Skip to content

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.

Selecting a note in Agent Chat, such as a file on a tool card, a drawing on a Copilot card, a link in a reply, or Open note in a change review, switches to the tab where that note is already open, including one in a popout window. A new tab opens only when the note isn’t open. To open another tab anyway, Cmd-click (Ctrl-click on Windows and Linux) or middle-click it.

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

AgentBest starting pointWhere model access comes from
OpenCodeRecommended for most peopleCopilot-hosted models, your API providers, or local OpenAI-compatible APIs
ClaudeYou already use Claude CodeYour Claude Code installation and Anthropic account
CodexYou already use the Codex CLI and Codex accountYour Codex CLI login through the codex-acp adapter

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

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 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 installation 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.

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.

Claude Code saves what it learns about you and your vault in memory files outside the vault and reads them in later chats. Auto memory in Claude’s settings tab is Default, which follows your Claude Code setting. Choose On or Off to override it for this vault’s new chats. Once Claude has saved memory for the vault, select Open under Memory files to review it. Deleting a chat doesn’t remove what Claude saved from it.

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:

PlatformDownload formatExtraction
Linux.tar.gzSystem tar, compatible with GNU tar.
macOS.zipBuilt-in tar, which uses bsdtar and supports ZIP extraction.
Windows.zipBuilt-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 installation 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.

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.

After execution, OpenCode steps keep Execute and list the connector tools OpenCode reports, such as Execute · Notion · Search. A block that runs several tools lists each distinct tool once. Approval cards and steps without reported tool names show Execute; expand their details to inspect the code.

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.

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 chat to that agent. A chat opened with Continue from here in a new chat keeps its earlier messages, title and chat note when it switches. Once you send a message, the chat 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 chat 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:

ChoiceWhat it does
SafeAllows edits under the agent’s approval rules
PlanDrafts a plan and waits for your approval before editing when the current agent supports this choice
AutoReduces 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. The one-time allow is highlighted. Hover an “always” option to see what it covers and how long it lasts: Claude names its rule and the settings file it saves to, and Codex and OpenCode say when they save a rule for later chats. Claude offers “Allow always” only when it has a rule to save. To review or remove the rules Claude, Codex, or OpenCode saved, open that agent’s tab in Copilot settings and select Edit approved rules. A Codex chat that is already open keeps a removed rule until you start a new chat. Stop cancels the agent’s unanswered requests. A running subagent’s requests stay open.

Approved rules can be shared with other chats using the same agent settings file or project. For file-backed rules, a removed rule stays on the approved-rules page, crossed out, with Undo until you leave the page. A restoration preserves other changes made since removal. If the file cannot be safely updated, the rule’s row shows an error and keeps Undo available to retry.

When Claude or Codex asks a set of questions, answer the current tab and select Next. Clicking an answer to a pick-one question moves to the next tab for you. 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.

Agent Chat keeps each conversation separate:

