Agent Chat

Agent Chat is the default Copilot experience on desktop. It gives an AI agent a working view of your vault so it can answer questions, use tools, and make permissioned changes while you follow the work in chat.

Agent Chat follows a growing response until you scroll up to read earlier messages. Select the round down-arrow near the bottom of the chat, or scroll to the end, to follow new output again.

Quick Chat remains available for lightweight conversation and is the main chat experience on mobile. For multi-step work, Projects, Skills, or file changes, start with Agent Chat.

Your sent messages display Markdown formatting, including headings, lists, code blocks, links, and image embeds. Copy and Edit keep the original Markdown text. Links and image embeds in saved chats resolve from the conversation note, even when another note is open.

Choose an agent

Open Settings → Copilot → Basic → Agents. Configure at least one agent, then choose the Default backend for new chats.

Agent	Best starting point	Where model access comes from
opencode	Recommended for most people	Copilot-hosted models, your API providers, or local OpenAI-compatible APIs
Claude	You already use Claude Code	Your Claude Code installation and Anthropic account
Codex	You already use the Codex CLI and Codex account	Your Codex CLI login through the codex-acp adapter

For opencode and Codex, Managed by Copilot uses the version Copilot tests. My own binary remains available, but other versions may not work correctly with Copilot. The Configure dialog reminds you of this when you choose your own binary.

A one-agent chat can work without a Copilot license when you bring your own model access. An eligible paid plan adds Copilot-hosted models and cloud-backed features. Compare Copilot plans.

opencode

opencode is the most flexible choice because it can use Copilot-hosted, BYOK, and local models.

1. In the opencode tab, open Configure, choose Managed by Copilot, then click Download & install. Progress shows elapsed time while connecting, downloaded bytes during transfer, then extraction, verification, and activation.
2. If you already installed it, choose My own binary in Configure, then select Auto-detect or enter the absolute path and click Apply.
3. Enable the models you want to see and choose a Default model.

Managed installs download and verify a pinned OpenCode release directly. You do not need to install npm, Bun, Node, or a separate archive tool. Copilot accepts your own OpenCode 2.0.21 or newer binary.

If OpenCode’s internal service stops responding, Copilot refreshes it. The failed turn explains the problem; send the message again when the agent is ready. If OpenCode cannot reopen the conversation, Copilot starts a new chat. If the service fails before it answers any chat, Copilot stops retrying and shows the error; check the Copilot log for the cause.

There are three ways to provide model access:

• Copilot-hosted: add your license under Basic → Copilot License. Models included with your plan appear automatically.
• Your API key: open Settings → Copilot → BYOK, add a provider, and configure its models. Copilot stores supported secrets in the Obsidian Keychain.
• Local: add an OpenAI-compatible endpoint from software such as Ollama or LM Studio under BYOK.

Claude

The Claude backend runs through Claude Code on your computer:

1. Open Basic → Agents → Claude → Configure.
2. Select Auto-detect, or enter the absolute path to the claude executable.
3. Select Sign in if Claude Code is not already authenticated, then finish in your browser. If the page does not open, select Open sign-in page. You can cancel and try again.
4. Enable the models you want and choose a default.

Claude Code stores the credentials on your computer. Copilot checks its sign-in status when the login finishes. To run the login command yourself, use Sign in using a terminal instead.

Claude models and billing come from your Claude Code account. Models added under BYOK do not join the Claude model list.

Codex

The Codex backend uses @agentclientprotocol/codex-acp, which includes a compatible Codex CLI:

Managed Codex downloads use the adapter version pinned by Copilot. Copilot uses the system tar -xf command to unpack the runtime on all supported desktop platforms:

Platform	Download format	Extraction
Linux	.tar.gz	System tar, compatible with GNU tar.
macOS	.zip	Built-in tar, which uses bsdtar and supports ZIP extraction.
Windows	.zip	Built-in tar.exe, which uses bsdtar and supports ZIP extraction.

The archive format and extraction command are separate: bsdtar can unpack ZIP files, while GNU tar does not support ZIP. See bsdtar’s supported formats and Windows tar documentation. Windows includes tar.exe starting with Windows 10 version 1803. If installation reports that tar is missing, install it and retry; on macOS or Windows, use bsdtar so ZIP extraction works.

For an existing or manually installed @agentclientprotocol/codex-acp adapter, Copilot checks compatibility and reports any required update in Configure. The adapter version and bundled Codex CLI version are separate.

