Agent Console

by Vinod Panicker
5
4
3
2
1
Score: 63/100

Description

Finally put your second brain to work. Agent Console connects your Obsidian vault to the best AI models – Claude Code, Codex, Gemini CLI, Kiro CLI, or any custom ACP agent. Point them at the notes that matter and build alongside them: docs, slides, projects, even code. Each project leaves your notes richer for the next.

Reviews

No reviews yet.

Stats

19
stars
4,516
downloads
3
forks
167
days
3
days
3
days
414
total PRs
1
open PRs
3
closed PRs
410
merged PRs
9
total issues
0
open issues
9
closed issues
1,514
commits

Latest Version

3 days ago

Changelog

Agent Console 2.5.0 starts your saved prompts from a link or a browser bookmark, gives every chat a link you can paste into your notes, and shows which chat needs you at a glance. It also runs Kiro CLI on its newer V3 engine.

What’s new

  • Start a prompt from an email, a doc or your browser – an obsidian://agent-console link opens a new tab with one of your saved prompts in the message box, ready to send. Copy link on a prompt chip makes one for you. Add the bookmarklet from the guide, and one click hands a web page’s selected text, address and title to a prompt. A link can only start a prompt that’s already in your vault, and it waits for you to press Enter unless it names only the prompt and you’ve ticked that prompt’s uri send box. The text a link hands over is cleaned of invisible characters, such as zero-width and tag characters, so the text you read before sending is the text the agent gets. (#323, #387, #407)
  • Link to any chat from your notes – right-click a tab and choose Copy link to chat to copy a Markdown link to it, which uses an obsidian://agent-console-chat address. Paste it into a meeting note, a to-do in your task app or a project page. One click later switches to that chat, or reopens it if you closed it, even after a restart. A link opens only chats in this vault’s history and never sends a message. See Chat links. (#406)
  • Your Kiro agents load fully, on Kiro CLI’s V3 engine – new Kiro chats start on V3, where Kiro agent configs that need it load and their hooks run in Agent Console chats. The notes and files you attach reach the agent. On an older Kiro CLI, chats keep working on V2, and a tip lists what kiro-cli update adds. If a chat won’t start on V3, set the agent’s Kiro CLI engine to V2 (legacy). Upgrading? On V3, --trust-all-tools and --trust-tools no longer apply, and Kiro’s own permission rules decide what runs without asking. --agent, --model and --effort still work. To stop being asked about a tool, choose Always allow for this vault. To allow everything, turn on Auto-allow permissions (Settings → Permissions). See the Kiro CLI setup guide. (#312, #315, #337, #341, #344, #373)
  • See which chat needs you – a tab shows a tick when a reply finished while you were away, a ring with ! when a turn stopped early, and a dashed ring while it connects. The tick and the ! ring clear when you look at the tab. A speech bubble with ! shows while a failed message’s error is on screen, and stays until you close the error or send again. The icons are sharper and keep their color in the tab list. (#333, #342, #352, #375)
  • See every chat’s status in a slim column – a new Tab display setting (Settings → Tabs) moves the tabs into a narrow column of status icons on the right edge of the chat view. Hover a tab for its name and status, or drag the edge to show names. When names show, hover a tab for its ×. Delete closes the focused tab at any width. New installs start on it, and existing installs keep the tabs on top. A one-time “What’s new” card lets you try the column in one click, and switch back just as fast. See Vertical tabs. (#361, #363, #364, #368, #372)
  • Idle tabs sleep to save memory – a background tab with nothing going on frees its agent after an hour. An agent with its tools can hold several hundred megabytes, so a day with many tabs no longer piles up gigabytes. A sleeping tab shows a moon and keeps its conversation on screen. Type or send and it picks up the same chat, with the model, effort and mode it had before it slept. (#331, #360, #393)
