# SwarmPane: full reference > SwarmPane is a desktop workspace for AI coding agents. It runs the coding CLIs you already use in real terminals, beside your project files, Git and approvals. This file collects what https://swarmpane.com publishes about SwarmPane (key facts, FAQ, documentation and journal articles) as Markdown for language models and other tools. The shorter index is https://swarmpane.com/llms.txt. ## Key facts - Modes: Agent, Code, Chat, Workflow. Coding sessions run in real terminals on your computer, next to project files, Git, a browser preview and approvals. - Coding CLIs: Claude Code, Codex, Gemini CLI, Cursor Agent, GitHub Copilot CLI, OpenCode, Grok Build, Kimi CLI, Aider, Droid run in Code mode terminal panes. Claude Code and Codex also power Agent and Chat conversations. You install and sign in to each CLI with your own provider account. - Workflows: chain up to 16 steps; each result passes to the next agent, approval steps pause the run, and runs are saved. - Jev Conductor: Jev Conductor (beta) routes a goal to one of your enabled coding agents, runs the checks you choose, and when a check fails hands the work to another agent with the evidence attached. You review the result before accepting it. One-time allowance of 5 million routing input tokens with SwarmPane Pro. Coding-provider usage is separate. Conductor beta is subject to availability. - Platforms: SwarmPane is built for macOS 13 or later. Windows (x64) and Linux (x64) are test builds while platform testing continues. Native voice requires macOS, and there is no iOS or Android app. macOS status: public release in preparation; see https://swarmpane.com/downloads for installers. - Pricing: SwarmPane Pro has a planned launch price of $29/month. Checkout is not open yet, and creating an account is free. Provider subscriptions and API usage are billed separately by each provider. Packaged desktop access requires an eligible active Pro subscription. - Voice: OverVoice dictates into a conversation or terminal pane on macOS (default shortcut Command + Shift + Space). The transcript lands as an editable draft, and nothing is sent until you send it. - Data: projects, conversations and workspace settings stay on your computer. The local core only accepts connections from the same machine. ## Not available yet - Public checkout is not open yet; the $29/month Pro price is a planned launch price. - Google and Apple sign-in are not available yet. - Windows and Linux builds are test builds; feature parity is still being validated. - Native voice currently requires macOS. - Cloud voice needs SwarmPane Pro and an internet connection; without them, the app can transcribe on the Mac instead. - There is no iOS or Android app. - Coding-provider subscriptions and API usage are billed separately by each provider. - Projects, conversations and workspace settings stay on your computer; the website account does not sync them between computers. - Website demos use sample files and simulated activity. - There is no hosted public API for starting agent work. ## Pricing SwarmPane Pro: $29/month. Planned launch price. Checkout is not open yet. Everything in Pro: - Agent, Code, Chat, and Workflow modes - Your installed coding CLIs: Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, and more - 5 million Jev tokens included: a one-time routing input-token allowance for Conductor (beta) - Files, Git, browser preview, MCP tools and approvals in the same window - 1,500 credits a month · cloud voice to text, about 25 hours (one credit is one minute, billed by the second; credits reset each billing period and don’t roll over) SwarmPane Pro has a planned launch price of $29/month. Checkout is not open yet, and creating an account is free. Provider subscriptions and API usage are billed separately by each provider. One-time allowance of 5 million routing input tokens with SwarmPane Pro. Coding-provider usage is separate. Conductor beta is subject to availability. ## Platforms SwarmPane is built for macOS 13 or later. Windows (x64) and Linux (x64) are test builds while platform testing continues. Native voice requires macOS, and there is no iOS or Android app. - macOS (.dmg / .zip): Public release in preparation. macOS 13+ runtime minimum. Apple silicon tested; universal packaging configured. - Windows (.exe): Test-build target. x64 installer target. Platform acceptance testing in progress. - Linux (.AppImage / .deb): Test-build target. x64 packages. Distribution compatibility under review. ## Workspace modes ### Agent: Agents with a job and a memory. Create teammates with their own instructions, skills, and long-term memory. Give one a task in your project and follow every turn, with your approval before it edits files. - Instructions, skills & memory - Claude Code or Codex, per turn - Guided approval before edits ### Code: Real terminals, side by side. Run Claude Code, Codex, Cursor, Gemini CLI, and more in real terminals. Arrange them in Solo, Pair, Workbench, or Swarm layouts with files, Git, and a browser preview beside them. - Persistent terminal sessions - Files, diffs & commits - Browser preview & project search ### Chat: Think it through first. A