Plexiform puts a small traffic light on your desktop for supported coding-agent sessions: green while it works, amber when it is your turn, red when it is stuck on you. Claude Code first, with Codex, Cursor and Gemini reporting in through hooks. The team board hands cards to agents and shows where each one got to.

Free Plexiform beta. Connect your existing AI tools; their charges and limits still apply. The first open takes a few extra clicks: here's how.

Give a card to an agent and watch it happen.

  1. Press Tackle with AIOn any card on your team's board. It runs on your machine, in your repo.
  2. Watch the card moveIt goes to In progress the moment the agent starts, and and the card stays in step with the agent.
  3. Supported input requests are visibleWhen a configured agent reports a current input request, its card and widget show that it needs you.
  4. Amber means your turnWhen it finishes there's a pull request to review and a handover with what it did.

Three lamps show reported agent activity.

You shouldn't have to tab through terminals to find the one that's waiting. Press a lamp to see what the character does.

A team board that lives next to your agents. Beta

Cards, handovers and run progress sit in the same app as the lamps. When an agent dies mid-run, the card says so, and the handover tells you where it got to.

The board: cards in To do, In progress, Review and Done. Each card shows who is running it, what it needs, and what it costs. A card's handover: the current hypothesis, the plan, what is done, and the files touched, so anyone can pick it up. The table view: every card in a sortable, filterable list. The dashboard: cycle time, cards done per week, who finished the work, where cards wait, and cost per card.

The real board, with sample data for an imaginary shop.

  • Cards run on the assignee's own Mac, in their own repo, with their own Claude or Codex login. AI execution uses the assignee's provider login; cards, run metadata and selected task messages sync through your team's hub.
  • Every run keeps a handover: a short narrative, recent actions, and the files and branch involved.
  • Supported permission requests can appear in the widget. Answering is off by default and must be enabled.
  • Text passes a filter for credential-like strings and home-folder paths before it leaves your machine.

Agents that can talk to each other. Coming

Your Claude can ask a teammate's Claude a question, or hand a finished piece of work to the agent that needs it, through the board, with a human able to read every message.

AnaClaude · reviewing MeiClaude · testing KofiCodex · migrating
  • Ana Claude · reviewing
  • Mei Claude · testing
  • Kofi Codex · migrating

An illustration of how it is planned to work. Not shipped yet.

GitHub, Slack and Sentry today. Linear and Jira next.

GitHub pull requests and builds light the widget today. The rest are on the way. Plexiform is the hub; these are the things it listens to.

Plexiform GitHub Slack Sentry Linear Jira
  • GitHub In the beta
  • Slack In the beta
  • Sentry In the beta
  • Linear Coming
  • Jira Coming

Solid outline: in the betaDashed outline: coming

One desk for supported agents.

Claude Code has the deepest support today. Configured hooks can report Codex, Cursor and Gemini activity to the same lamps. Activity visibility does not make every provider messageable; supported runners and explicitly configured channels have separate capabilities.

It works with the logins you already have. There is no shared key and no account to swap into.

Claude CodeCodexCursorGemini
The Usage tab's headline tiles: cost, turns, tokens and cache hit rate, each with its change and a trend line.The Usage tab's cost chart: daily cost stacked by model, with one spike.
The Usage tab in the app, showing sample data.

Know what your agents cost, and where. Beta

Plexiform keeps its own daily record of your usage, so history reaches back past the month of transcripts Claude Code keeps. See spend by model, project and hour; watch cache hit rate; get a nudge when today is well above your usual.

  • Set a daily or weekly budget and a runaway-session alert.
  • On a subscription, amounts are shown as API-equivalent dollars and labelled as not a bill.
  • Anything it can't price is counted and shown as unpriced, never dropped.

A character that's yours.

Buddy comes with friends. Give them costumes, eye moods and poses, or turn into a ghost when you hit a usage limit. Click one. More characters are coming, and so is a way to hatch your own. Hatch: coming

Where your data is stored and shared.

The desktop app sends no telemetry or analytics, and the widget needs no account. Its usage reader calculates tokens and cost locally without copying full conversations. Local Tasks does keep task text, messages, AI activity and replies; pending widget requests and session reports can also contain bounded content.

Your team's hub receives cards, handovers, run progress and signed-in account details. In the release candidate, supported sessions in signed-in team workspaces are automatically shared to watch, and the owner can allow sending for a session (terminal Claude channels you attach on purpose allow it from the start); personal sessions stay separate. Viewers can watch, and members with write access can message. Processing a message uses the session owner's provider quota and permissions. You can stop a private-session share or disable automatic sharing.

Live session messages and replies are readable by the hub operator while relayed. Board task-inbox messages, receipts and replies are stored durably as team history; queuing does not prove agent acknowledgement or resume a run. Update and download requests reveal your IP address and app version. Cloudflare serves this website; a recent check found a Web Analytics beacon in the served privacy and download pages.

This draft summary describes prepared release-candidate sharing behavior, not a claim that the candidate is already published. Legal review, hosting analytics verification and release acceptance are pending.

Read the privacy notice

  • Voice questions are turned into text on your device and are off by default.
  • Calendar access is off by default, and event titles aren't read unless you turn that on.
  • Crash reports are recorded and scrubbed on your Mac only. Sending them is not built yet, and when it is, it will be opt-in.
  • Your rules, settings and cameo photos stay in a folder on your Mac.

Free for you and your team.

The desktop app and team features are free. Agent Mail, including Ask my AI, stays free. Connect your own AI tools and accounts; their subscriptions, usage charges and permissions still apply.

Each team has its own private boards and members. Free teams support up to 25 members and 10 boards, with fair-use storage and signup limits.

Public registration is being prepared. Existing team access still uses invitations.

  • Published beta installers are available for macOS and Linux. Windows is unavailable while its runtime acceptance is pending; check the download page for published builds.
  • A phone companion, to approve a permission prompt when you're away from your desk, is coming.
  • Every change is written up in the changelog, and the docs list what it adds to Claude Code's settings.

Questions people ask first

Which agents does it work with?

Claude Code has the deepest support. Configured hooks can report Codex, Cursor and Gemini activity. Messaging requires a supported current runner or an explicitly configured session channel; hooks alone do not provide it for every provider.

Does it read my prompts or my code?

The usage reader reads Claude Code logs locally and keeps counts and costs. Tasks keeps task text, sent messages, AI activity and replies locally; temporary widget requests may include tool input or questions. Content you send to a team board or supported shared session can be read by its hub operator, and task-inbox messages and replies are stored there. Your chosen AI provider handles prompts and code under its own terms.

What does the app send out?

The desktop app sends no telemetry or analytics. Team boards receive the content you share; the release candidate automatically shares supported sessions associated with a signed-in team workspace. The hub operator can read shared messages and replies, and task inboxes retain them. Updates reveal IP address and app version. Website hosting includes an injected Cloudflare Web Analytics beacon. See the privacy notice for the boundaries.

Why does the first open take extra steps?

The beta isn't signed with an Apple Developer ID yet, so macOS can ask for confirmation in System Settings, including after updates. The steps are on the download page.

Is there a Windows or Linux version?

Linux AppImage and .deb beta installers are published. Windows is unavailable while its runtime acceptance is pending. The download page lists the published builds; it does not mean the prepared release candidate is already downloadable.

Who is Buddy?

The pixel character who holds up the traffic light.

How do I get in?

You can install the desktop app now. Public registration is being prepared; existing team access still uses invitations.

Hear when public signup opens.

Get an email when free public registration opens. We'll only use your email about Plexiform's launch.