  • Tell whether an answer is 30 seconds or 30 minutes old – each message shows when you sent it, and each reply shows when it finished and how long the agent worked, such as 9:18 AM · 6m. Time spent waiting on your approval is left out. While a reply runs, the line under the chat reads like Working 4m · last update 10s ago, or names the tool that’s running, and each finished tool call shows how long it ran. Hover a time to see how long ago it was. The times are facts: Agent Console doesn’t guess whether an agent is stuck. See Times in the chat. (#400, #404, #412)
  • Each prompt runs on the agent and mode it needs – add agent: and mode: to a prompt note, and the prompt runs with that agent and mode. When your current tab uses something else, the prompt opens a new tab instead of changing yours. (#321)
  • Know which chat you’re in, even with icon-only tabs – the chat header shows the chat’s title. After the title, in gray, it adds only what the message box doesn’t already show, such as the agent’s name on Claude Code. “[Agent Console]” is gone from the header, and the tab’s icon shows whether the chat is connected. If the agent you chose couldn’t load, its name shows struck through, then the one that’s running. Hover the header for the full title, agent, engine, model, effort, mode and folder. (#312, #326, #347, #397)
  • Jump to the start of a long reply – when the agent finishes a long reply, an up arrow at the top of the chat takes you to where its final response starts, so you can read it from the top. The Jump to start of latest reply command does the same, even while the agent is still writing. See Reading long replies. (#403)
  • Lighter on your computer while agents work – the dot grid that showed an agent at work is now the working face, the Agent Console icon, animated. It blinks and glances while the agent works and smiles when the reply is done. It costs about as much as an idle chat, around 3 ms of processing per second, down from 76–92 ms for the dots. Older messages also no longer redraw each time a reply starts or ends. (#414, #416)
  • Attach files from a button – a paperclip above the send button opens your file picker. Picked files work like dropped ones: images go in as images when the agent can see them, and as files otherwise. (#384)
  • Use Antigravity CLI out of the box – Google’s successor to Gemini CLI is now a built-in agent. It appears in the first-run panel and in every agent list, with its own settings page and setup guide. The Gemini CLI setup guide is now marked legacy and points to Antigravity CLI. (#382)
  • Kiro CLI V3 remembers your setup – the model and effort you pick from the toolbar carry over to your next chat. A chat you reopen keeps the model, effort and mode it was using. (#313, #314)
  • Sign in to MCP servers from the chat – on Kiro CLI V3, a server that needs sign-in prompts you as soon as Kiro reports it. Servers that fail to start are listed at the top of the chat, with Kiro’s reason. (#328)
  • Decide how far “Always allow” reaches – on Kiro CLI V3, Always allow and Always deny ask whether to remember your choice for this vault or just this chat. For shell commands, you also choose the exact command or every command that starts the same way. (#330)
  • Fewer agents at launch – opening Obsidian with restored tabs starts only the agent for the tab you’re looking at. The others start when you open them. With 6 restored tabs, 1 agent starts at launch instead of 6. (#316)
  • Notes reach the agent in full, once – a pinned note’s content goes with the first message it appears in, so the agent can start without opening the file. Later messages send a link. @-mentions are no longer cut off at 10,000 characters: a note up to 16,000 characters is sent whole, and a bigger one is sent as a link the agent reads when it needs to. Session logs showed agents re-reading a cut-off note 24–44% of the time, against 8% for a note sent whole. (#345)
  • Choose how many chats to keep – a new Max saved chats setting (Settings → Sessions) keeps 25 to 500 chats, or all of them, with 100 as the default. Past the limit, the chat you used least recently leaves your history, and its saved messages go to your trash instead of being deleted. Chats open in a tab are never removed. The new Clean up unused chat files command clears files left behind by earlier versions. (#351)

