Most AI coding agents understand your repository the same way a junior developer would on their first day — by scanning raw files in bulk. CodeGraphContext (CGC) changes that by turning your codebase into a queryable knowledge graph. When you pair it with RooCode and Regolo.ai’s serverless inference, you get full architectural awareness without bloating the context window or the invoice.
The real problem with file-based context
When a coding agent needs to fix a bug or plan a refactor, the standard approach is to shove as many relevant files as possible into the prompt. A 5,000-line Python file is roughly 40,000–60,000 tokens; at 100 requests per day, that alone adds up to hundreds of dollars per month on a single feature. The model is also working blind to structure: it does not know which functions call which, where circular dependencies hide, or what is genuinely dead code.[1]
This is not just a cost problem, large and unstructured context leads to what researchers call “context rot” — model accuracy degrades as prompts grow.
A pre-structured graph is a fundamentally different proposition, in a recent benchmark across 7,928 queries and 45 domains found that a graph-based retrieval system uses 11× fewer tokens than standard RAG (269 tokens per query vs. 2,982) while producing answers that are 4× more accurate.
The token-efficiency advantage also compounds as query complexity increases, with F1 scores rising from 0.374 at single-hop to 0.772 at five-hop reasoning tasks, while RAG peaks at two hops and then declines.
What CodeGraphContext actually is
CodeGraphContext is a dual-mode tool: a CLI for developers and an MCP (Model Context Protocol) server for AI agents. It parses your source code with Tree-sitter and SCIP to extract semantic relationships — function calls, class hierarchies, module imports, circular dependencies — and stores them in a graph database. You can choose between KùzuDB (embedded, zero-ops), FalkorDB (high-performance), or Neo4j (enterprise-scale with visual dashboards).
Once indexed, the graph is queryable in two ways. From the terminal, you run commands like cgc analyze callers my_function --all to get a precise list of every direct and indirect caller, with zero false positives from comments or string matches. From an AI agent, the MCP server exposes tools such as find_code, analyze_code_relationships, execute_cypher_query, and find_dead_code that let the model navigate your codebase symbolically rather than by ingesting file blobs.
A live file watcher keeps the graph synchronized with local changes automatically. For sharing or CI pipelines, you can package an indexed codebase as a .cgc bundle for instant loading without re-indexing.
Installation in five minutes
Install from PyPI:
pip install codegraphcontext
cgc --versionCode language: Bash (bash)If the cgc command is not found, patch the PATH:
curl -sSL https://raw.githubusercontent.com/CodeGraphContext/CodeGraphContext/main/scripts/post_install_fix.sh | bash
source ~/.bashrc # or ~/.zshrcCode language: 1C:Enterprise (v7, v8) (1c)Run the interactive setup wizard, which guides you through database selection and IDE configuration:[8]
cgc setupCode language: Bash (bash)The recommended local option uses Docker to spin up a Neo4j instance automatically. Once setup is complete:
cgc start # starts the MCP server
cgc index /path/to/your-project # builds the initial graphCode language: Bash (bash)To enable live monitoring of a directory:
cgc watch /path/to/your-projectCode language: Bash (bash)Connecting CodeGraphContext to RooCode and Regolo
RooCode is an autonomous coding agent that runs inside VS Code and speaks the OpenAI-compatible API format — which means it connects directly to Regolo.ai’s serverless GPU inference with no extra configuration. Here is the full setup.
Step 1 — Install RooCode.
Open the VS Code Extensions view (Ctrl+Shift+X), search for “Roo Code”, and install it.
Step 2 — Configure Regolo the provider.
Open the RooCode panel, select “openai compatible” as the API provider, and enter the following:
- Base URL:
https://api.regolo.ai/v1(verify against the latest Regolo.ai docs) - API key:
YOUR_REGOLO_API_KEY - Model: replace
MODEL_ID_PLACEHOLDERwith your chosen model from the Regolo.ai catalogue
This gives RooCode access to open models running on Regolo’s GPU infrastructure — no idle cost, and with zero data retention on inference content, which matters when your prompts contain proprietary source code.
Step 3 — Install CodeGraphContext
Install in seconds with pip and unlock a powerful CLI for code graph analysis.
pip install codegraphcontextCode language: Bash (bash)Step 4 — Index your project.
In the project path, open termianel and type the following command to index your current project:
codegraphcontext index .Code language: Bash (bash)CodeGraphContext will parse the repository and build the graph, RooCode can issue graph queries instead of raw file reads.
Step 5 – Setup MCP for AI Assistant
Configure your AI assistant to use CodeGraphContext
Setup: Run the MCP setup wizard to configure your IDE/AI assistant:
codegraphcontext mcp setupThe wizard can automatically detect and configure:
- VS Code
- Cursor
- Windsurf
- Claude
- Gemini CLI
- ChatGPT Codex
- Cline
- RooCode
- Amazon Q Developer
- Kiro
Upon successful configuration, codegraphcontext mcp setup will generate and place the necessary configuration files:
- It creates an
mcp.jsonfile in your current directory for reference. - It stores your database credentials securely in
~/.codegraphcontext/.env. - It updates the settings file of your chosen IDE/CLI (e.g.,
.claude.jsonor VS Code’ssettings.json).
Start: Launch the MCP server:
codegraphcontext mcp startCode language: Bash (bash)Use: Now interact with your codebase through your AI assistant using natural language! See examples below.
Step 6 – Setup Regolo as provider
Use open source models to develop your application through Regolo Core Models or publish your own from Huggingface url repo.
Here the detailed guide to How to Use Roo Code with Private EU Inference in VSCode.
