posts/ideas/bring-your-own-agent-acp-portable-editors.md

Table of Contents

Post Idea: Bring Your Own Agent — Breaking Out of Editor Lock-In

Status: Idea Date: 2026-05-22 Source: AI-Powered Development slide series (Parts 1-3), Engineering Excellence talks

Hook

Most AI coding tools lock you into their model AND their agent. Cursor gives you Composer. VS Code gives you Copilot. You pick the editor, you get whatever agent they baked in. But what if you could bring any agent into any editor?

The Story

At Qwestly, we built a stack that's completely decoupled:

Any Editor → ACP Protocol → Any Agent → BYOM (any model)

We showed the team how to run Claude Code inside vim (via Avante.nvim), inside Zed (via ACP), and inside Cursor (via MCP config). Same agent, same MCPs, same skills — just different editors.

The key insight: tools and skills are harness-provided, not model-provided. Swap the backend model (Claude → DeepSeek) and you keep the full tool suite — filesystem, shell, git, MCPs. You just change the reasoning engine.

Technical Details

  • ACP (Agentic Control Protocol) — open standard from Zed/Anthropic/Cursor, like LSP for AI agents
  • Avante.nvim — Neovim plugin that spawns any ACP agent inside vim (with file tree, git diffs, split panes)
  • Zed — Rust-native editor with native ACP support, GPU-accelerated, parallel agents
  • Provider-switching aliases — one shell alias per backend (vim-cursor, vim-claude, vim-codex, vim-ollama)
  • MCP portability — MCP config lives with the agent (~/.mcp.json), not the editor, so it travels everywhere

The Vela Observation

Vela's "lost a friend" reaction to switching models ties in here. When you BYOA/BYOM, the agent stays the same (same UI, same tools, same interaction pattern), but the model voice changes. This surfaced a UX question: should agents abstract away model personality? Or is personality part of the workflow?

Angle

This isn't just for engineering teams. Any team using AI tools benefits from a portable agent setup. The protocol layer (ACP) is the unlock — once you stop hard-coding the agent-editor relationship, the whole ecosystem opens up.

Potential Structure

  1. The problem: editor lock-in (cursor → composer only)
  2. ACP as the solution (the "LSP for agents")
  3. Our stack in practice (vim, zed, cursor, all running same agent)
  4. BYOM: swapping models behind the agent (DeepSeek costs $0.435/M vs Opus $5.00/M)
  5. The human factor: Vela's "lost a friend" and what it says about personality vs protocol
  6. Key takeaway: decouple editor, agent, and model

Tags

engineering ai tooling acp zed vim deepseek devtools