Bring the harness you already use.
Plug in opencode, Claude Code, or Codex. Copilot routes your work through whichever you pick, adding vault context, permissions, and Obsidian-native output on top. Your existing subscriptions come with you. No second bill. No second key.
Skills and MCPs, in one place.
You shouldn't have to maintain three copies of skills and MCP servers, one per agent. In v4, Copilot manages them for you.
Add a skill or MCP server once. Toggle it on for any agent (opencode, Claude Code, Codex) with a click. Mix and match per agent. Change your mind anytime.
Your vault. Not your filesystem.
By default, agent harnesses can read anything on your filesystem. Copilot restricts them to your active vault.
Every harness Copilot launches is sandboxed before it starts. The agent is programmatically blocked from reading anything outside your vault. That includes ~/Documents, SSH keys, and any other personal files on your machine.
Nothing changes behind your back.
No agent should edit your notes without showing you the change first.
Every write (create, rename, modify, delete) is staged as a reviewable diff. Approve, reject, or send feedback. Review is the default. Auto-apply is an opt-in for the workflows you trust.
Use the model you trust, for the task at hand.
Different work wants different models. Copilot v4 lets you pick.
BYOK continues to work with opencode. Use whichever provider fits the job in front of you: Anthropic for nuanced reasoning, OpenAI for breadth, Google for long context, your own deployment for whatever reason. Your keys, your bill, your call.
Offline mode, local model only.
Point Copilot at a local model and your vault stays on your machine.
Offline mode routes all model inference to a local, OpenAI-compatible runtime (Ollama, llama.cpp, your own server). No vault content is sent to a remote model. No notes, no metadata, no embeddings.
Built for knowledge workers, not codebases.
You can point a CLI coding agent at your vault. People do. It works the way a hammer works on a screw.
Coding agents are tuned for codebases. Copilot v4 is tuned for the way knowledge workers think in Obsidian.
Both can ride the same agent harness. Only one was built to be your thinking partner.