Everything Overdeck does

Overdeck runs and directs AI coding agents across harnesses, models and machines. Here is every part of it, each with a link to the docs that cover it. Questions? Ask in the Overdeck Discord.

Conversations and code

Talk to agents and see what they changed, beside the chat.

  • Conversations

    Talk to an agent in a rendered view, or switch to the raw terminal of the same session. Hand off to a new conversation to change the model.

    Docs: Conversations
  • The turn card

    Each turn that edits files ends with a card listing the files, the lines added and removed, and whether the commit is pushed.

    Docs: The turn card
  • Code beside the chat

    Open a file beside the conversation, with a blue bar on each line that conversation changed.

    Docs: File explorer
  • The full diff window

    All turns shows only this conversation's edits. You can also compare the branch with main, or any two points.

    Docs: Full diff window
  • Edit, then a pull request

    Your edits go to an edit workspace on its own branch and reach main only when a pull request merges.

    Docs: Editing and Create PR
  • File viewers

    Images, PDFs, CSV, SQLite, notebooks and Mermaid diagrams open in the dashboard. Notebooks never run code.

    Docs: File viewers
  • Open in Overdeck

    pan open opens a file or folder from your terminal in its project. An optional file manager entry adds Open in Overdeck to the right-click menu.

    Docs: Open in Overdeck
  • Keyboard shortcuts

    Press ? anywhere outside a text field to see every shortcut for the page you are on.

    Docs: Keyboard shortcuts
  • New conversation with options

    Pick the project, machine, model, harness, effort, skills and a linked issue before the first message.

    Docs: New conversation with options
  • Handoff and fork

    Carry a conversation's context into a new one with pan handoff, or start fresh. This is also how you switch models.

    Docs: Handing off
  • Steer a running turn

    Interrupt a Claude Code agent or a Pi conversation mid-turn, queue a message, or hold it until your usage limit resets.

    Docs: Sending to a busy agent
  • Scheduled conversations

    Send a stored prompt to a conversation on a schedule with pan conv schedule.

    Docs: Scheduled runs
  • Subagents beside the chat

    Claude Code and Codex subagents are listed beside their parent with their transcripts. Codex app-server subagents also get their own composer.

    Docs: Subagent activity
  • Bookmarks, citations and attachments

    Bookmark messages, quote part of a reply into your next prompt, and attach files and images.

    Docs: Bookmarks and citations
  • Answer prompts from the dashboard

    Permission prompts, questions and plans to approve open as dialogs and wait in Needs you, including those from subagents.

    Docs: Needs you
  • Settle and snooze

    Split conversations into Needs you, Inbox, Snoozed and Settled, and settle them automatically after idle days or when the pull request merges.

    Docs: Settling and snoozing
  • Composer commands

    /pan <verb> runs Overdeck commands from the composer on every harness, and / lists your skills on Claude Code.

    Docs: Composer commands
  • Native terminal UIs

    Attach Codex or OpenCode's own terminal UI to a running conversation from its Terminal tab.

    Docs: Native Codex CLI in Terminal
  • Command palette and search

    Ctrl+K searches issues, conversations and commands. Semantic search over conversations needs an OpenAI key.

    Docs: Keyboard shortcuts
  • Markdown preview

    Markdown files get a rendered preview in code and diff tabs, with front matter shown as a list and working relative links.

    Docs: File explorer
  • Simple and Advanced mode

    New profiles open in Simple mode, and Home shows a Get set up checklist of what still needs attention.

    Docs: Quickstart
  • Voice input and read-aloud Opt-in

    Dictate into the composer, and have the dashboard speak agent activity through a local text-to-speech voice.

    Docs: Conversations
The full diff window on All turns, with lua/reachable_hostiles.lua in a split view: the old file on the left and the new file on the right, with 16 added lines highlighted.

Issues and the pipeline

Hand over an issue and let agents carry it to a pull request you merge.

  • The issue pipeline

    pan start plans the issue if needed, makes a workspace and starts a work agent. Agents build, review and test the work, and you click Merge.

    Docs: pan start and agent commands
  • Planning

    A plan agent reads the issue and writes a PRD and a checklist plan the work agent follows. Run it with pan plan, or let pan start do it.

    Docs: Agent commands
  • Issue trackers

    Issues come from GitHub, Linear, GitLab, Rally or Jira.

    Docs: Issue trackers
  • GitHub identities

    Overdeck acts as the GitHub App bot where the App is installed and as your own gh account everywhere else.