focused conversation for planning, drafting, and questions. Attach files and images, switch providers as you go, and pick up where you left off. - Saved conversations - Files & images as context - Readable code in every reply ### Workflow: Chain agents into a process. Connect up to sixteen steps on a canvas. Each result moves to the next agent, approval steps pause the run for you, and every run is saved. - Visual step editor - Output passed between agents - Approval steps & run history ## Frequently asked questions ### What is SwarmPane? SwarmPane is a desktop workspace for AI coding agents. It runs the coding CLIs you already use, like Claude Code and Codex, and keeps their conversations, live terminals, project files, browser previews, and approvals together. ### How much does SwarmPane cost? SwarmPane Pro has a planned launch price of $29/month. Checkout is not open yet, and creating an account is free. Your coding providers bill their own subscriptions and usage separately. ### Which coding agents can I use? SwarmPane detects installed CLIs including Claude Code, Codex, Cursor Agent, Gemini CLI, GitHub Copilot CLI, and Aider. Install and sign in to the provider you want to use. ### What are the 5 million Jev tokens for? One-time allowance of 5 million routing input tokens with SwarmPane Pro. Coding-provider usage is separate. Conductor beta is subject to availability. Jev uses these tokens to choose agents and decide when to retry or hand off. This is not a monthly refill and does not cover coding-agent tokens. ### Does Pro include coding-agent usage? Your coding CLI provider account, subscription, API key, and usage are managed separately by that provider. The included Jev allowance covers Conductor routing only. ### What does Conductor do? You describe the objective and success criteria. Jev selects from your available agents, runs your verification commands, and can retry or hand off with the previous context. You review the result before accepting it. Conductor is currently in beta. ### Where does the code run? Coding sessions run through CLIs installed on your computer, in terminals connected to your project folders. You can inspect files and changes in the same workspace. ### Which operating systems does SwarmPane target? macOS, Windows, and Linux desktop builds are configured. Windows and Linux are test-build targets with platform validation in progress. Native voice currently requires macOS. Check Downloads for verified public installers; iOS and Android apps are not available. ### Can I cancel my subscription? Yes. When checkout is available, you can manage or cancel your Pro subscription from your account. The planned launch price is shown until public checkout opens. ## Documentation ### Overview: A better place to start. URL: https://swarmpane.com/docs Learn your workspace, connect your tools, and take your first task from an idea to a reviewed result. #### Your first session SwarmPane is a desktop workspace for installed coding agents. Your SwarmPane account provides access to the app; your coding-provider accounts run the AI work. Start with a small project and a task whose result you can check. Related: [Follow the getting started guide](https://swarmpane.com/docs/getting-started), [Check installer availability](https://swarmpane.com/downloads) #### Four ways to work - Agent — give a persistent teammate instructions, context, and a task. - Code — work with project folders, terminal panes, files, Git, and browser previews. - Chat — explore questions and keep the conversation with its context. - Workflow — connect steps into a repeatable sequence and inspect each result. Related: [Explore the interactive demo](https://swarmpane.com/#demo) #### Know your build SwarmPane has desktop build targets for macOS, Windows, and Linux. Public installer availability and feature readiness are separate: Windows and Linux are test-build targets, and voice currently depends on macOS. SwarmPane does not currently have an iPhone or iPad app. Related: [Compare platform status](https://swarmpane.com/docs/platforms) ### Getting started: Your first workspace. URL: https://swarmpane.com/docs/getting-started Install the app, connect your accounts, and start a small, reviewable coding task. #### Before you start - Check Downloads for a verified installer for your operating system. A test build is not a public release. - Install and authenticate the coding CLI you plan to use. Provider subscriptions and API usage are separate. - Choose a project folder you own and can safely edit. Keep changes under version control. Related: [Downloads](https://swarmpane.com/downloads), [Platform details](https://swarmpane.com/docs/platforms) #### Connect your SwarmPane account If checkout or public downloads are not yet available, use the browser demo to explore the interface. Creating an account does not start a paid subscription. 1. Create an account or sign in on the website. 2. Open the desktop app and choose Log in. 3. Compare the connection code in the browser with the one in your app, then approve your connection. 