1. Open Basic → Agents → Codex → Configure.
2. Choose Download & install under Managed by Copilot. Copilot shows elapsed time while connecting, downloaded bytes during transfer, then labels extraction, verification, and activation while it prepares Codex. It keeps your current installation until the replacement is ready. You do not need Node.js or npm.
3. Click Sign in, finish authentication in your browser, and return to Obsidian. If the browser does not open, click Open sign-in page. You can cancel or retry sign-in. Existing credentials are reused when you use the same Codex profile.
4. If you prefer your own adapter, use My own binary to Auto-detect it or enter its absolute path. You remain responsible for upgrades to that binary.
5. Enable the models you want and choose a default.

The Authentication section shows your signed-in ChatGPT email and plan when Codex provides them. Select Sign out beside the account to switch accounts. API-key authentication and accounts without an email show the signed-in badge without an email. An installed agent shows Sign in required until it is authenticated.

You can also choose Sign in on the Agent Chat status card. For terminal login, run your configured adapter with cli login using the same CODEX_HOME as Copilot.

Switching to your own Codex binary removes unused managed downloads. Your custom binary and account credentials remain on your computer. Cancel is available during downloads; configuration changes finish before another action can start.

If your adapter is below the supported minimum, select Configure in Settings to manage it. Agent Chat also offers Upgrade, with shared progress and errors. Copilot uses the login stored by the bundled Codex CLI. Models added under BYOK do not join the Codex model list.

For Windows-specific installation help, see Windows setup for Agent Chat.

Start a chat

Select the Agent Chat ribbon icon or run Open Copilot Agent Chat Window from the command palette. From an open note, select the Copilot icon beside the note’s view controls to open Agent Chat with that note attached to your next message. If Agent Chat is already open, the icon brings it into view; otherwise, it opens in the right sidebar. If the default agent is missing, its custom binary is outdated or unusable, or it is not signed in, Copilot opens Select your agent with the reason and a Configure action. You can also choose another working agent and select Start chat. If an agent is still starting, wait for that launch to finish before starting another agent. Configure remains available while you wait. If you already have a chat open, Copilot keeps the conversation and your unsent draft visible and shows a warning above the chat.

An empty Agent Chat shows a fixed hint: “Ask anything • @ to add context • / for commands”.

Agent Chat groups consecutive tool calls and reasoning into a compact activity row. The row reports the total tool commands, distinct files read or edited, and recorded reasoning time. Open it to inspect every step.

Copilot checks for updates in the background when the plugin loads, including when Obsidian starts. If a newer version is available, a notice offers View release notes once per release. This is remembered separately from dismissing the Agent Chat home banner. You can keep using or close Obsidian while the check runs. Copilot reads the version from the published release’s manifest, and Settings and Agent Chat share that check.

Copilot also checks for the latest release whenever you create a new Agent Chat tab with +, so you can discover updates without reloading the plugin. When a newer Copilot release is available, the global Agent Chat home shows an update banner along the bottom of the pane. The banner stays above the home tabs when space is tight. Select See what’s new to read the release notes, or dismiss the banner for that release. When the release has a demo video, the banner also shows its thumbnail; select it to watch the video. Project homes and active conversations do not show it.

The release notes open with the newest release. If you skipped earlier stable releases, their notes follow below it, newest first, so you can catch up on every change since your installed version. The dialog shows up to ten releases and leaves out prereleases. If GitHub’s release list can’t be loaded, only the newest release appears.

Models, effort, and permissions

Each agent has its own model list. The models shown in one agent do not automatically appear in another.

• Set the model and effort used by new chats under Settings → Copilot → Basic → Agents.
• Use the controls beside the composer to change the model or effort for the current chat.
• Before the first message, choosing a model from another installed agent switches the empty chat to that agent. Once a conversation has started, it stays with its agent.
• Effort appears only when the selected agent and model support it. Higher effort can improve difficult reasoning but may take longer and use more of your account allowance.
• Models with effort controls always use a concrete effort. A supported saved effort is kept; otherwise Copilot chooses the lowest supported effort and updates the saved default. Models without effort controls omit effort. Agent default clears the saved model preference; the new session shows the agent’s model and its resolved effort.
• Reloading an agent keeps each open chat’s model and effort. If that model is no longer enabled for the agent, or is still waiting for its API key, the chat falls back to your saved default, or to one of your enabled models.