    Docs: GitHub organizations and identities
  • Workspaces

    Each issue gets its own git worktree and feature branch, so several agents work at once without conflicts.

    Docs: Workspaces
  • Specialist agents

    Review and test agents (the test agent also runs Playwright UAT) take a finished build to an approved, green pull request. You click Merge.

    Docs: Convoys
  • Verification gate

    Typecheck, lint, the tests you touched and a plan check run before review, queued so the machine does not overload.

    Docs: Quality gates in the project registry
  • Conflict repair

    An approved pull request that turns conflicting gets one repair instruction, and a stalled reviewer is started again.

    Docs: Convoys
  • One merge gate

    Every merge path asks the same questions: approval at the head commit, green CI, mergeable, and the UAT verdict.

    Docs: UAT batch trains
  • Close-out

    pan close checks the definition of done and cleans up the workspace and branch.

    Docs: Workspaces
  • Pipeline journal and pan show

    Each issue keeps an append-only journal. pan show <id> prints its state, journal and health.

    Docs: Convoys
  • Issue view, board and pipeline pages

    A conversation-first issue view with Discussion, PRD, Changes and Checks tabs, plus a kanban board and a pipeline page by phase.

    Docs: Mission Control
  • Decisions from agents

    Agents ask you a question with two to four options through pan ask. It lands in Needs you and on your phone.

    Docs: Push notifications
  • CI and review findings go back to the agent

    Failing CI checks, with a log excerpt, and CodeRabbit findings go to the work agent to fix.

    Docs: Auto-merge and CI
  • Review convoy

    In full review mode, four reviewers run in parallel and one writes the summary.

    Docs: Convoys
  • Merge trains Opt-in

    Keep one assembled, testable batch of every ready feature, so you can try the result and land exactly what you tested. Turn on the Merge train switch on the Flywheel page.

    Docs: UAT batch trains
  • Strikes

    Small fix pull requests from pan strike get one quick review. With auto-merge on, they merge ahead of the train.

    Docs: UAT batch trains
  • Auto-merge Opt-in

    Scheduled merging for approved, green pull requests, with a banner that lets you cancel first.

    Docs: Auto-merge
  • Swarm

    Run one work agent per plan item, in parallel, on the same issue, each in its own worktree.

    Docs: Swarm
  • Gauntlet lanes

    Launch builder, critic and verifier lanes from any harness with pan lane, and collect their reports.

    Docs: Lanes
  • Projects and polyrepos

    Add, clone or create projects. One project can span several repositories, with a git worktree in each per workspace.

    Docs: Project registry
  • Docker workspaces and local HTTPS

    Give each workspace its own Docker Compose stack, served over HTTPS at a *.localhost address.

    Docs: Docker and HTTPS setup
  • Order books Early

    A queue of issues with lanes, order and completion rules that Overdeck enforces.

    Docs: Order books
  • The Flywheel Early

    Works across issues: picks what to do next, starts the agents and watches each pull request. Today it is the loop Overdeck uses to build itself.

    Docs: The Flywheel vision
  • Backlog sequencer and parked issues

    Order the backlog the Flywheel works from, and see which issues it parked, why, and what releases them.

    Docs: Order books
An order book for the file explorer work: eight issues in Lane A, parallel up to 2, all landed. Lane B, strictly serial, is empty.

Harnesses and models

Use the coding agents and models you already have.

  • Harnesses

    Claude Code (the default), Codex, Pi, Kimi Code natively or over ACP, OpenCode, Grok Build, Muse Code and Prime Agent.

    Docs: Harnesses
  • Model routing and presets

    Set a model for plan, work, review, test and the other roles through three workhorse slots, or apply a preset: Anthropic defaults, an Anthropic cost-saver pilot, or OpenAI defaults, which are not yet eval-tested.

    Docs: Model presets
  • Tiered execution Opt-in

    Plan difficulty picks a crew model for each issue, and a frontier model supervises the cheaper tiers.

    Docs: Model presets
  • Permission modes

    New installs run Claude Code with an auto-approve hook instead of skipping every permission check.

    Docs: Conversations
  • Harness updates and version checks

    npm-installed harness CLIs update automatically, and Overdeck refuses a launch on a CLI too old for the model.

    Docs: Harnesses
  • Effort

    Choose the reasoning effort for each conversation and each role.

    Docs: Effort
  • Local models Experimental

    Run Claude Code agents on a model served by a local Ollama, with no provider API key.