4. Return to the app. Packaged desktop access requires an eligible active Pro subscription. Related: [Create an account](https://swarmpane.com/signup), [Account and billing help](https://swarmpane.com/docs/accounts) #### Run a first task 1. Switch to Code and add a workspace using your project folder. 2. Open a coding-agent pane and select the intended provider account. 3. Ask for a small change with a clear finish line. For example: “Explain the failing test, make the smallest fix, and show the diff.” 4. Inspect changed files and terminal output. Run the relevant checks before accepting the result. Related: [Learn Code mode](https://swarmpane.com/docs/code) ### Platforms: Your workspace. Your OS. URL: https://swarmpane.com/docs/platforms What macOS, Windows, and Linux builds support today, and what remains in testing. #### macOS The macOS packaging targets DMG and ZIP installers. The runtime minimum is macOS 13; the current local verification is on Apple silicon. Universal packaging is configured for Apple silicon and Intel, but an Intel target alone does not establish hardware verification. Public Mac downloads require signing, notarization, and release verification. Local test DMGs are not a substitute for a verified public installer. Related: [Check Mac download availability](https://swarmpane.com/downloads) #### Windows An x64 Windows EXE installer target is configured. Windows builds are test candidates while installation, terminal behavior, and feature parity are validated on Windows. Native voice capture is not available in Windows test builds. Use keyboard input for prompts. A working website demo is not evidence that a Windows installer has passed acceptance testing. #### Linux Linux x64 packaging targets AppImage and Debian packages. These are test candidates; support for individual distributions and desktop environments has not been established. Native voice capture is not available in Linux test builds. Check the release notes for the exact artifact and tested environment before installing. #### Phones and tablets The website and sample demo can be viewed on mobile. SwarmPane is a desktop application; there is no iOS or Android app release. Phone companion and remote pairing are preview features, not a supported mobile product. #### Download availability The Downloads page links an installer only when a verified public download URL has been configured. An unavailable download is shown as a status, rather than a button that cannot deliver a file. Related: [Downloads](https://swarmpane.com/downloads), [Development changelog](https://swarmpane.com/changelog) ### Agent mode: Give the work a teammate. URL: https://swarmpane.com/docs/agent Set up persistent agents with a clear brief and keep the result easy to review. #### Start with a useful brief Agent mode organizes work around persistent teammates. Give an agent a name, instructions, and the context it needs. Describe the outcome and any constraints before asking it to act. - Outcome: what should be different when the task is done? - Context: which project, files, or prior decisions matter? - Evidence: what should the agent show so you can verify the result? #### Keep context intentional Saved instructions, memories, and skills help an agent repeat your working preferences. Review that context when the task or project changes. A remembered instruction should not accidentally authorize a new external action. #### Review the handoff Inspect the task status and output. A completed turn means the run ended successfully; use file changes, tests, and other evidence to decide whether the task meets your standard. Failed and canceled runs need their own follow-up. Related: [Write a verifiable assignment](https://swarmpane.com/blog/agent-workflows/), [Tasks and routines](https://swarmpane.com/docs/routines) ### Code mode: Stay close to every line. URL: https://swarmpane.com/docs/code Use real terminals, project files, Git, and browser previews in the same workspace. #### Open a project Add a project folder in Code mode. Its terminal panes and file tools work against that workspace. Use Solo, Pair, Workbench, or Swarm layouts to arrange the work you need to see. #### Choose a coding CLI The Code workspace detects installed coding CLIs and opens terminal sessions. Detection tells you a command is available; it does not confirm provider sign-in, credits, or access to a particular model. Claude Code and Codex are supported by the conversational runner. Other CLIs can be used in terminal panes when installed; support varies between terminal, conversation, and workflow features. Related: [Provider accounts](https://swarmpane.com/docs/accounts) #### Inspect the result Agent edits affect real files in the desktop app. The website demo uses an isolated sample workspace and simulated terminal responses. 1. Open Files to inspect or edit a project file. 2. Review Git status and the diff to see what changed. 3. Use Browser for a project preview when its development server is running. 