  • Select + for another chat. Each tab keeps its own history, draft, attachments, and queued follow-ups.
  • Select New chat to reset the current tab.
  • Copilot Navigator is a preview for Supporters; without preview access, its icon and the Open Copilot Navigator command open a dialog about it. If you are a Supporter, select Already a Supporter? Enter your license key and enter your key in Copilot settings; Copilot Navigator opens once your access is active. If your preview access ends, Copilot Navigator offers See plans instead of your chats, which stay under Recent chats in Agent Chat. Select the Copilot Navigator icon at the left end of the tabs, or run Open Copilot Navigator from the command palette, to list your saved chats in Copilot Navigator in the left sidebar. While the sidebar is open the tabs are hidden, and closing it brings them back. Select a chat to open it; the open chat is highlighted. A chat’s status shows at the right end of its row: a spinner marks a chat that is responding, an alert icon marks one waiting for your permission, and a green dot marks one that is Running. A bold title marks a chat with a result you have not seen. Leaving a chat before you send a message in it shuts it down, unless you typed text or attached a note in it. A chat you sent a message in keeps running until you shut it down, so its changes and an unseen result are still there when you come back. Select New chat at the top to start a new global chat; it is listed under Recents once you send its first message.
  • The sidebar groups chats into Pinned, Projects, and Recents. A pinned chat is listed only under Pinned, with its Project’s name beside it. Each Project is a folder that opens to show its five newest chats; select Show more to list the rest of the Projects or of a Project’s chats. Recents lists chats outside any Project and loads more as you scroll. Projects and chats are listed newest first and keep their place when you open them. Select a group’s name to fold or unfold it. Select a Project to open its home; to fold or unfold its chats, hover it and select the arrow at the right end of its row. Folds stay after you reload Obsidian. Hover a Project and select + to start a chat in it. Select + beside Projects, type a name, and press Enter to create a Project; press Escape to cancel. A folder shows the most urgent status of the chats it hides: waiting for permission, then an unseen result, then responding.
  • Hover a chat in the sidebar to swap its status for Pin chat, Archive chat, and …. Select … (or right-click) for its other actions: rename, delete, open the chat note, and Shut down for a running chat. Hover a Project and select … (or right-click) to reveal, edit, archive, or delete it. Select the search icon to filter chats by title across every group. Select the … at the top and choose View archived to see archived chats and Projects; they are not listed in the sidebar.
  • While the sidebar is open, the Agent Chat start page shows only the greeting, the message box, and Relevant Notes, and a Project’s start page shows only its Files. To leave a Project, select another chat or New chat in the sidebar. Closing the sidebar brings back the Recent chats and Projects lists.
  • Select Stop to cancel the current turn. Messages the agent has not read yet return to the message box, with their notes, web pages, agent mentions, and images, so you can edit or resend them. They come before the current draft, and custom commands return as their expanded text. Subagents and background commands keep running; stop each one from the running-tasks list.
  • 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. A long chat opens at its latest messages; scroll up to load earlier ones.
  • Hover a chat in Recent chats and select Pin chat to keep it at the top of the list, above a divider. The most recently pinned chat comes first, and pinned chats keep their order when you use them or change how history is sorted. Select Unpin chat to return it to its usual place. The pin is saved in the chat’s Markdown note, so it lasts across restarts and appears in both the global and the Project list. Pinning a chat that has no note yet saves it as a note first, even when autosave is off. Chat history inside a conversation ignores pins.
  • The bar above the composer holds New chat, Save chat as note (when autosave is off), Chat history, and More actions (⋯), and starts with a task count, such as 2 tasks running, while background tasks run. More actions has Copy chat link, Add to bookmarks, and Report an issue. Copy chat link and Add to bookmarks stay dimmed until you send the chat’s first message.
  • Hover a chat in Recent chats, select More actions, and choose Archive chat to put it away without deleting it. Archived chats disappear from Recent chats, Chat history and the Project list, even when pinned. Select View archived below the list to unarchive a chat or open it. Sending a message in an archived chat unarchives it.
  • To archive or delete several chats at once in Recent chats or Copilot Navigator, Shift-click to select every chat between two chats, or Cmd-click (Ctrl-click on Windows) to add or remove one chat. Right-click a selected chat and choose Archive N chats or Delete N chats, then confirm. Click a chat or press Esc to clear the selection.
  • To bookmark the open chat in Obsidian’s Bookmarks pane, select More actions above the composer and choose Add to bookmarks. To bookmark a chat without opening it, hover it in Recent chats, select More actions (or right-click the row), and choose Add to bookmarks. The bookmark uses the chat’s title, and clicking it opens the chat in Agent Chat. A chat that has no note yet is saved as one first, even when autosave is off. Choose Remove from bookmarks to remove it. Deleting the chat from Recent chats also removes its bookmark, while renaming the chat keeps the bookmark’s title. While the Bookmarks core plugin is turned off, Add to bookmarks is dimmed, with an info icon explaining why. Selecting it opens Settings → Core plugins filtered to Bookmarks so you can turn it on.
  • Select More actions above the composer and choose Report an issue to report a problem with the current chat, with its screenshot and agent details ready. See Reporting from Agent Chat.
  • Select More actions above the composer and choose Copy chat link 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. To copy a link without opening the chat, hover it in Recent chats, select More actions, and choose Copy chat link. A chat that has no note yet is saved as one first.
  • To branch off from an earlier point, hover a reply and select Continue from here in a new chat. Copilot saves the chat as a note, saves the conversation up to that reply as a new chat note, even when autosave is off, and opens it in a new tab with the same title, agent, model and Project. The new chat shows that conversation as plain messages with their attached notes and images, without tool steps, and your first message there passes it to the agent, along with an instruction not to read the original chat’s note, which holds the later messages. The original chat stays as it was. The option is hidden while the chat is responding.
  • Properties you add to an Agent chat note stay when Copilot saves the chat again. Copilot updates only the properties it manages, such as backendId, sessionId, projectId, model, usage, carriedMessages, continuedFrom, and tags.
  • A spinner marks a chat while its agent is responding or one of its background tasks runs. Closing a chat’s tab keeps it running once you have sent a message in it; a chat you leave or close without sending a message shuts down. When the response stops but the chat is still running, a small blue dot appears on the bottom-right corner of the chat icon. Hover the row and select Shut down to stop it. Saved history stays available to reopen, and other chats keep running. If the agent can’t shut down a single chat, or shutting down fails, the chat stays marked running with an error.
  • Deleting an Agent chat removes Copilot’s chat note and its entry in Copilot’s chat history. Claude Code, Codex, and OpenCode may keep their own copy of the conversation on your computer. Copilot does not delete or change those files, so you can still resume the chat in that agent’s own tool. The confirmation and the notice after the delete both say this. Deleting an open chat closes it and stops any reply in progress.
  • 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.