    Docs: Local models
  • Workers

    pan worker run hands one brief to a separate, issue-linked agent and prints its report.

    Docs: Workers
  • MCP server for agents

    Claude Code, Codex, OpenCode and Kimi Code (over ACP) agents start with Overdeck's MCP server. Other harnesses get the matching pan commands. A project can add its own MCP servers.

    Docs: MCP server
Settings, Model Routing: preset buttons, three workhorse slots set to Claude Opus 5.5, Claude Opus 5.5 and Claude Sonnet 5.5, and the Plan role set to the expensive slot.

Machines, devices and access

Run agents on any machine you link and reach them from anywhere.

  • Overdeck Anywhere

    Link your machines. Each lists the other's conversations next to its own, and Pull here moves a running Claude Code or Codex conversation to the machine you are on.

    Docs: Linked machines
  • SSH machines Testing

    Start or adopt Overdeck on a machine you reach over SSH, or install a lasting service with npx @overdeck/core connect.

    Docs: pan machines
  • Fly.io machines

    pan machines add-fly starts a Fly.io VM that runs Overdeck and links it. You pay Fly directly.

    Docs: Linked machines
  • Fleet updates

    pan machines versions shows what each linked machine runs, and Update all waits for running turns.

    Docs: Linked machines
  • Remote access

    Pair a laptop, tablet or second desktop with a machine's dashboard. Each device gets its own revocable session.

    Docs: Remote access
  • Scoped API tokens

    pan token create makes a bearer token with only the scopes a program needs.

    Docs: pan token
  • Tailscale Testing

    Turn on protected HTTPS access to the dashboard and find other Overdeck machines on your tailnet.

    Docs: Tailscale
  • Session Vault

    Saves your agents' conversations, encrypted, to a git remote you own, so another machine can continue them.

    Docs: Session Vault
  • Shared sessions Early

    Share a live Claude Code conversation by link. Viewers join from their own Overdeck, with npx, or in a browser, and you decide who gets in. Sharing needs an overdeck.ai account, and accounts are invite-only for now.

    Docs: Shared sessions
  • overdeck.ai account Early

    pan account login signs in with GitHub for the hosted directory and shared sessions. Accounts are invite-only for now.

    Docs: pan account
  • Phone web app

    The dashboard installs on a phone as a web app at /m: Needs you, conversations and the pipeline, with three-tap answers.

    Docs: Browser on the phone
  • iPhone and Android apps Testing

    Native apps that pair by QR code over the encrypted channel, with sealed push and an iPhone Live Activity. In TestFlight and Play internal testing.

    Docs: Phone
  • Push notifications

    Get a phone notification when an agent or conversation needs you, sent by your own machine.

    Docs: Push notifications
  • Desktop app

    A desktop app for Linux, macOS and Windows from GitHub Releases. It can also open another machine's dashboard.

    Docs: Downloads in the quickstart
The Pull here dialog: this machine becomes the owner, the other machine stops running the conversation, and uncommitted and unpushed changes come with it. Machine names are blurred.

Seeing and steering the work

Know what every agent is doing and what needs you.

  • Mission Control

    The top bar counts the agents working and what needs you. All projects lists every project's conversations and opens up to four side by side.

    Docs: Mission Control
  • Project home

    Actions, five facts, what needs you, the pipeline, and recent commits on main with the conversation or agent behind each one.

    Docs: Project home
  • Needs you

    Questions, plans to approve, permission prompts and decisions wait in one list, each with one action to answer or open it.

    Docs: Mission Control
  • Costs at API rates

    Each conversation row shows its changes and its cost. Costs are shown at API rates; plan costs may differ.

    Docs: Mission Control
  • Activity feed

    A feed of what agents and conversations did, across projects.

    Docs: Activity feed
  • Appearance and themes

    Light and dark color themes, two dashboard styles, and VS Code theme import.

    Docs: Appearance
  • Agents page Experimental

    What is live, what needs you and what is waiting in the pipeline, with finished work in History. Turn it on in Settings, Experimental.

    Docs: Agents page
  • More experimental views Experimental

    Costs, Metrics, Health, Resources and Sessions pages, shown when Settings, Experimental features is on.

    Docs: Mission Control
  • Context preview

    See the layered launch context an agent will get, at /context.

    Docs: Context layers
  • God View Experimental

    An ambient map of the live pipeline: where issues are moving, which hooks are firing and where work has gone stale. Turn it on in Settings, Experimental.