4. Run project checks in a terminal and inspect the output before committing. #### Try it before installing The browser demo uses the app interface with sample files. Try help, ls, cat src/App.tsx, or npm test in a sample terminal. Native dialogs and provider authentication require the desktop app. Related: [Open the demo](https://swarmpane.com/#demo) ### Chat mode: Keep the conversation. URL: https://swarmpane.com/docs/chat Ask questions, refine a brief, and return to the thread with its context intact. #### Start a conversation Choose Chat for a focused conversation. Select the provider and account you intend to use, add relevant context, and write your request. Your SwarmPane account and the selected coding-provider account serve different purposes. #### Attach only what helps Give the agent the files or context needed for the question. Say whether you want explanation, a plan, or an implementation. A useful review request identifies the change and the evidence you want back. #### Return to your work Conversation drafts stay with their conversations. Switching workspace modes preserves the mounted workspace state, so you can inspect code and return to the discussion. Desktop conversation history is stored locally. Signing into the same SwarmPane account on another computer does not currently synchronize your threads or project files. Related: [Accounts and local data](https://swarmpane.com/docs/accounts) ### Workflows: Make the next step clear. URL: https://swarmpane.com/docs/workflows Connect work into a sequence with explicit inputs, outputs, and review points. #### Build a small sequence Start with a short sequence such as inspect → implement → review. A workflow can have up to sixteen steps. Model and provider availability depend on your installed tools and accounts. 1. Open Workflow and create a workflow for your project. 2. Choose the step type and the installed provider or configured tool it needs. 3. Write an instruction that names the expected result. Configure edit permission deliberately. 4. Add the next step and make its dependency on earlier output explicit. 5. Run the workflow and inspect the result of each step. #### Place approval where it matters A human-approval step pauses a run for a decision. Use it before a consequential action, and inspect the proposed change and prior output before continuing. Per-step editing settings control intended access for that step; they are not a guarantee that every external CLI or tool has identical permission behavior. #### Handle failures explicitly A stopped or failed workflow is different from a completed one. Read the step error, correct the configuration or underlying issue, and decide whether to run again. Check for partial changes before retrying commands that modify files or external state. Related: [Troubleshooting](https://swarmpane.com/docs/troubleshooting) ### Voice & dictation: Speak. Refine. Send. URL: https://swarmpane.com/docs/voice Understand voice availability, permissions, and where your transcript goes. #### Current availability The native voice implementation uses a macOS helper. Windows and Linux test builds do not support native voice capture. The website voice example is simulated and does not use your microphone. Voice availability can vary by build. If your installed version does not expose voice controls, the website example does not enable them. Check the build’s release notes. #### Dictate into the right place 1. Select the conversation or coding pane that should receive the words. 2. Use the voice control or configured shortcut when available. The default Mac shortcut is Command + Shift + Space. 3. Allow microphone and speech permissions when macOS requests them. 4. Stop recording and review the transcript in the destination draft. 5. Edit names, commands, and punctuation before sending. #### Speech engines and permissions The macOS system speech engine requests on-device recognition where supported. Whisper requires the configured local model and runtime. An unavailable engine or missing model can prevent recording or transcription. If the shortcut does nothing, check the app’s recording state and shortcut configuration. If permissions were denied, review SwarmPane in macOS Privacy & Security settings before retrying. Related: [Voice troubleshooting](https://swarmpane.com/docs/troubleshooting#voice) ### Accounts & billing: Your accounts, explained. URL: https://swarmpane.com/docs/accounts Separate workspace access, provider usage, and local project data. #### Your SwarmPane account Use your SwarmPane account for website sign-in, desktop pairing, and subscription access. Create an account with an available sign-in method, and complete email verification when requested. Packaged desktop access requires an eligible active Pro subscription. Creating an account alone does not purchase a plan. Checkout availability and the final price are shown before payment. Related: [My account](https://swarmpane.com/account), [Pricing](https://swarmpane.com/pricing) #### Your coding-provider accounts Claude Code, Codex, and other coding providers manage their own sign-in and usage. Install the CLI and authenticate with that provider. Where account profiles are supported, name the profiles and select the intended account in the pane or conversation. SwarmPane Pro does not include provider subscriptions or API usage. A provider limit or sign-in error must be resolved with that provider account. Local activity counts are not the same