When a conversation grows long, choose /compact at the start of an empty message to have the agent summarize it and free up room in its context window. The / menu offers /compact only at the start of a message, and it replaces any Skill or command named exactly compact. Anything you type after /compact is sent with it, so an agent that accepts compaction instructions, such as Claude, can use them.

  • Attached notes, web pages, images, and agent mentions are not sent with /compact. Project context waits for your next regular message.
  • A queued /compact runs as its own turn instead of joining other queued messages. If you stop the current turn, the other queued messages return to the message box without it.

Agents also compact on their own when the context window fills up. Either way, the chat shows the progress in line with the conversation:

StatusMeaning
Compacting conversation…The agent is summarizing the chat after /compact
Context is full, compacting…The agent started compacting on its own
Context compacted or Context compacted automaticallyThe summary is done. When the agent reports them, the token counts before and after follow the label
Couldn’t compactCompaction failed. The reason appears with the row or below it
Compaction cancelledYou stopped the turn before compaction finished
Nothing to compact yetThe chat is too short to summarize

When a finished compaction reports its new size, the context ring drops to that size instead of keeping the reading from before it.

Reopening a Codex or OpenCode chat from chat history shows its past compactions again, each in the place it happened. A compaction that never finished shows as Compaction cancelled. Reopening does not change the context ring.

A model can only write so much in one reply. When a reply runs out of room partway through a step, such as writing a long note, that step never runs. Its row shows as failed instead of spinning, and Copilot asks the agent once, in the same reply, to continue in smaller steps. Copilot never continues a turn you stopped or one the model refused, and it continues only once per message.

If the reply still stops early, it ends with a Stopped early notice that names the cause: the model hit its output limit, the agent hit its step limit, or the model’s safety system stopped it. The last step may not have run. Send another message to continue.

A step cut short by Stop shows as stopped. Any other step the agent left unfinished when its turn ended shows as failed, including in chats you reopen from history.

You can type and send while the agent works. The button beside the message box is Stop while the agent’s reply runs and the box is empty. Otherwise it is Send. With Claude, the message waits above the message box under The agent reads these at its next step, then moves into the chat when the agent reads it. Select Withdraw queued message to take it back into the message box, or Send now to stop the agent’s current reply so it reads every waiting message right away. Background subagents and commands keep running. With Codex and OpenCode, the message waits with Sends when the agent finishes.

Claude runs subagents and long commands in the background unless it needs to wait on one, so you can keep chatting. A task appears as soon as Claude starts launching it. When one finishes, Claude replies with its result on its own. The chat’s spinner stays on until every subagent has finished and Claude has replied about it.

  • Hover the task count above the composer, such as 2 tasks running, to see each running task.
  • Select a task, or its card in the chat, to open its view: a subagent’s steps or a command’s output. Select Back to chat or press Escape to return.
  • Select Stop task on a task to end only that task. It turns on once the task starts.

When a turn finishes after the agent has edited your vault, a Files changed card closes that turn. It lists one row per note, with the number of added and removed lines, and a new or deleted badge for notes the agent created or removed. The card totals the changes in its header and scrolls when a turn touched many files. A subagent’s edits show on a card in that subagent’s view instead.

Select a row to open that note’s before and after as a single tab. Removed text is struck through, added text is highlighted, and everything else reads as the note normally does, including the blank lines between paragraphs, so you can see the exact wording the agent changed without leaving Obsidian. Use Open note in its header to open the note itself and edit it there. Files that are not notes, such as canvases, are shown as a plain line-by-line comparison.