For CLI Toolkit Mode
Start using immediately with CLI commands:
# Index your current directory
codegraphcontext index .
# List all indexed repositories
codegraphcontext list
# Analyze who calls a function
codegraphcontext analyze callers my_function
# Find complex code
codegraphcontext analyze complexity --threshold 10
# Find dead code
codegraphcontext analyze dead-code
# Watch for live changes (optional)
codegraphcontext watch .
# See all commands
codegraphcontext helpCode language: Bash (bash)The token savings benchmark: what graph context actually costs
Graph-based context retrieval is not just a qualitative improvement — the numbers are concrete and reproducible.
Tokens per query. The CKG benchmark across 7,928 queries shows that a pre-structured knowledge graph consumes an average of 269 tokens per query versus 2,982 for standard RAG and 3,450 for Microsoft GraphRAG. That is an 11× reduction while simultaneously delivering 4× better answer quality (F1 score 0.471 vs. 0.123 for RAG). The advantage grows with query complexity: multi-hop reasoning tasks, exactly the kind an agent uses when tracing a call chain or planning a refactor, are where graph retrieval pulls furthest ahead.
Cost per query at real pricing. Consider a code assistant processing a repository that produces a 4,000-token context per query. Without graph-based retrieval, a typical setup costs roughly $0.15 per query; with graph-based context and prompt caching, that drops to around $0.02 per query — an 87% reduction. At 100 daily queries for a single developer, the difference is approximately $130 per month saved on one workflow.
Prompt caching on top of graph context – the two optimizations stack, because graph-derived context is structurally stable across requests (the same function relationships return the same structured data), the system prompt prefix is highly cacheable. A comprehensive study across 500+ agentic sessions with 10,000-token system prompts found that prompt caching reduces API costs by 45–80% and time-to-first-token by 13–31% across OpenAI, Anthropic, and Google.
The best-performing configurations achieved 78–81% cost reduction on GPT-5.2 and Claude Sonnet 4.5. Using Open Source models you can increase the costs even more.
In a real production scenario, ProjectDiscovery’s agentic security platform went from a 7% cache hit rate to 84% after restructuring its prompts to separate stable system content from dynamic tool results — cutting total LLM spend by 59%. For their most complex tasks (67.5 million tokens across 1,225 steps), the cache hit rate reached 91.8%. The takeaway for CGC users: because graph query results are deterministic and compact, they make excellent stable context that benefits maximally from prompt caching.
| Retrieval method | Tokens per query | F1 score | Notes |
|---|---|---|---|
| Standard RAG | 2,982 | 0.123 | |
| Microsoft GraphRAG | 3,450 | 0.120 | |
| Pre-structured graph (CKG) | 269 | 0.471 | |
| Code assistant (no cache) | ~4,000 | — | $0.15/query |
| Code assistant (graph + cache) | ~400–600 | — | $0.02/query |
The practical recommendation is to structure your Regolo.ai prompts with CGC-derived context in the stable system-prompt prefix and the user’s natural-language question at the tail. This makes the cacheable portion maximally large and maximally reusable.
FAQ
Does CodeGraphContext support languages other than Python?
Yes. CGC uses Tree-sitter under the hood, which supports 14+ languages including JavaScript, TypeScript, Java, Go, Rust, C++, and more. The graph model is language-agnostic: nodes represent functions, classes, and modules regardless of language, and the same Cypher queries work across a polyglot codebase.
Will CGC break the prompt cache if the graph results change?
Graph queries return deterministic, structured data for a given codebase state. As long as you place the graph-derived context in the stable portion of the system prompt (before any dynamic user content), it behaves as an excellent cache prefix. The live file watcher only updates the graph when files actually change, so for most queries during a session the cached prefix remains valid.
What database backend should we use for a team setup?
For local single-developer use, KùzuDB (embedded) requires no infrastructure. For a shared team environment, Neo4j AuraDB (hosted) lets everyone connect to the same graph, and CGC’s .cgc bundle format makes it easy to share a pre-indexed snapshot without requiring each developer to re-index.
Is Regolo.ai compatible with all RooCode features?
RooCode accepts any OpenAI-compatible provider. All core features — natural language file manipulation, terminal execution, web browser automation, Custom Modes, Boomerang Tasks, and MCP integration — work regardless of which compatible provider you configure. Check the Regolo.ai documentation for the current list of available models, as the catalogue evolves.
Does this setup work for non-coding repositories (e.g., infrastructure-as-code)?
CGC is optimized for source code with Tree-sitter parsers. For Terraform, Kubernetes YAML, or similar structured configuration files, the graph indexing may be partial depending on available language parsers. The Cypher query layer, however, can be adapted to any graph that CGC builds, so infrastructure-as-code files with supported parsers would benefit from the same dependency-tracing capabilities.
What about data privacy when using CodeGraphContext with Regolo?
Only the graph query results — compact structured data about relationships — are sent to the inference provider, not raw source files. Combined with Regolo.ai’s zero data retention policy and european data center hosting, this means your codebase content is never stored externally and never leaves European jurisdiction. For teams working on proprietary or sensitive codebases, this is a meaningful architectural difference compared to providers with data retention policies.
Start your free 30-day trial at regolo.ai and deploy LLMs with complete privacy by design.
👉 Talk with our Engineers or Start your 30 days free →
- Discord – Share your thoughts
- GitHub Repo – Code of blog articles ready to start
- Follow Us on X @regolo_ai
- Open discussion on our Subreddit Community
Built with ❤️ by the Regolo team. Questions? regolo.ai/contact or chat with us on Discord