Bug fixes

  • A chat that has to start over keeps its conversation – sometimes an agent can’t reopen a chat’s session: it no longer has it, a V3 chat was reopened on V2, a reload started fresh, or a retry followed a failed start. The new session now gets the most recent 10,000 characters of the conversation, without tool output or images, plus Agent Console’s instructions, with your next message. The chat keeps one history entry and its title. (#341, #343, #346, #349, #355, #357)
  • A chat only the agent kept says when it’s gone – opening a chat that exists only on the agent, not in Agent Console, now says when the agent can’t open it any more, with a Close button, instead of opening an empty chat. (#356)
  • Reload works on every chat – Reload connects a chat you just opened from Session history, a sleeping tab, or a tab restored after a restart. After a Reload fails, sending a message or clicking Reload again picks the same conversation back up once the agent can start. On Kiro CLI V2, Reload picks the conversation back up instead of starting fresh. A message you send during a reload goes out once the agent is back. Reload also keeps the tab’s own model, effort and mode, not the ones you last picked in another tab. (#325, #327, #377, #383, #393)
  • Your message isn’t lost when the agent fails to start – Send stays available after a failed start and retries the connection. Choosing Edit on a message waiting to send no longer loses it. While a message waits to send, its attachments are locked in with it: you can’t add or remove a file until it sends, and adding one tells you to wait. Before, an added file vanished and a removed one was sent anyway. (#322, #384)
  • Session history shows your words and your titles – restored agent chats show only what you typed, not Agent Console’s hidden instructions, and keep their title. The agent’s session list shows your title for chats you renamed. Starting a new chat right after a reply no longer overwrites the chat before it. Closing a tab right after a send no longer leaves an empty chat in Session history. Empty chats no longer pile up in the agent’s list on Kiro CLI V3 and Claude Code. (#329, #336, #348, #362, #366, #370)
  • Session history matches the folder you’re in – the agent’s session list starts on this folder’s chats, and Only this folder follows the folder the chat runs in. Fork is hidden on chats only the agent has, where it used to open a blank tab. The agent’s list also opens on a tab that hasn’t connected yet. (#318, #319, #340, #350)
  • Note links work from a chat in any folder – in a chat that runs in a folder other than your vault, such as one started with New chat in directory…, pinned and @-mentioned notes pointed at paths that don’t exist there. A note sent as a link couldn’t be opened at all. Note links now point at the real file in your vault. (#394)
  • Scrolling holds your place – dragging the scrollbar, clicking its track, or pressing PageUp or the arrow keys no longer snaps you back while a reply streams. A tab you scrolled up in returns to where you were reading. Switching back to a long chat no longer flashes the middle of it first, and resizing the pane keeps the text you were reading in view. (#353, #365, #401, #402)
  • Notifications open the right chat – clicking a notification that a reply is complete in macOS Notification Center now switches to that chat’s tab, even minutes after it appeared. Before, a late click brought Obsidian forward but left you on the tab you were on. (#399)
  • Keyboard users keep their place – pressing Tab onto a message’s copy button now shows the button, instead of landing on one you can’t see. When the Jump to latest pill disappears while it has focus, focus moves to the message box instead of being lost. (#405, #409)
  • One set of MCP servers per tab – starting a new chat in the same tab no longer stacks another set of MCP servers, about 15 processes each time on Kiro CLI V2. On V2, a new chat restarts Kiro, which takes a few seconds, and Rename with AI is grayed out. Closing a tab now shuts down all of its MCP servers. (#320, #324)
  • One update notice, and it stays reliable – with many tabs open, a new release shows one “Update available” notice instead of one per tab. A relaunch within 6 hours reuses the last update check, so the check stays within GitHub’s limits and the notice no longer goes missing. (#307, #308)
  • Clearer errors – on Kiro CLI V3, a busy model now reads Service Busy instead of “Authentication Required”, and other Kiro service errors read Agent Error. A Kiro agent set to V3 on a Kiro CLI that’s too old says Kiro CLI needs an update. (#344, #374)
  • Smaller fixes
    • Closing a tab closes its notifications, so an old notification no longer points at a tab that’s gone. (#309)
    • Choice buttons with hyphenated names show as buttons, not code. (#310)
    • Diffs show whitespace edits, multi-line blocks and new files correctly. A permission prompt goes away when its agent stops. Selecting a tool call’s title no longer collapses it. Flipping a setting no longer jumps the settings page to the top. (#311)
    • A shell command on Kiro CLI V3 shows its output once, not twice. (#332)
    • A finished or failed tool call no longer goes back to looking like it’s still running. (#400)
    • A failed tool call shows a flush circled ✕ instead of a tall pink pill. (#338)
    • Settings search finds rows inside the built-in agent pages. (#339)
    • Hovering the header shows the model you switched to on Antigravity CLI. (#379)
    • Antigravity CLI no longer shows Agent Console’s instructions in your message. (#358)
    • An OpenCode default agent survives a restart, and its name shows instead of an id. (#382)
    • Choosing V2 (legacy) for a Kiro agent runs V2, even when Kiro CLI’s own engine setting is V3. (#359)
    • Upgrading users no longer see the offer to import settings from another plugin. (#367)

Known issues

  • On Kiro CLI V3, an MCP sign-in that takes longer than 60 seconds fails, because Kiro counts it against the server’s connection timeout. MCP Tools shows how to raise it.
  • #409 changes how the message box handles a held Enter: an Enter you hold down on another control, such as the Jump to latest pill, no longer sends your draft when focus moves to the box. This was tested on macOS only. If a single press of Enter stops sending on Windows or Linux, please open an issue.

Install

Settings → Community plugins → Browse → “Agent Console” → Install → Enable. Existing users update via Settings → Community plugins → Check for updates.

README file from

Github

Agent Console

The AI agent plugin built for Obsidian power users.

