Redcon: Context Budget Management Tool for AI Programming Agents, Reducing Token Usage by 80%

Redcon (formerly ContextBudget) is a context management tool for AI programming agents. It addresses the pain point of high token consumption caused by irrelevant context in AI programming assistants through intelligent file sorting, multi-strategy compression, and token budget control. It can reduce token usage by 80% without compromising code understanding quality, supports MCP integration and VS Code extension, and adopts a deterministic design and dual-license model.

When using AI programming assistants like Claude Code and Cursor, there are two inefficient approaches:

1. Approach 1: Throw the entire codebase (a medium-sized microservice repository can reach 200,000 tokens) to the model, costing several dollars per request;
2. Approach 2: Let the agent search blindly, requiring multiple tool calls and easily missing key files. Redcon proposes a third approach: sort and compress files using deterministic algorithms before sending, and only package the most relevant content.

Redcon Four-Step Pipeline

1. Scan: Incrementally scan the code repository and cache file metadata (path, language, import relationships, etc.);
2. Sort: Determine file relevance through multi-dimensional scoring (keyword matching, import graph analysis, file role recognition, Git history analysis);
3. Compress: Select strategies (full content, code snippets, symbol extraction, summary) based on file type/importance, supporting mainstream languages;
4. Package: Select the optimal file combination within the token budget and output machine/human-readable results.

MCP Integration (Pull Mode)

Expose 5 tools for agents to call on demand: redcon_rank (file sorting), redcon_overview (repository structure), redcon_compress (single file compression), redcon_search (range search), redcon_budget (budget planning), completing exploration with approximately 5000 tokens.

Redcon provides a VS Code extension with the following features:

• Sidebar conversation: Enter tasks to view the packaging process in real time;
• Dashboard: Display token budget usage, compression strategy distribution, etc.;
• Status bar: Show current budget usage and risks;
• CodeLens: Display compression strategy and token count at the top of files;
• File decoration: Show relevance score badges in the file browser.

Redcon is suitable for the following scenarios:

• Large codebases: Microservices, monolithic applications (many files, complex dependencies);
• Multi-repository workspaces: Unified processing of cross-repository tasks;
• CI/CD integration: Automatic auditing of context growth;
• Cost-sensitive scenarios: Reduce operating costs of token-based APIs (e.g., OpenAI, Anthropic).

Redcon emphasizes determinism (same input produces same output), with the following advantages:

1. Reproducible: Easier debugging and testing, no random issues;
2. No embedding dependencies: Runs locally, no additional API costs;
3. Privacy-friendly: Code does not leave the local environment, suitable for sensitive repositories. Compared with RAG solutions, the rule + heuristic approach is better in terms of cost and privacy.

Redcon adopts a dual-license model:

• Core engine, CLI, plugins, cache: MIT open source license;
• Gateway, control plane, proxy middleware, LLM integration: Commercial license. Individuals/small and medium-sized enterprises can use core functions for free; enterprise-level functions require commercial authorization.

Redcon is a pragmatic AI engineering approach. By filtering and compressing at the input end, it allows AI agents to focus on key content, improving API call efficiency and cost savings. For developers who use AI programming assistants daily, it is a worthwhile optimization solution. Although it does not completely change the workflow, it can make each call more efficient.