How a local CLI tool gives AI agents better file context without cloud costs

If you’ve ever watched an AI coding agent spiral into chaos after reading 500 irrelevant lines of code, you’re not alone. For many solo developers, the real bottleneck isn’t the agent’s intelligence—it’s the way these tools blindly consume files, flooding token budgets and cluttering context windows before any real work begins.

That’s the problem one developer set out to fix. After months of frustration watching token bills balloon and agents lose track of what they were supposed to do, they built unerr, a lightweight CLI that gives AI agents "eyes" on your codebase without ever sending your files to the cloud.

Instead of letting an agent read an entire file—only to discover halfway through that the code is irrelevant—unerr intercepts those requests and serves up only the structural context the agent actually needs. The result? Faster responses, lower costs, and a lot less frustration when debugging.

The hidden cost of blind file reads in AI coding agents

Most AI coding agents today operate on a simple but inefficient principle: when they need to understand a file, they read the whole thing. Whether it’s a 20-line utility or a 500-line module, the agent doesn’t know what’s important until it’s already consumed the tokens.

For solo developers and small teams, this adds up fast. Every unnecessary file read inflates token usage, slows down responses, and clogs the agent’s context window with irrelevant noise. Tools like Graphify and RTK exist to help, but many rely on post-processing or external LLMs to compress context after the damage is done. That’s not just slow—it’s expensive.

The developer behind unerr wanted something different: a way to pre-filter file context entirely on your machine, using only deterministic logic. No cloud dependencies. No extra API calls. Just a tool that runs locally and gives agents the right information before they even ask for it.

How unerr turns raw files into structured intelligence

At its core, unerr acts as a smart intermediary between your AI agent and your codebase. It doesn’t replace the agent—it enhances it by controlling what gets fed into its context window.

When you run unerr, it starts by indexing your repository using Tree-sitter to parse code into an Abstract Syntax Tree (AST). It then stores this structure in an embedded local database (CozoDB), building a map of your codebase’s entities, dependencies, and relationships.

When your agent requests a file, unerr doesn’t serve the raw contents. Instead, it returns:

• The exact structural entities (functions, classes, variables) referenced in the request
• Caller and callee relationships that show how modules interact
• Dependencies that are actually relevant to the task at hand

This means your agent gets the context it needs without ever touching irrelevant code. The tool doesn’t guess what’s important—it calculates it, deterministically, before the agent even starts reading.

Why this matters for solo developers and small teams

For developers working alone or in small teams, token efficiency is everything. Every extra 1,000 tokens costs money. Every slowdown in agent response time breaks momentum. And every irrelevant file read introduces noise that can derail the entire session.

unerr addresses all three pain points:

• Cost savings: By filtering context upfront, it reduces token usage by up to 70% in early tests
• Faster responses: Agents spend less time parsing noise and more time solving problems
• Reliability: Since everything runs locally, there’s no dependency on external APIs or cloud services

The developer behind the project emphasized that this approach is especially valuable for developers building on limited budgets or working in restricted environments where cloud-based AI tools aren’t an option.

Ready to test it yourself? Here’s how to try unerr

If you’ve ever wished your AI agent could read code smarter, unerr is worth a try. The tool is open for beta testing, and the developer is actively looking for feedback from a wide range of environments.

To install unerr, run:

npm install -g @unerr-ai/unerr

Once installed, you can use it as a local proxy between your agent and your codebase. The setup requires no accounts, no API keys, and no cloud processing—just a few seconds of indexing time before your agent starts working.

The developer is particularly interested in hearing about:

• Installation issues across different operating systems and Node setups
• Whether the pre-filtered context matches what your agent actually needs
• Any unexpected behavior or edge cases that haven’t been caught in testing

If you run into a bug or just want to share your experience, the developer is monitoring responses closely. Given the "Works on My Machine" curse is real, broad testing is the only way to ensure unerr delivers on its promise.

The future of local-first AI development tools

As AI coding agents become more powerful, the tools that support them need to evolve too. The shift from cloud-dependent workflows to local-first development isn’t just a trend—it’s a necessity for developers who value control, cost efficiency, and reliability.

Projects like unerr are paving the way for a new generation of tools that give AI agents the context they need without the overhead. Whether you’re debugging a legacy codebase or building a new feature from scratch, having the right context at the right time can make all the difference.

If you’re curious about how local context optimization could change your AI coding workflow, now’s the time to try unerr. The project is still in its early stages, but the potential is clear: smarter agents, lower costs, and a lot less frustration when your codebase grows.

The developer plans to expand unerr’s capabilities based on community feedback, with potential integrations for more programming languages and deeper analysis tools. For now, though, the focus is on making sure the core experience works reliably across as many setups as possible.