The permission picker shows only choices supported by the current agent:

Choice	What it does
Safe	Allows edits under the agent’s approval rules
Plan	Drafts a plan and waits for your approval before editing when the current agent supports this choice
Auto	Reduces approval prompts according to the current agent’s automatic permission rule

opencode supports Safe and Auto. Claude supports Safe, Plan, and Auto. Codex shows the choices supported by the installed adapter. Claude also has an Auto mode permissions setting that controls how much Auto may approve.

With Codex, Safe uses the adapter’s Ask for approval preset. It can edit files in the working directory; the adapter asks before editing external files or using the internet. Auto uses automatic review within the same workspace-write sandbox, without enabling full access. Plan uses Auto’s approval preset while Codex drafts a plan and waits for your approval. Once you approve it, Codex continues in Auto mode to implement the plan.

When Codex presents a plan, you can Approve it, Reject it, or type feedback in the plan card. Feedback starts a new chat turn so Codex can revise the plan; review the new proposal before approving it. If you ask Codex to save the plan to a note, approve the revised plan to let it create the note.

When an action needs approval, Agent Chat displays a Permission required card with the proposed change or tool input. Choose one of the temporary or persistent allow or deny options offered by that agent. Stopping the turn cancels unanswered requests.

When Claude or Codex asks a set of questions, answer the current tab and select Next. On the final tab, Submit becomes available after every question has an answer. You can use the tabs to review or skip ahead. Cancel dismisses the entire request. If Codex withdraws or times out the question, its card disappears.

Codex can ask questions while the permission picker is on Safe, Plan, or Auto. You do not need to switch to Plan to answer a question.

OpenCode does not show question cards yet. If it asks a clarifying question in chat, answer in the message box.

Permission and question cards stay in a scrolling action area above the message box until you answer them, even while the response continues streaming.

Your vault or project is the agent’s working directory, not a security sandbox. Auto or bypass permissions can reach other files and services available to the agent or your account. Use Safe for unfamiliar work and review persistent permissions carefully.

Context and history

Agent Chat keeps each conversation separate:

• Select + for another session. Each tab keeps its own history, draft, attachments, and queued follow-ups.
• Select New Chat to reset the current tab.
• Select Stop to cancel the current turn. Queued follow-ups return to the message box, with their notes, web pages, agent mentions, and images, so you can edit or resend them. Queued content comes before the current draft, and custom commands return as their expanded text.
• Use Recent Chats from the Agent Chat home screen, or Chat History inside a conversation, to resume saved work. The Recent Chats list can be scrolled or searched.
• Select Copy Chat Link beside New Chat to copy an obsidian:// link to the current conversation. Paste it into another note to reopen the chat in the same vault. The link points to the conversation’s Markdown note, so Copilot saves the note first if it does not exist yet, including when autosave is off. The link keeps working after the note is renamed.
• A spinner marks a chat while its agent is responding. When the response stops but the session remains live, a small blue dot appears on the bottom-right corner of the chat icon. Hover the row and select Close session to release that session. Saved history stays available to reopen, and other chats keep running. If the agent does not support closing individual sessions or release fails, the chat stays marked live with an error.
• Add the active note, selected text, other notes, folders, a Copilot Web Viewer tab, or supported images. You can also mention a note with [[Note title]].
• Hover the context ring beside the send controls to see how much of the model’s context window is in use. The ring stays empty until the agent reports usage, and a stopped response keeps the last reported reading. If the connected account reports usage limits, the same panel shows the available limit and reset time.

You can send an image without adding text, for example when the agent asks for a screenshot. Choose a model that supports images.

Attachments apply to the next message. For instructions and context that should be reused, create a Project or add rules to AGENTS.md. See Context and Mentions for every context option.

Uploaded images are embedded in saved conversation notes. Copilot stores the image files under <Copilot folder>/copilot-conversations/attachments/, not in your vault attachment folder, and saves each unique image once per conversation. Identical images in the same conversation reuse the saved attachment, including when it has been deleted. Later saves preserve attachment links updated by Obsidian or an attachment organizer; deleting a saved attachment leaves a missing image instead of recreating the file.

Type / to insert an enabled Skill or Copilot command. For a quick question or rewrite beside the current selection, use Quick Ask.

Multi-agent answers

With active Plus access, type @, open Agents, and mention one or more other installed agents in the same prompt. Copilot sends the same question, that turn’s attachments, and a bounded slice of the visible conversation to each mentioned agent. With one other agent mentioned, only that agent responds and its answer is shown directly. With two or more, they answer in parallel and the current agent summarizes their answers without producing a separate answer of its own.

