Working on a large Rust codebase. The token problem is real — Claude Code will happily spend $5 of context just trying to understand how two modules relate before writing a single line. And once context compaction kicks in, it's even worse — the agent loses the thread completely and starts grepping the same files again from scratch.
Approaches I've tried:
Feeding CLAUDE.md / architecture docs manually — helps, but gets stale fast. Cursor's built-in indexing — breaks on monorepos, and I don't love proprietary code going to their servers. Basic MCP server with grep — works for exact matches, useless for semantic queries.
Eventually built something more serious: a local Tree-sitter indexer that builds a knowledge graph of file relationships and exposes it via MCP so agents query semantically instead of grepping blind. One tool call instead of 15 grep iterations. Published it here: https://github.com/Muvon/octocode
But genuinely curious what others are doing before I go deeper on it.
Three specific questions:
1. How do you handle the "ripple effect" problem — knowing that changing one file semantically affects others that aren't obviously linked?
2. Do you trust closed-source indexing with proprietary code, or have you gone local-first?
3. Has anyone gotten GraphRAG-style relationship mapping to work in practice at scale, or is it still mostly hype?
I've been working on Octomind, an AI development assistant that addresses two problems I kept running into:
1. *Cost blindness* - Most AI tools don't show you what you're spending until the bill arrives
2. *Model inefficiency* - Using GPT-4 for everything when simpler tasks could use cheaper models
*What makes it different:*
*Real-time cost display:*
Every interaction shows exactly what you're spending:
```
[~$0.05] > "Explain this algorithm"
[~$0.12] > "Refactor for better performance"
[~$0.18] > "Generate comprehensive tests"
```
*Intelligent model routing:*
Configure different models for different task types. Simple queries go to cheap models (Claude Haiku, GPT-3.5), complex reasoning goes to premium models (Claude 3.5 Sonnet, GPT-4).
*MCP server integration:*
This is the part I'm most excited about. You can add specialized AI agents through configuration alone:
```toml
[mcp.servers.code_reviewer]
model = "openrouter:anthropic/claude-3-haiku"
```
Now `agent_code_reviewer(task="review this function")` is available in your session. No custom code needed.
*Technical details:*
- Written in Rust for performance
- Supports 6+ AI providers (OpenRouter, OpenAI, Anthropic, Google, Amazon, Cloudflare)
- Session-based architecture with full conversation history
- Built-in development tools via MCP (file operations, shell commands, etc.)
*Real usage example:*
Last week I used it to refactor a complex authentication system. Total cost: $0.23 for what would have been 3+ hours of manual work. The cost visibility helped me optimize my prompts and model selection.
I'd love feedback from the HN community. What features would make AI development tools more useful for your workflows? Are you tracking AI costs in your projects?
After bouncing between ChatGPT, Claude, and countless VS Code extensions for months, I got frustrated with the constant context switching and re-explaining my codebase to AI. So we built *Octomind* - an open-source AI assistant that actually understands your project and remembers what you've worked on.
## What's different?
*No more copy-pasting code snippets.* Octomind has semantic search built-in, so when you ask "how does auth work here?" it finds the relevant files automatically. When you say "add error handling to the login function," it knows exactly where that is.
*Built-in memory system.* It remembers your architectural decisions, bug fixes, and coding patterns. No more explaining the same context over and over.
*Real cost tracking.* Shows exactly what each conversation costs across OpenAI, Claude, OpenRouter, etc. I was shocked to see I was spending $40/month on random API calls before this.
*Multimodal support.* Drop in screenshots of error messages or UI mockups - works across all providers.
## The workflow that sold me:
```
> "Why is this React component re-rendering so much?"
[Finds component, analyzes dependencies, explains the issue]
> "Fix it"
[Implements useMemo, shows the diff, explains the change]
- *Does this solve a real pain point for you?* Or are you happy with your current AI workflow?
- *What's missing?* We're thinking about adding team collaboration features
- *Performance concerns?* It's built in Rust, but curious about your experience
The whole thing is Apache 2.0 licensed on GitHub. Would love to hear what you think - especially if you try it and it doesn't work as expected.
# Your AI coding assistant finally has a full "memory" of your codebase
Most of us have experienced that frustrating moment when working with Claude or another AI coding assistant:
"Can you help me fix this authentication bug?"
"Sure! Can you show me your authentication code first?"
"I just showed it to you two messages ago..."
This limitation has been driving me crazy. Our AI coding tools are brilliant but handicapped by their inability to see the big picture. They're like trying to fix a complex engine while only looking through a keyhole.
So I built *Octocode* to solve this fundamental problem.
## What makes Octocode different
Rather than just throwing more tokens at the problem, Octocode creates a genuine understanding of your entire codebase:
1. *Natural language code search*: Ask "where do we validate email addresses?" instead of trying regex patterns
2. *Complete codebase awareness*: It remembers every function, relationship, and dependency across your entire project
3. *Intelligent context management*: Instead of expensive token stuffing, it creates smart summaries that actually work better
4. *Tool integration*: Works with Claude Desktop, VS Code, and other AI assistants you already use
## Real results from daily use
I've been using this daily for several months, and the quality difference is profound:
- Before: Constantly re-explaining my own code structure
- After: AI immediately understands full context and relationships
- Before: Getting suggestions that would break other parts of the codebase
- After: AI understands dependencies and potential side effects
- Written in Rust for performance
- Uses vector embeddings to understand code semantics
- Works with 50+ AI models through OpenRouter
- Minimal resource usage (only processes what changed)
- Privacy-focused: your code stays local
I built this out of necessity - AI coding tools are amazing, but they've been operating with severe blindness. The difference now is like working with a senior developer who knows your codebase inside-out versus explaining your project to a new hire every day.
Questions or feedback? I'd appreciate hearing about your experience with AI coding tools and what problems you're looking to solve.
This is interesting — has anyone tried something like this with Elasticsearch or maybe Milvus/Meilisearch/Typesense? I know those are popular for vector searches, but I haven't seen anything specific for improving GitHub searches like this.
I'm curious how this compares performance-wise, especially when it comes to large repositories with tons of issues and PRs. Also, how scalable is it? I feel like semantic search has a lot of potential here, but does anyone know if GitHub itself has plans to integrate something similar?
I'm thrilled to see the direction we'll take with this. We initially started with just a few features, primarily because we're eager to understand what you want from a service like this and then refine it based on your needs. Don't hesitate to ask questions or make suggestions.