To undo part of a turn, hover over a stretch of changes and select Revoke. That stretch goes back to what the note said before the turn, and the rest of the note stays as it is. Revoke is unavailable for a stretch that you or a later turn edited after this turn, and for notes the agent created or deleted. The agent is not told about a revoke.

To undo everything the turn did to the note, select Revoke all in the tab’s header. An edited note goes back to its text from before the turn, a note the agent created is deleted according to your Files and links > Deleted files setting, and a note the agent deleted is recreated. Revoke all is unavailable once you or a later turn edits the note, or, for a deleted note, once another file takes its place.

Select text in the tab to attach it to your next message. Removed and added words in the selection stay marked, so a question such as “why did you change this?” is about exactly that change.

In a long note, a narrow strip along the left edge of the tab shows where the changes are. The strip stands for the whole note, with one mark for each change at its position in the note: green where text was added, red where it was removed, and blue where both happened. Select a mark to jump to that change.

The header shows which change you are at, such as “1 of 6”. Use the up and down arrows beside it to move to the previous or next change; moving past the last change returns to the first. The Go to next change in file diff and Go to previous change in file diff commands do the same, so you can assign them hotkeys in Obsidian’s settings.

Selecting the same row again returns you to the tab that is already open. This review lasts until you close Obsidian or reload the plugin, which drops the cards and the diff tabs. Earlier turns show no card when you reopen a saved chat.

In a Project chat, the Project info menu beside the Project name lists every file the chat has changed so far. See See what a chat changed.

The card covers the notes an agent changed with its editing tools. Changes an agent makes by running a shell command, such as deleting a note with rm, are not listed. Treat the card as a review aid rather than a complete audit of the turn.

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, YouTube transcripts, and X posts. When Plus is inactive or no license key is configured, Copilot automatically makes these Skills unavailable and removes their managed files and agent links. Settings shows why they are unavailable. Your saved choices are kept and apply again after Plus is activated.

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

In new installs, obsidian-markdown starts turned on for every agent, obsidian-bases for Codex and OpenCode, and json-canvas for OpenCode. Turn them on for other agents in Built-in skills.

While Miyo document processing is on, which is the default, agents read PDF and EPUB files through Copilot’s Miyo parse tool, which asks the connected Miyo to parse them and does not need Plus access. The tool is available only while Miyo is connected. With a remote Miyo server, files in the vault parse on that server, and files outside the vault can’t be parsed. If Miyo cannot parse a file, the agent tells you and reads the file with its own tools. With Miyo document processing off, agents read documents with their own tools.

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.

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.

While Command line interface is on in Obsidian’s Settings → General, Claude Code, Codex, and OpenCode reach the running Obsidian app through Copilot’s Obsidian CLI tool. Ask about open tabs, daily notes, typed properties, tasks, backlinks and unresolved links, or Bases results, or create a note from a template and rename notes with their links updated. There is no skill to turn on. Depending on the permission mode, the agent asks before each command. The tool refuses to reload Obsidian or turn off Copilot, because that would end the chat.

While the Excalidraw plugin (2.29 or later) is enabled, Claude Code, Codex, and OpenCode read, draw, and edit Excalidraw drawings through Copilot’s Excalidraw tools. The Excalidraw plugin builds each drawing, so labels stay inside their shapes and arrows stay attached when you move a shape. There is no skill to turn on.

  • Ask for a diagram, flowchart or chart, for example “Draw a diagram of the steps in [[Release process]]”. The agent draws it in Excalidraw and embeds it in the note it belongs to. Ask for a Mermaid diagram or a canvas to get that instead. Creating a drawing never overwrites an existing file.
  • Select shapes in an open drawing to point the agent at them, for example “make these red”. The selection is attached to your next message, like selected text. Clicking empty canvas removes it.
  • When the agent edits a drawing, Obsidian opens it in a tab and saves each change.
  • After drawing or editing, the agent looks at a picture of the result and fixes overlaps before it answers.
  • In the chat, a Copilot card marks each drawing the agent draws, edits or reads. Click the drawing’s name to open it.
  • Codex asks you to approve each change to a drawing unless the chat is in Auto.