as live provider quota. #### Local data and multiple computers Project folders, conversation history, and workspace configuration live on the computer running the app. Website account sign-in does not currently synchronize those files between computers. Keep backups appropriate to your projects. #### Manage access and billing Open My account to inspect your plan or access the billing portal when available. Canceling SwarmPane does not cancel your coding-provider subscriptions. Related: [Password recovery](https://swarmpane.com/forgot-password), [Cancellation and refunds](https://swarmpane.com/refunds) ### MCP, skills & plugins: Connect with context. URL: https://swarmpane.com/docs/mcp Understand MCP connections, tool access, and what a Preview label means. #### Inspect the connection MCP connections make tools available through a configured server. Review the server command, arguments, environment, and intended access before enabling a local connection. A connection can depend on installed software and separate service credentials. Keep secrets out of shared instructions and screenshots. #### Discovery is only the first check Discovering tool names means the server responded. It does not establish that every action is appropriate or will succeed. Inspect a tool’s input and result, and review consequential actions before running them. The connector catalog can include Preview entries. A Preview listing is not an authenticated connection or proof that a working adapter is available. #### Skills, plugins, and access Skills supply reusable instructions; plugins can package additional capabilities. Read what they add and which services they require. Installing instructions does not automatically grant a provider account or permission to act in external systems. Related: [A guide to MCP boundaries](https://swarmpane.com/blog/mcp-with-boundaries/) ### Tasks & routines: Make repeatable work useful. URL: https://swarmpane.com/docs/routines Keep recurring responsibilities specific, observable, and easy to review. #### Define the task first Use a task to name an outcome and track its status. Give the agent enough context to work and an explicit standard for completion. A recurring responsibility should start as a task you can run and verify once. #### Schedule with clear expectations Routines provide a schedule for agent work. Confirm the intended agent, instructions, project, and timing before enabling a routine. The local application needs to be running for local work to execute; this is not a promise of an always-on hosted agent. Provider authentication, usage limits, and device availability can interrupt a run. Review completed and failed work instead of assuming that a schedule guarantees an outcome. #### Keep the output reviewable - Ask for a concise summary and the evidence behind it. - Keep external actions behind an explicit review decision. - Update or disable a routine when its project or responsibility changes. Related: [Workflows](https://swarmpane.com/docs/workflows) ### Troubleshooting: Get back to building. URL: https://swarmpane.com/docs/troubleshooting A practical checklist for installation, sign-in, agents, terminals, and voice. #### The app will not install or open Confirm that the installer matches your operating system and architecture. Read its release status: a local test artifact is not a signed public release. Windows and Linux candidates still require platform testing. If a verified installer is not available, use the browser demo. Do not bypass an operating-system warning to treat an unverified artifact as a production release. Related: [Installer status](https://swarmpane.com/downloads) #### Desktop sign-in is not completing - Use the same account you created on the website. - Start a fresh connection from the desktop app if the code expired. - Check that the browser connection code matches your app. - Check your account’s subscription status; website login and desktop entitlement are separate. #### An agent is missing or returns an error - Check that the CLI is installed and runs in your normal terminal. - Sign in through the provider and confirm that the account can run a request. - Select the intended account in the workspace. - Read the exact provider error. Authentication, model access, and usage limits have different remedies. #### Voice does not start or produces no text - Check that you are using a macOS build with voice controls. - Review microphone and speech permissions in system settings. - Confirm that the selected engine and any required local model are available. - Check the selected destination and recording status before retrying. #### Share a useful report Include your app version, operating system, the steps to reproduce, and what you expected. Remove credentials and private project content from logs or screenshots. Related: [Contact and support details](https://swarmpane.com/contact), [Development changelog](https://swarmpane.com/changelog) ## Journal ### Give every agent a job you can verify URL: https://swarmpane.com/blog/agent-workflows/ Published: 2026-09-26 · Workflow · 5 min read A practical way to turn vague AI coding requests into assignments with evidence