Version Obsidian downloads License Build Docs

📦 Install from Obsidian  ·  📖 Documentation  ·  💬 Discussions

Why Agent Console

Most AI agents start every chat knowing nothing about you. Agent Console runs your agents inside your Obsidian vault. @mention a note and the agent starts from it, and it can follow the links. Ask it to write back – capture a decision, draft a note – and your second brain grows as you work.

When an agent asks to edit a file or run a command, you choose Allow or Reject. Keep the prompts you retype as notes, and fire one from a chip, a ! search, or a bookmark in your browser. Keep several jobs going, each in its own tab, with an icon that shows whether it’s busy, done, or waiting on you.

The shift in how you work: Tell the agent what you want done, not how to do it. With the right skills, it works out the steps, and you stay on the work itself.

See it in action

Features

  • Work from the notes you’ve already written – @mention a note and the agent reads it, and it can follow the links to related notes. Send images and files to the agent too, by copy-pasting them or with the attach button
  • Let your second brain grow itself – ask an agent to write back: capture decisions, draft notes, and link ideas straight into your vault, without the manual upkeep
  • Fire your go-to prompts in one click – keep kickoffs like “debrief this meeting” as notes and launch them from a chip or a ! search. show when puts a chip only on the notes it fits, and a prompt can pick its own agent and mode
  • See which chat needs you at a glance – each tab’s icon shows busy, waiting on you, finished while you were away, stopped early, or a failed message. New installs show the tabs as a slim column of icons; drag it wider for names
  • Stop waiting on one agent before starting the next – run several chats side by side, each in its own tab: one researching, one drafting, one fixing a script
  • Use the agent you’ve already set up – Claude Code, Codex, Kiro CLI, OpenCode, Antigravity CLI, Gemini CLI, or any custom agent built on the Agent Client Protocol. A fresh install picks one it finds, and a chat can switch agents with its conversation
  • See and approve what runs – commands the agent runs show in the chat with their output. When an agent asks before editing a file or running a command, you see the diff or the command and choose Allow or Reject
  • Your chats come back – tabs, conversations, and half-typed messages survive a restart, and Reopen closed session tab brings back a chat you closed. Search every past chat by title or by what was said in it
  • Your context budget goes further – a pinned note goes in full with your first message, then as a link, so a long chat uses 62–77% fewer context tokens than re-sending the full note each turn
  • Agents know they’re in Obsidian – each agent gets a short briefing on the first message so it writes links, math, and diagrams that render and works with your notes; you choose exactly what it’s told
  • Drive it all from the keyboard – every control is keyboard-operable with a clear focus ring; bind hotkeys to switch tabs, show the tab list, and fire quick prompts
  • Know which tab just finished – the “agent finished” notification names the tab it came from, and clicking it jumps there
  • Tabs name themselves – each tab picks up a short title from what you’re working on; rename any one yourself, drag to reorder, right-click to close
  • Scroll up to read while the agent is still typing – whether you use the wheel, scrollbar, or keys, the stream won’t pull you back down, and tabs keep their place
  • Idle tabs don’t hold memory – restored tabs don’t start their agent or MCP servers until you type, and a tab left alone for an hour puts its agent to sleep
  • Choose how many chats to keep – past that number, the least recently used goes to the trash instead of being deleted; closing a panel with several chats asks first
  • A working directory per chat – point each chat, or a specific agent, at the folder it should work in, and that choice sticks across restarts
  • Pick the right model for each task – switch modes and models per chat without restarting
  • Read the conversation, not the logs – tool calls render as one summary row you click to expand or collapse, and a failed call is flagged so you don’t miss it
  • Your agent’s slash commands, built in – type / in the message box to see and run the commands your agent offers
  • Answer with one click – when an agent offers a few options, it can show them as buttons; click one to send your pick as a normal message
  • Every link the agent shared, in one list – the header’s link button collects the notes and pages it pointed you to, newly created files first
  • Your MCP tools come along – whatever MCP servers your agent uses keep working in Agent Console with no extra setup
  • Use it in your language – Agent Console follows Obsidian’s language, with Korean, Chinese, and Japanese alongside English
  • Switching from Agent Client? Bring your setup – import its agent definitions, defaults, and API keys in one click
  • Looks like part of your Obsidian – it adopts Obsidian’s own styling and your theme’s variables instead of hardcoded colors and chrome, so any custom theme restyles it like a built-in panel

What you can do with it

  • Do several things at once – ask one agent to research a topic, another to prep your next meeting, and a third to draft a newsletter
  • Vibe-code a script or tool – describe what you want, get working code back, and review the diff
  • Pull in your notes – type @ and a note’s name, like a meeting note or a project page, and the agent works from it
  • Make your vault fill itself – with the right skills, agents write meeting notes, research summaries, and action items straight into your vault
  • Compare two agents – give the same task to two different agents in side-by-side tabs and see which answer you like better
  • Stay focused – status icons tell you when an agent is ready or still working, so you don’t break your flow checking on it

The pattern: tell the agent what you want done. Switch tabs. Come back when status says ready.

