Obsidian Copilot V4 is an Obsidian community plugin that brings AI agents, chat, search, and editing tools into your vault. Its main feature is now Agent Chat, a desktop workspace that can run opencode, Claude Code, or Codex with permissions you choose.
That is a meaningful change from the earlier version of this article. Copilot is no longer best described as a chat box with Smart Vault Search. It is closer to an agent workspace that also keeps the smaller chat and editing tools.
If you are comparing the wider field first, my guide to Obsidian AI plugins separates full assistants, search tools, and smaller writing helpers. This page focuses only on Copilot V4.
What Obsidian Copilot V4 actually does
Obsidian Copilot lets an AI work with the notes and context inside your vault without making you copy everything into a separate browser chat. The exact abilities depend on the agent, model, context, and permission setting you select.
In Agent Chat, an agent can inspect notes, use tools, create Obsidian files, and continue a multi-step task across several turns. Copilot keeps this work inside an Obsidian interface while the connected agent does the reasoning and tool use.
The plugin still includes Quick Chat, Commands, Quick Ask, and local-first Miyo search. Those smaller tools matter when you want a rewrite, explanation, saved prompt, or quick conversation rather than a full agent session.
Why Agent Chat is the center of Copilot V4
Agent Chat is the main Copilot V4 workspace for tasks that need several steps. You can open separate sessions, add vault context, choose a model, and select how much freedom the agent has before it asks for approval.
This makes it useful for work such as reviewing unfinished notes, organizing research, drafting from existing material, or creating a plan from several files. The value is continuity. The agent can keep working with the same project context instead of starting cold in every message.
Permission choice deserves attention. A read-only review and a task that edits files are different risk levels. Start with the smallest permission level that can complete the job, then inspect proposed changes before granting broader access.
Obsidian Copilot only has useful context when you give it the right notes, folders, selections, or project material. More context is not always better. A smaller, relevant set makes the task easier to inspect and reduces accidental exposure.
A good first agent task has a visible finish line. Ask for a review and a proposed plan before asking for edits. Once the plan looks right, let Obsidian Copilot change a small set of files and review the result in Obsidian’s file history or your normal backup system.
Which agents can Obsidian Copilot use?
Copilot V4 currently offers three main agent paths. The easiest choice depends on which tools and subscriptions you already use.
opencode
opencode is the recommended starting point in the official setup guide. Copilot can download and manage its binary, then connect it to Copilot-hosted models, your own provider key, or a compatible local model.
Claude Code
If Claude Code is already installed, Copilot can detect the command-line tool and use its existing login. You do not paste the Claude Code account key into Copilot. Manual configuration is available when automatic detection fails.
Codex
Codex connects through the supported @agentclientprotocol/codex-acp adapter and uses the login held by its bundled Codex command-line tool. The current guide requires version 0.0.45 or newer of that adapter.
The practical benefit is that you can bring an agent you already pay for into Obsidian. Provider terms and usage limits still apply, and each path has its own installation requirements.
Obsidian Copilot does not erase the differences between these agents. Their model choices, tool behavior, login limits, and approval flows can differ. Pick the backend you already understand, then test it on a reversible vault task before widening its access.
Does Obsidian Copilot work on mobile?
Copilot works on desktop and mobile, but Agent Chat itself is desktop-only because its backends run local processes. On a phone or tablet, Quick Chat, Copilot Commands, and Quick Ask remain available for shorter jobs.
This split matters if mobile is your main Obsidian device. You can still ask questions and work with selected text, but the opencode, Claude Code, and Codex agent sessions require the desktop app.
For a mixed-device workflow, use Obsidian Copilot Agent Chat for heavier desktop work and the mobile tools for quick questions or edits. Do not plan around handing the same local agent process from a computer to a phone.
How do Projects, Skills, and Commands work?
Projects give ongoing work its own instructions, reusable context, and chat history. That keeps one topic or client separate from unrelated vault work and reduces the need to explain the same background in every new session.
Skills are reusable instruction packets that can be enabled for installed agents. Copilot also includes skills for Obsidian Markdown, Bases, Canvas, and the Obsidian CLI. A skill can appear across opencode, Claude, and Codex rather than being rebuilt for each one.
Commands are saved prompts for repeatable tasks. You can run them from Agent Chat with a slash, from the editor, or from Obsidian’s command palette. Quick Ask handles smaller questions beside the text you are already reading or editing.
This is where Obsidian Copilot becomes more useful than a generic chat window. A Project holds the working boundary, a Skill defines how a job should be done, and a Command gives you a fast way to start the same job again.
Is Obsidian Copilot free?
Yes. The open-source plugin works without a Copilot license when you use a supported agent account, your own provider key, a local model, or an OpenAI-compatible endpoint.
Free use includes normal single-agent chats, Projects with Markdown context, custom Skills and Commands, Quick Chat, Quick Ask, and local Miyo search. Your chosen provider or agent subscription may still have its own costs and limits.
Paid Copilot access adds hosted models and cloud-backed features. Some higher-tier features, including multi-agent work, require an active eligible plan. Check the official pricing page for current entitlements because models and limits can change.
Where does your data go?
Your vault remains a collection of local files, but prompts and any included note context are processed by the route you choose. That may be a CLI agent, cloud provider, local endpoint, or a Copilot-hosted service.
A local model keeps a different boundary from a hosted model. A Claude or Codex login also follows that provider’s terms. Review the provider and Copilot privacy details before adding private notes or granting an agent permission to edit files.
Before using Obsidian Copilot with sensitive material, confirm which model route is active and which files are in context. Also check whether a hosted feature is involved. The plugin interface alone does not tell you that every request stays local.
Obsidian Copilot vs a raw MCP setup
I already run an Obsidian MCP server to connect an outside AI client to my vault. MCP gives me a protocol layer and more control over the client, server, tools, and scope. It also leaves more setup and maintenance in my hands.
Copilot gives you a finished Obsidian interface around several agent paths. It handles chat sessions, Projects, shared Skills, saved Commands, permission selection, and smaller in-editor tools in one plugin.
Choose Copilot when you want the shorter route to agent work inside Obsidian. Choose a raw MCP setup when you want an outside client, a specific server design, or tighter control over how the connection is assembled. Running both can make sense when their jobs differ.
How do you set up Obsidian Copilot V4?
- Install and enable Obsidian Copilot from the Community Plugins browser.
- Open Settings, Copilot, Basic, then Agents.
- Download managed opencode, auto-detect Claude Code, or configure Codex through the supported adapter.
- Open the Agent Chat window from the ribbon or command palette.
- Select an installed agent, model, and permission level, then start with a concrete outcome.
The official Copilot V4 setup guide has the current commands and manual paths. Use it for installation details instead of relying on an older tutorial, especially for the Codex adapter.
Who should use Obsidian Copilot?
Copilot is a strong fit if your notes already live in Obsidian and you want an agent to work with them through a familiar interface. It is also a practical entry point if building and maintaining your own MCP connection sounds like unnecessary work.
It is a weaker fit if you work mainly on mobile, want no model provider to process any note context, or need a custom tool boundary the plugin does not expose. In those cases, smaller local tools or a carefully scoped outside connection may fit better.
Where this fits in my workflow
The useful lesson in Copilot V4 is that the agent matters less without durable instructions, context, permissions, and a repeatable process around it. Those are also the pieces I am building into Tandri.
If you want a connected system for managing projects, decisions, content, and AI-assisted work across tools, the Tandri page explains the approach and the current waitlist.