Connect services such as Notion, Linear, and Todoist under Settings → Copilot → Connectors, and Claude Code, Codex, and OpenCode can use them in your chats. See Connectors. OpenCode calls connected tools individually, so each step and permission request names its tool. Native OpenCode code execution still uses Execute.

Agents read their configuration when they start: API keys, enabled models, instructions, skills, and connectors. A change you make while a chat is open does not reach the agent already running. Agent Chat says so with a settings 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.

Turning on Self-Host Mode never waits: it restarts the agent straight away, so no later step in the conversation can browse outside the new boundary. Changes to Miyo’s Search scope, the Miyo folders ticked under Current vault, Maximum results and Search variants apply to the agent’s next search, with no reload.

Claude Code, Codex, and OpenCode publish through Copilot’s own OpenArtifacts tools, which send your Copilot Plus license key to api.openartifacts.ai over HTTPS. There is nothing to install and no skill to turn on. Every chat has the tools; without Copilot Plus, the agent tells you that OpenArtifacts comes with Copilot Plus and gives you an upgrade link, and nothing is published.

Ask the agent to publish a Markdown note or an existing HTML file. For Markdown, the agent renders the note to HTML and publishes it in the same turn, without writing a file. For an HTML file, it publishes that file’s contents; only that page is uploaded, so relative local images or styles in it will not load. A page the agent creates is always private: only you, and anyone you name, can open it. Copilot opens a new page in your browser, already signed in, and to make it open to anyone with the link, choose Anyone with the link in Share on the page. Updating a page that is already published keeps who can open it.

After publishing a Markdown note, Copilot saves the 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. Ask the agent to withdraw a page to take it down; it confirms first, and Copilot removes the note’s link to that page. For HTML, keep the returned OpenArtifacts link to update or withdraw the page later; publishing the HTML file again creates a new page.

Earlier versions shipped an openartifacts-publish skill. Updating Copilot removes it.

Private sharing is included with every plan that can publish to OpenArtifacts: Plus and Supporter, including existing Believer licenses. If your plan cannot publish, nothing is published, public or private, and you get an upgrade link. If your plan lapses after you share a page, private pages stay private and invited people keep access. You can still see who has access, remove people, decline requests, and make a page public, but making a page private or adding a person tells you private sharing needs a paid plan, and nothing is changed.

When you name people, the agent adds each address after the private publish. Only you and those people can open it, and they sign in with the address that was invited. Open in the Share with OpenArtifacts dialog opens your page without signing in. If the server does not support private documents, or your plan does not include private sharing, the agent tells you and publishes nothing.

You can also ask the agent later to add or remove a person, show who has access and who has asked to open it, or open one of your pages in your browser. It changes one address at a time, and approves or declines a request only when you say so.

People you add can comment on the page by default. Ask the agent to let someone only read it, and it adds them as a viewer; ask again to let them comment. A public link never lets a stranger comment.

Ask the agent to address the comments on a page. It lists the open threads, revises the note so each comment is addressed, publishes the update, then replies to each thread it addressed, naming the new version, and resolves it. Threads that ask a question or need your decision stay open, and the agent raises them with you. Replies and resolutions show as your agent.

The Publish file to OpenArtifacts command opens the Share with OpenArtifacts dialog.

  • Link access chooses Anyone with the link or Only those invited. Choose it before the first publish, or change it later from a published note’s dialog. To invite people or answer requests, use the Share button on the published page.
  • Copy link (Copy right after you publish or update) copies the page link, and Open opens the page for you without a sign-in.

Without a Copilot Plus license key, the dialog opens on OpenArtifacts access required with an Upgrade link. If your key cannot publish to OpenArtifacts, publishing shows the same screen with OpenArtifacts’ reason, and nothing is published, whichever Link access you chose. If your plan has lapsed, choosing Only those invited shows the plan message with an Upgrade link and changes nothing; Anyone with the link keeps working. If OpenArtifacts does not support private documents yet, the dialog says so, and Update, Delete, Copy link, and Open keep working.

If your Copilot license is not linked to an OpenArtifacts account, the first new page you publish, from the dialog or the agent, adds one line offering to link it, with a Link account link. Linking gets you an email when people you share with comment or ask for access. Copilot shows this once per vault, whichever way you publish, and records that in .openartifacts/link-offer-shown; delete that file to see the offer again.

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. Chats with only images use the same fallback title in tabs and Recent chats until a text or agent-generated title is available.

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 Update 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.