    Docs: God View
  • Hygiene check

    pan hygiene audits commits, pushes, pull requests, agents, branches, workspaces and disk in one report.

    Docs: Hygiene check
  • Safe shutdown

    pan shutdown lets every agent finish its current tool call, then stops everything and tells you it is safe to reboot.

    Docs: Safe shutdown
  • Emergency stop

    Stop every agent from the agents menu or with one keyboard shortcut.

    Docs: Keyboard shortcuts
  • Low-disk alerts

    Below 5 GiB and again below 2 GiB of free disk, Overdeck adds a Needs you item and sends a phone push.

    Docs: Push notifications
  • Recovery after restarts

    After a terminal-backend restart you can resume every conversation, and dead work agents resume on their own.

    Docs: After a terminal-backend restart
  • Resource governor

    Memory thresholds that scale with RAM, and CPU pressure checks, refuse new lanes and warn before agent starts. Conversations are never blocked.

    Docs: Memory governor
  • Terminal backend

    Every agent runs in a Herdr pane by default. tmux is still supported when you ask for it.

    Docs: Terminal backend
Four conversations side by side, each with its own composer, on two harnesses.
Costs appear on screen at API rates; plan costs may differ.

Skills, memory and knowledge

Skills, memory and rules for the agents Overdeck launches.

  • Skills

    pan sync installs Overdeck's skills for Claude Code, Codex and Pi.

    Docs: Skills
  • Skills on or off at every level

    Turn each skill on or off globally, per project, per issue or per conversation.

    Docs: Skills
  • Skill packs

    Try a third-party skill collection, such as Atlassian Rovo for Jira and Confluence, for one project, one issue or one conversation, pinned to a commit you approve. For Claude Code and Codex.

    Docs: Skill packs
  • Project memory

    Short facts committed in your repository, so every managed harness sees the same list on every machine.

    Docs: Project memory
  • Memory search

    Overdeck extracts observations from each workspace, and pan memory search finds past decisions and fixes.

    Docs: Project memory
  • Context layers

    Rules for every machine, one project or one machine, delivered only to sessions Overdeck launches.

    Docs: Context layers
  • Project knowledge

    Connect an Open Knowledge Format bundle. Agents get cited concepts in context, a local viewer shows the bundle, and /okf publish builds a static site.

    Docs: Project knowledge

Integrations and automation

Connect Overdeck to the rest of your tools.

  • Webhook triggers

    Let GitHub, Stripe, CI or a chat bot message a conversation or start one, through a secret URL on your machine.

    Docs: Webhook triggers
  • Connect a bot

    Give a chat bot or a hosted assistant scoped, revocable access to Overdeck through its MCP server.

    Docs: Connect a bot
  • Semgrep Opt-in

    Off by default. When on, the verification gate scans each branch and findings go to the security review. Optional switches add an MCP server and a scan after each Claude Code edit.

    Docs: Semgrep
  • Computer use Opt-in, Linux

    Agents can read and operate a local desktop app: the accessibility tree, clicks, typing and screenshots.

    Docs: Computer use
  • Commit-status CI

    CI that reports commit statuses, such as Dagger Cloud, counts as the test job.

    Docs: Auto-merge and CI
  • Linear sign-in for agents

    One click connects Linear's MCP server for your agents.

    Docs: Issue trackers
  • Deft Directive

    pan skills deft use <issue> runs one issue on Deft's own pipeline.

    Docs: Skill packs
  • Agent guardrails

    Agents cannot set operator-only labels, commit on the wrong branch or edit the primary checkout.

    Docs: Workspaces
  • Jev judgment client Opt-in

    Typed judgment calls to TypeSafe's Jev model. Off by default.

    Docs: Jev

Setup and help

Install, check and fix Overdeck.

  • Quickstart

    npx @overdeck/core starts Overdeck and opens the dashboard. Needs Node.js 22 or newer, git, and Claude Code signed in.

    Docs: Quickstart
  • pan doctor and pan diagnose

    pan doctor checks your system and dependencies. pan diagnose collects a redacted bundle and starts an agent on the problem.

    Docs: Core commands
  • The pan CLI

    Start agents, message them and move conversations from your shell.

    Docs: CLI reference
  • Telemetry

    What anonymous telemetry Overdeck sends and how to turn it off.

    Docs: Anonymous telemetry

Try it on your own code.

Free and open source. One command starts it. Read the quickstart.