A local code graph for AI agents that grounds reasoning in actual project structure and refuses speculative hallucinations.

Key facts

Overview

Sonde is a minimalist open-source tool designed to prevent a specific failure mode in AI agent code reasoning: when agents (whether Claude, GPT, or other LLMs) are asked to refactor or analyse code, they often hallucinate about the existence, location, or signature of functions, classes, and APIs that do not actually exist in the codebase. This leads to agent-generated patches that fail to apply, contradictory instructions, or broken code suggestions. Sonde addresses this by building a queryable graph of a project's actual code structure and exposing it as a tool that agents can call before proposing changes. The author, Anish Moncivarghese, presents Sonde as an experimental research project on GitHub. The tool parses source code (language support unspecified but likely Python and JavaScript initially) into an abstract syntax tree, constructs a local graph of symbols (functions, classes, modules, imports), and exposes a query interface. An agent can call functions like "list_functions_in_module" or "get_signature_for_function" and receive ground-truth answers rather than LLM hallucinations. Sonde runs entirely locally with no cloud calls, improving privacy and latency. The project is open-source and free with no commercial offering. There is no packaged distribution, web interface, or managed service. Setup requires cloning the GitHub repository, understanding the code, and integrating the graph-building and query APIs into a custom agent orchestration loop (using tools like LangChain or direct LLM function-calling). Documentation is minimal—the README provides an overview of the concept but lacks integration examples, API reference, or guidance on polyglot or monorepo support. Sonde is a research-level tool, not a competitor to Cursor or Copilot. Those mainstream tools are already trained on code and have some implicit understanding of structure. Sonde is instead a grounding mechanism for teams building bespoke agents or testing whether explicit structure graphs reduce hallucination rates. Its immediate utility is limited to researchers and teams building custom agent systems. The lack of production maturity, integrated UX, or commercial support means it remains a proof-of-concept.

Who it is for

• Best for: AI agent systems that need to reason about code structure safely during refactoring, analysis, or generation tasks, particularly when hallucinations about missing functions or file paths have caused agent errors in the past..
• Avoid if: You are using standard IDE-based AI assistants (Cursor, Copilot) which already have code awareness; you need a commercially supported product; or your project uses uncommon languages or monorepo structures not yet validated by Sonde..

Who this is for

AI researchers and ML engineers experimenting with agentic code reasoning. Teams building bespoke autonomous refactoring or analysis tools may find this useful if they are implementing their own agent orchestration. Software architects evaluating how to ground LLM reasoning in actual system knowledge, rather than training or prompt-injection, may see value in the graph-based approach.

Who should skip this

Most professional developers should skip this. It is not integrated with Cursor, VS Code, or any commercial IDE. Teams without systems programming or ML infrastructure experience should avoid; integration requires understanding of how agents invoke tools. Any team prioritising speed-to-productivity over research-grade robustness should wait for a mature, productised version or recommend developers use Cursor instead.

Verdict

Sonde identifies a genuine problem—LLM hallucinations about code structure—and proposes an elegant local-graph-based solution. However, the project is too early and research-oriented for mainstream adoption. It is valuable for ML engineers exploring grounding techniques but unsuitable for commercial or team-based development workflows. Recommend for academic study; monitor for future maturation into a productised tool.

Sonde FAQ

What is Sonde? +

Sonde is a minimalist open-source tool designed to prevent a specific failure mode in AI agent code reasoning: when agents (whether Claude, GPT, or other LLMs) are asked to refactor or analyse code, they often hallucinate about the existence, location, or signature of functions, classes, and APIs that do not actually exist in the codebase. This leads to agent-generated patches that fail to apply, contradictory instructions, or broken code suggestions. Sonde addresses this by building a queryable graph of a project's actual code structure and exposing it as a tool that agents can call before proposing changes. The author, Anish Moncivarghese, presents Sonde as an experimental research project on GitHub. The tool parses source code (language support unspecified but likely Python and JavaScript initially) into an abstract syntax tree, constructs a local graph of symbols (functions, classes, modules, imports), and exposes a query interface. An agent can call functions like "list_functions_in_module" or "get_signature_for_function" and receive ground-truth answers rather than LLM hallucinations. Sonde runs entirely locally with no cloud calls, improving privacy and latency. The project is open-source and free with no commercial offering. There is no packaged distribution, web interface, or managed service. Setup requires cloning the GitHub repository, understanding the code, and integrating the graph-building and query APIs into a custom agent orchestration loop (using tools like LangChain or direct LLM function-calling). Documentation is minimal—the README provides an overview of the concept but lacks integration examples, API reference, or guidance on polyglot or monorepo support. Sonde is a research-level tool, not a competitor to Cursor or Copilot. Those mainstream tools are already trained on code and have some implicit understanding of structure. Sonde is instead a grounding mechanism for teams building bespoke agents or testing whether explicit structure graphs reduce hallucination rates. Its immediate utility is limited to researchers and teams building custom agent systems. The lack of production maturity, integrated UX, or commercial support means it remains a proof-of-concept.

How much does Sonde cost? +

Sonde pricing: Open-source; free. Always confirm current pricing on the official site, as plans change.

Does Sonde have a free tier? +

Yes. Sonde offers a free plan or free credits you can use to evaluate it.

What is Sonde best for? +

AI agent systems that need to reason about code structure safely during refactoring, analysis, or generation tasks, particularly when hallucinations about missing functions or file paths have caused agent errors in the past..

When should you avoid Sonde? +

Avoid Sonde if: You are using standard IDE-based AI assistants (Cursor, Copilot) which already have code awareness; you need a commercially supported product; or your project uses uncommon languages or monorepo structures not yet validated by Sonde..

What are the main pros of Sonde? +

Reduces AI agent hallucinations about code structure by enforcing queries against actual parsed project graphs; Local execution with no external API calls or cloud dependency, improving privacy and latency for agents; Directly addresses a known failure mode of LLM-based code assistants: inventing APIs, functions, or file paths that do not exist.

What are the main cons of Sonde? +

Early-stage research code with no packaged distribution, integration examples, or production documentation; Unclear how it scales to large codebases or polyglot projects with mixed languages; Requires integration into custom agent workflows; not a drop-in plugin for existing IDE tools or agent frameworks.

Does Sonde have an affiliate program? +

No public affiliate program is listed for Sonde at the time of review.

How is Sonde rated? +

WireTensors rates Sonde 3.3 out of 5, based on capability, value, and fit for its intended use case.

What category does Sonde fall under? +

Sonde is categorised under coding on WireTensors.

When was this Sonde review last verified? +

This review was last verified on 2026-08-31 against the vendor's official site.

Sources

• Sonde — official website — verified 2026-08-31