and a clear finish line. #### Start with an outcome “Improve this app” is an understandable goal, but it gives an agent too many possible directions. Begin with a concrete outcome: “The voice shortcut should open a recorder in Code mode, and the transcript should appear in the selected terminal without running it.” A specific outcome makes it easier to choose the right files, test the change, and know when the work is done. A useful brief says who is affected, what happens today, what should happen instead, and which boundaries matter. If you want the agent to draft a command but wait for you to press Enter, say that explicitly. The boundary is part of the feature. #### Ask for evidence, not a confident summary Good agent work leaves traces you can inspect. For a bug fix, ask which failure was reproduced, which path changed, and what test or manual check passed. For a design change, ask to see the page at desktop and phone widths. For a workflow, ask for the state transitions and the failure path. Keep the evidence close to the work. A terminal transcript, file diff, and approval request in the same workspace make review faster than a polished paragraph that says everything looks fine. #### Use a small finish line Split large ambitions into reviewable steps. A useful sequence is: reproduce the problem, fix the smallest cause, test the affected path, then decide what remains. That sequence does not stop you from building a powerful product. It keeps each addition trustworthy. When a change touches permissions, accounts, payments, or deployment, make the final action visible. The agent can prepare the implementation and checks; you should be able to inspect what will happen before anything goes live. ### When speaking helps you code URL: https://swarmpane.com/blog/voice-for-coding/ Published: 2026-09-26 · Voice · 4 min read Use voice to capture intent and feedback quickly while keeping command execution under your control. #### Speak the idea, then inspect the text Voice is useful when the thought is larger than the sentence you want to type. Explain a bug while looking at the screen. Describe the expected behavior. Dictate a review comment after reading a diff. These are all natural speech tasks. Code and terminal input need one extra step. Recognition can mishear a flag, filename, or symbol. A safe voice workflow inserts the transcript as editable text and waits for you to submit it. That keeps dictation fast without turning recognition errors into executed commands. #### Keep the recording state obvious A shortcut should tell you whether recording started, whether it is processing, and where the text will land. If speech recognition fails or permission is denied, the app should say so in place. A recorder that simply disappears leaves you guessing about the microphone, the shortcut, and your words. A compact floating control can help when you are in another app, but it should not cover the thing you are reading. Clear state and predictable placement matter more than a large animation. #### Pick the right destination A transcript meant for a coding terminal belongs in that terminal draft. A transcript meant for an agent conversation belongs in the chat composer. Changing modes after recording can be surprising, so choose the destination before you speak and show it while recording. Voice works best as another input method for an existing workflow. You should be able to edit, discard, or submit the words exactly as you would with a keyboard. ### Connect tools without losing track of access URL: https://swarmpane.com/blog/mcp-with-boundaries/ Published: 2026-09-26 · MCP · 6 min read How to think about local MCP servers, tool discovery, permissions, and review before an agent acts. #### Know what the server can reach An MCP server gives an application a structured way to discover and call tools. The interesting question is not only whether it connects. It is what those tools can read or change on your computer or in connected services. Before adding a local server, inspect the command that starts it, the arguments passed to it, and the accounts or folders it can access. Treat the server process as software you are choosing to run. Give it only the scope it needs for the job. #### Separate discovery from action A connection test should tell you that the process starts and expose the names and schemas of available tools. That is useful evidence, but it is not proof that every tool call is safe or will succeed. An action call needs its own arguments, result, and error display. When tools write files, send messages, or change external state, show the proposed action before it runs. The more consequential the action, the more specific that review should be. #### Design for failure Local servers can fail to launch, wait indefinitely, return an error, or produce a result too large to display. A good client sets time limits, closes the process when finished, and gives a clear error instead of showing a permanent “connected” badge. Keep a short path back to control: remove a server, retry discovery, and see which tool was called. A reliable integration is one whose behavior you can understand when something goes wrong.