This is useful for research, second opinions, and reviews. Mentioning only the current agent behaves like a normal turn.

Each answer appears in its own tab. A direct one-agent turn shows only that agent. A turn with two or more answerers opens with Summary first. If one answerer fails in a multi-agent turn, Copilot keeps the successful answers and summarizes what completed.

Multi-agent answers are designed for read-only research, not edits. Copilot denies explicit vault edit, delete, and move tools, along with tools it cannot classify. Retrieval Skills can still run their own scripts under the agent’s permissions, so multi-agent answers are not a security sandbox. Use only trusted Skills, and use a regular single-agent turn when you want files changed.

The default model and effort saved for each mentioned agent are used for its answer. If a multi-agent summary fails, its tab shows the error and any partial summary; the individual answers remain available. If an agent is not installed or ready, configure it before adding it to the prompt.

Skills across agents

Skills are reusable instruction packets built around a SKILL.md file. One Skill can be made available to opencode, Claude, and Codex without maintaining three copies.

1. Open Settings → Copilot → Skills.
2. Find a Skill and toggle the opencode, Claude, or Codex icons for the agents that should use it.
3. Type / in Agent Chat to choose it, or describe the task and let the agent select an enabled Skill.

Shared Skills live under <Copilot folder>/skills/. Copilot links them into the native folders used by each agent: .opencode/skills/, .claude/skills/, and .agents/skills/. Skills already present in those native folders also appear in the settings list.

The Built-in Skills table contains Copilot’s bundled skills with read-only previews. Turn off a whole skill to remove its generated files, or turn off an individual agent to exclude it. These choices survive restarts and updates. Copilot generates built-in skills only when an available agent can use them. Setting up another agent enables eligible built-ins for that agent unless you previously opted out. Your own skills remain editable in Your Skills.

Custom Skills and built-in Obsidian Skills are free. Active Plus access adds cloud-backed Skills for web research, PDF reading, YouTube transcripts, X posts, and OpenArtifacts.

The web search, web fetch, and PDF reading Skills (copilot-web-search, copilot-web-fetch, and copilot-read-pdf) start turned off in new installs, so agents use their own web and PDF tools. Turn them on in Built-in Skills to route those requests through Copilot instead.

Built-in skill preferences record only the skills and agents that are turned off. New skills use the defaults, and content updates keep your saved choices. Copilot fully manages built-in skill folders. Updating a built-in skill replaces its entire folder; removing it deletes the folder. Both operations remove any files you added inside it. Keep custom skills and files in separate folders. When a skill is retired, Copilot also clears its saved preferences. A renamed skill starts with fresh defaults. If file permissions prevent removal, Copilot reports the failure and keeps your disabled preference. After you restore access, Copilot retries cleanup when skills refresh or the plugin reloads.

In Self-Host Mode with OpenCode selected, Agent Chat’s built-in web-search Skill uses the search provider selected under Settings → Copilot → Self-Host. Provider credentials stay inside Obsidian rather than being passed to OpenCode, and the feature does not require Obsidian’s command line interface. Copilot disables OpenCode’s native web-search and web-fetch tools so they cannot bypass that route. Full-page web fetching is unavailable through OpenCode in Self-Host Mode because the supported search providers do not share a page-fetch interface; Agent Chat can still use the configured provider’s search results. If the web-search Skill is turned off, turn on copilot-web-search in Built-in Skills to give OpenCode web search in Self-Host Mode.

If you set a custom OPENCODE_CONFIG_CONTENT for OpenCode, Self-Host Mode adds final deny rules for the native web tools in its top-level and per-agent permissions. Other permission rules in your custom configuration remain in their original order. Self-Host Mode does not start OpenCode with a custom configuration that uses the OpenCode 1 permission, tools, agent, or mode keys; write those settings as OpenCode 2 permissions and agents instead.

On Windows, creating the folder links may require Developer Mode or administrator access. If a sync service replaces a link, toggle that Skill off and on for the affected agent to recreate it.

Applying a settings change

Agents read their configuration when they start: API keys, enabled models, instructions, and skills. A change you make while a chat is open does not reach the agent already running. Agent Chat says so with a config has changed notice and a Reload action, and the chat keeps working in the meantime.

