TokenPilot: Local AI Agent Intelligence

TokenPilot is an open-source, local-first intelligence layer for AI coding agents. Modern coding agents are powerful, but they often waste expensive model context on work a developer’s machine can already do locally: scanning repositories, finding relevant files, detecting frameworks, preparing implementation plans, and rediscovering old project decisions. TokenPilot solves this by sitting before tools like Codex, OpenCode, Claude Code, Cursor, and other MCP-compatible agents. TokenPilot scans a repository locally, detects the stack, selects relevant files, compresses context, prepares an evidence-backed implementation brief, and saves reusable project memory. The coding agent can then focus on final implementation instead of spending paid tokens repeatedly understanding the same codebase. The MVP includes a TypeScript CLI, MCP server, Hono API, Next.js dashboard, local JSON memory, provider adapters, Docker support, and test coverage. It is not a SaaS and does not require login or subscriptions. Code stays on the developer’s machine unless remote model providers are explicitly enabled. For harder planning tasks, TokenPilot can connect to local models or AMD/Fireworks-compatible providers, making it a practical hybrid layer between local intelligence and remote AI acceleration. TokenPilot’s goal is simple: make AI coding agents cheaper, more private, and more context-aware.