Capy

Capy is a coding agent you hand tasks to and come back to as pull requests. Each task is a thread with its own Linux machine in the cloud, so 10 threads can work on 10 features at once and keep going after you close the laptop. Capy opens the PR, fixes failing CI and review comments, and merges when you say so. The desktop app can also run threads on your own Mac, PC or Linux server.

Most agentic IDEs, like Conductor or T3 Code, run Claude Code or Codex on your laptop in git worktrees. Capy runs its own agent, and by default on its own machines:

You can still bring the subscription you already pay for. Capy runs about 60 models from Anthropic, OpenAI, Google, xAI and open-weight labs, and connects to Claude Pro/Max, ChatGPT (Codex), GitHub Copilot and SuperGrok plans.

A thread is one conversation with one agent about one piece of work. You start it from New thread: pick the project, the model and where it runs, and describe the task.

The sidebar shows every thread with its status: Working, Waiting (on CI or a review), Failed or Idle, plus a Needs you flag when the agent has a question. You can write to a thread while it works in 3 ways:

A thread has tabs for Overview, Diffs, Files, Setup, Browser and its PRs, and since 0.8.0 you can split them into panes. Select text in a reply to Fork the thread, fork with a summary, or open a Side chat for a quick question that doesn't derail the main one.

A thread can start subagents for parts of the job, nested up to 3 levels. A read-only subagent shares the parent's machine; one that writes code gets a fresh machine and can open its own PR. They run side by side or one after another, and each can use a different model.

A Captain is a long-lived agent for work bigger than one thread: a feature that takes 8 PRs, the bug queue, an area of the codebase. You tell it what it's for, and it starts threads, reads their reports and keeps a digest. The Needs you list at the top is the part to read: merges to approve, decisions to make, questions it can't answer alone.

Each Captain keeps its plan, rules and notes in its own Drive (AGENTS.md, CREW.md, NOTES.md, DIGEST.md, skills). Start one with New Captain, or turn a busy thread into one with Promote to Captain. In Slack, @Capy captain= sends a message straight to it.

Capy has a built-in issue tracker with For you, Assigned to me and Mentioned views, sub-issues and blocking relations. Captains file and pick up issues there, so you can see what the crew is doing without reading every thread. If your team lives in Linear, delegate a Linear issue to Capy instead.

Every PR Capy opens has its own page in the app, next to the thread that wrote it. The Diff tab shows the files with a Reviewed checkbox on each, and the thread stays open beside it, so you can ask "why did you change the retry delay?" and get an answer from the agent that made the change.

The Guide tab walks you through a PR in reading order, one chapter per change, with the before and after code. It helps most on large PRs, where reading the diff file by file hides why the changes go together.

When you're done, Review posts a GitHub review: Comment, Request changes or Approve. Merge offers squash, merge commit and rebase, Enable auto-merge to merge when checks pass, and Force merge. Stacked PRs work too, including GitHub's native stacks.

A few rules keep the agent in check:

Capy also has a review agent for any PR, written by people or other agents. It posts inline GitHub comments with a category (Bug, Risk, Maintainability, Refactor) and a severity, by default from Medium up. Turn it on per repo (Off, Once, Every push) or call it with an @capy review comment. It reads your AGENTS.md and flags code that breaks the rules written there.

Each project sets a machine size, and setup and dependencies are baked into a snapshot so new threads start ready to run tests. Machines bill by the hour while awake and sleep about 2 minutes after the thread stops working. A sleeping machine costs nothing.

All machines run in US East (New York). Each thread gets a live preview URL for the app it runs, like web-3000.preview.capy.ai, and Keep awake 4h holds a machine up while you test.

The desktop app for macOS, Windows and Linux pairs your computer as a device. In the composer, Run on picks the cloud, This Mac or another paired machine, and you choose a New worktree, an existing worktree or the current checkout. Local runs use no machine time.

A thread that started on your laptop can Continue on Cloud when you need to close the lid, and come back later. Staged, unstaged and untracked files move with it; gitignored files like .env stay on your computer. To turn a server or a spare Mac mini into a device without the desktop app, run the headless Dock CLI:

There's no native phone app. Open capy.ai in the phone's browser and add it to the Home Screen to get push notifications when a thread needs you.

The model picker lists about 60 models: Claude (Opus 5.5 is the default), GPT-6 and GPT-5.x, Gemini, Grok, Kimi, GLM, DeepSeek, Qwen and MiniMax. You pick a model per thread, and a subagent can use a different one. On Capy credits, models cost each provider's list price, with no markup.

To use a plan you already pay for, open Settings → Models → Add a connection:

When a subscription runs out mid-task, Capy asks whether to wait for the limit to reset or switch to credits.

Automations run a prompt on a cron schedule, an incoming webhook, a GitHub event or a Slack event, such as a reaction in a channel. You can group events so a burst of them starts one run, and keep the Slack posts quiet.

Other ways in:

On Pro and above, a thread can attach a cloud iPhone or Android phone through Revyl. Capy builds your app, installs it on the phone, taps through it and reads the logs, and you watch the phone right in the thread. Useful for React Native and native apps, where a cloud Linux machine can't run a simulator. A phone sleeps after 60 minutes.

Every plan includes unlimited members and all models. Plans differ in credits, machine sizes and how many machines run at once.

Annual billing takes 20% off. You can also pay as you go from a $5 balance with auto-reload. Unused credits don't roll over. YC companies get $1,000 in credits, and open-source projects can ask for a sponsored allowance. Capy is SOC 2 Type II and doesn't train on your private code.

Capy changes fast. In February 2026 it split work between a read-only Captain planner and a Build agent; in April Build was retired; in August the app was rebuilt around threads; in October Captains came back as orchestrators. Some of Capy's own comparison articles still describe the old flow, so trust the docs and the changelog over older reviews.

Cloud machines are Linux only and US East only, with no SSH access, no swap and no KVM, so Android emulators and multi-node Kubernetes don't run there. Unpushed work is lost if a machine is rebuilt, and a cloud workspace asleep for more than 14 days is cleared. There's no plan mode: ask for "a spec, then wait for my go-ahead" in the message or in your instructions.

Costs have 2 parts. Your subscription or credits pay for the model, and credits always pay for cloud machine time and reviews. A Large machine at $0.40 an hour is cheap per thread, but 20 threads that run CI loops all afternoon add up, so watch the usage page in the first weeks. Independent reviews are still rare; most comparisons online are Capy's own.

Codex cloud is the closest pick if you're all in on ChatGPT: cloud sandboxes, PRs, and no extra bill on top of your plan, but only OpenAI models. Devin offers the same "hand it a task, get a PR" loop with managed sessions and playbooks. Cursor's cloud agents fit if your team already works in Cursor. Capy's edge is the mix: any model or subscription, a machine per thread, Captains, and review and merge in one app.

If you'd rather keep agents on your own machines, T3 Code runs Claude Code, Codex and others from one window and drives them from your phone, and Herdr does the same in the terminal. Conductor was the popular Mac pick, but Meta bought it in November 2026 and is shutting down its cloud. For a side-by-side of all of them, see Top new agentic IDEs.