

• Average token savings: 57%
• Cost reduction: 25%
• Speed improvement: 23%
• Tool call reduction: 62%

Open-source, supports 20+ languages, works with Claude Code / Cursor / aider.

Today’s AI coding tools aren’t competing on models—they’re competing on engineering that “saves brainpower” for the models.

GitHub: https://github.com/colbychherry/codegraph…

Previously everyone obsessed over “how big is the model’s context window”—1 million, 2 million, 10 million tokens. But if you feed an AI 1 million tokens of code, it either gets lost or misses the point.
What really matters isn’t “can it fit”, but “can it find what it needs.”

This kind of approach (codegraph) will become more common: docs with “semantic search,” code with “graphs,” internal company data with “structured indexes.” Underneath AI, it’s the same idea—don’t make the AI read everything; let it know where to look.

That’s also why RAG hasn’t died. It’s just evolved from “coarse chunking” to “structured querying.”

For an AI, irrelevant code in context is just noise—too much noise leads to hallucinations. With this “map,” the AI operates like it has a god’s-eye view, knowing where the roads are instead of blindly searching in a jungle of code. That’s real Context Engineering.

Absolutely, blindly RAG-ing the entire repo is a waste of tokens. With a code graph, the AI knows which parts to look at—precision strikes instead of carpet bombing.

Hermes Agent’s SOUL.md lets you directly define the agent’s role and behavior.

5 templates you can copy and paste:

• Researcher — Verify facts, check sources, organize evidence, don’t guess.
• Content Creator — Learn your style, find trends, write posts, avoid repeating last angle.
• Operations Manager — Manage inbox, meetings, deadlines, report only what really matters.
• Sales — Find leads, draft outreach, follow up, log objections, push deals forward.
• Founder’s Chief Assistant — Prioritize projects, identify execution bottlenecks, track goal progress, keep direction on course.

The logic of SOUL.md: Instead of writing a pile of system prompts every time, define once who the agent is and how it works.

Which role do you need most right now?