NexCode

What it is

A local-first, multi-agent AI coding assistant for VS Code. NexCode routes tasks to specialized agents — planner, coder, reviewer, QA, security — each optimized for a specific role, with native streaming from every agent stage.

The Problem

Single-agent coding assistants treat every task the same — writing a function, reviewing security, planning architecture, and testing all go through the same pipeline. This produces slow, shallow responses for complex multi-step work.

My Role

Built the complete VS Code extension: the agent-core orchestration layer, multi-agent pipeline with six specialized modes, provider routing for Ollama and OpenAI-compatible endpoints, and the sidebar UI with session history and per-session model controls.

Approach

Designed a multi-agent core where different modes (auto, planner, coder, reviewer, QA, security) route to different models based on capability requirements. Native live token streaming from each agent stage — no end-of-run replay. Persistent memory with workspace and episodic modes. Tooling layer for filesystem, terminal, git, tests, local code search, and web search. Deny-by-default terminal policy with symlink-safe operations and multi-layer secret redaction.

Hardest Decision

Designing the agent handoff protocol so planner, coder, reviewer, and QA agents can share context without losing state. Each agent produces structured output that the next agent consumes — but the context window is finite, so decisions about what to compress, cache, or discard directly affect response quality.

Status

Open source on GitHub with 2,000+ tests across 68 test files. Implements six agent modes with native streaming and multi-provider support. Currently in beta.

What I'd Change

Would implement a full context persistence layer across sessions, add streaming tool output visualization, and build a plugin system for custom agent modes.

Stack