The notice names the agent that needs to reload. Changing Claude’s enabled models does not ask you to reload OpenCode. Settings an agent can apply without restarting do not need a reload notice.

Choose Reload when you want the change to take effect. The agent restarts and reopens every chat using that agent, keeping each chat’s history and anything you had typed but not sent. During a running turn the reload waits for that turn to finish. If the agent cannot reopen a conversation, that tab starts a fresh chat and the old one remains under Recent Chats. If a chat failed to start, correct its settings and choose Retry to apply them.

Two changes never wait. Narrowing Miyo’s Search scope and turning on Self-Host Mode both restart the agent straight away, so no later step in the conversation can search or browse outside the new boundary.

Related

• Getting Started
• Models, Effort, and Permissions
• Projects
• Instructions for Agent Chat and Quick Chat
• Copilot Commands and Quick Ask
• Copilot Plans, Privacy, and Self-Hosting

Publishing from Agent Chat

Copilot ships an openartifacts-publish skill to Claude Code, Codex, and OpenCode. Publishing needs a Copilot Plus license key in Copilot settings. Copilot passes that key to the agent process, and the skill’s bundled wrapper script sends it to api.openartifacts.ai over HTTPS. The wrapper needs only the shell tools already on your system: sh, curl, and awk on macOS and Linux, PowerShell on Windows.

Ask the agent to publish a Markdown note or an existing HTML file. For Markdown, it renders the note to HTML, using the note’s file name without .md, converted to title case, as both the browser title and a visible heading above the note body. If the note already starts with the same heading, Copilot keeps that heading instead of adding a duplicate. If you explicitly ask for no title, Copilot does not add the visible heading but keeps the browser title. Long titles wrap on narrow screens instead of being cut off. Copilot writes the complete page under .openartifacts/handoffs/ in your vault, tells you the path, and stops. For HTML, it preserves the existing file and passes that exact file to OpenArtifacts without converting, rewriting, copying, or deleting it; because only that file is uploaded, the agent warns before publishing if any relative local resource in an HTML attribute or CSS URL will not load. Open the HTML file in your browser to see the page as it will be uploaded; OpenArtifacts adds its own header and footer bylines when it serves the page. Reply that it should publish, or describe changes and the agent revises the generated Markdown handoff or waits for you to update and review the original HTML again. The agent never publishes in the same turn that produced the HTML.

After publishing a Markdown note, the agent saves the public link in the note’s openartifacts property, so publishing the note again updates the same page and the regular Publish file to OpenArtifacts command recognizes it as published. Notes published under the older symposium property keep working and move to openartifacts the next time they are published. If a property already holds something other than an OpenArtifacts link, the agent asks before touching it. Ask the agent to withdraw a page to take it down; it confirms first and removes only note properties whose links match the withdrawn page afterwards. For HTML, keep and supply the returned OpenArtifacts link when asking the agent to withdraw the page later; republishing the HTML creates a new link.

Copilot also includes research-memo as an optional theme. For a named theme, the skill checks .openartifacts/themes/<name>.md in your vault, then themes/<name>.md alongside the installed skill. If neither exists, the agent continues with readable defaults. A missing theme never blocks publishing.

Upgrading an agent

When an installed agent needs a supported version, select Configure in Basic → Agents. Agent Chat offers Upgrade if Copilot can upgrade that installation. Both surfaces show the same progress and failures, including upgrades started in Configure. If an upgrade fails, retry in Configure or use Retry in Agent Chat. A failed custom-path selection is reported in Configure and does not turn the upgrade action into a path-validation retry.

Image-only messages appear as “Image attachment” in the queue. Sessions with only images use the same fallback title in tabs and Recent Chats until a text or agent-generated title is available.

Managed runtime updates

When Copilot loads, managed Codex and OpenCode runtimes update to the version shipped with Copilot before that agent starts. Agent Chat shows download progress, then opens a new chat when the agent is ready. Other agents remain usable during the download. A failed download leaves the previous supported runtime available and retries on a later load after 24 hours. Configure lets you retry immediately or select your own binary, which Copilot does not update automatically.

Codex and OpenCode show Upgrade required when the installed runtime is below the minimum supported by Copilot. An unsuccessful update keeps that installation unavailable until you upgrade it; only an older runtime that still meets the minimum remains usable. Configure lets you retry immediately. A new minimum requirement permits a fresh automatic update attempt even if a previous attempt was postponed. Custom binaries require a manual update.