Start agent work from your other apps

Start a saved prompt from wherever the work shows up, with its agent, its mode, and your text already in place.

  • Hand off what’s in your browser – select part of an email thread or a GitHub issue, click a bookmark, and “draft a reply” or “fix this issue” opens in a new chat with that text
  • Start your day from a bookmark – it runs your daily review
  • Pick up where a script left off – have your script or terminal agent open the link when it finishes, and the follow-up waits in a new tab
obsidian://agent-console?vault=My%20Vault&prompt=draft-a-reply&selection=…&url=…&title=…&agent=…&mode=…

Copy link on a prompt chip gives you the link, and prompt links has the bookmarklet. A link only starts prompts already in your vault, and it waits for Enter unless it’s a plain link to a prompt you’ve explicitly set to auto-run (uri send).

Install

Agent Console is in the Obsidian Community Plugins store:

  1. Open Settings → Community plugins in Obsidian
  2. Click Browse and search for "Agent Console"
  3. Click Install, then Enable

Obsidian updates the plugin automatically when a new version is released. When an update is waiting, the chat header shows an Update pill – click it to jump straight to Settings → Community plugins, where you can update with one more click.

Quick start

You’ll need an AI agent installed on your computer. Popular choices:

  • Kiro CLI – Amazon’s agent; Agent Console runs its V3 engine, or V2 on older versions
  • Claude Code – Anthropic’s coding agent
  • Codex – the Agent Client Protocol Codex adapter
  • Antigravity CLI – Google’s command-line agent, successor to Gemini CLI
  • Gemini CLI – Google’s earlier command-line agent (legacy)
  • OpenCode – open-source agent; runs local models via ollama
  • Custom agents like Qwen Code and others

Once you’ve set up the agent:

  1. Open Settings → Agent Console
  2. Enter the path to the agent and any API keys it needs
  3. Click the robot icon in the ribbon to open the chat panel
  4. Click the + button to open more tabs as you need them

On a brand-new install, Agent Console detects an agent you already have, makes it the default, and opens the chat panel once – so steps 1–3 are often already done for you.

Configuration

Customize how each agent behaves under Settings → Agent Console – agent paths, modes, models, permissions, and tab behavior. Per-agent setup guides are in the documentation.

Hotkeys

Move between tabs with a keystroke instead of the mouse. Set your preferred bindings under Settings → Hotkeys.

How it works with your agent

Three things come together, and you control all of them:

  • Your notes give the agent context – it reads notes, mentions, and attachments from your vault, so it works with your knowledge instead of starting from scratch
  • Your agent stays itself – task instructions, custom rules, tools – Agent Console doesn’t get in the way of anything your agent already supports
  • Agent Console keeps it all in view – many tabs, each with a status icon, so you can run more at once without losing track

Whatever your agent can do, Agent Console lets you run several of those jobs at once.

Fixes that stay fixed

A plugin that runs live agent sessions in your vault has to be dependable. A few things keep it that way:

  • Every bug fix ships with a test – the test reproduces the bug first, so it stays fixed. There are 6,000+ tests across the plugin’s core logic
  • Every change is checked before it lands – each pull request runs linting, type-checking, the full test suite, and a build before it can merge
  • You can check the efficiency claim yourself – the “62–77% fewer context tokens” figure comes from a benchmark you can run (npm run bench:tokens), and a test keeps it in line with the shipped code
  • Slower context handling blocks a release – context-handling benchmarks run on every change, and a release can’t ship if they’re more than 15% slower than the saved baseline

Contributing

Issues and pull requests are welcome on the GitHub repo.

For a big new feature, please file an issue first so we can talk about scope. Bug fixes can go straight to a pull request.

License and Attribution

Apache License 2.0 – see LICENSE.

Agent Console is based on Agent Client by @RAIT-09, originally released under Apache-2.0. Changes are © Vinod Panicker. See NOTICE for full credits.

The Agent Client Protocol is developed by Zed Industries.