Capy
Scrapybara, Inc. · Capy: Cloud Coding Agent with Parallel Subagents, PR Review, and 50+ Models
Capy is a cloud coding agent that takes a task from the web app, desktop app, Slack, Linear, or its API, works on an isolated Ubuntu VM with your repositories cloned, and opens a pull request that it keeps fixing when CI fails or reviewers comment. It suits engineering teams that want background agents across many models on one shared credit balance, with no per-seat fees.
Best for
Engineering teams on GitHub that want to hand backlog tickets, migrations, and recurring maintenance to cloud agents across many models, with one shared credit balance, unlimited members, and pull requests that follow up on CI and review on their own
Not ideal for
Teams on GitLab or Bitbucket, organizations that must self-host or keep execution inside their own cloud, developers who want to run Claude models on a Claude subscription, and anyone who needs a predictable flat monthly cost
Who it's for
Engineering teams and technical founders on GitHub who want to delegate coding, review, and maintenance work to cloud agents and choose models per task
Capabilities
- Each thread runs on its own isolated Ubuntu 24.04 cloud VM with passwordless sudo, Docker, Node.js, Python, git, the GitHub CLI, and a desktop with Chrome for browser use and screenshots
- Subagents split work up to three levels deep, either sharing the parent's machine for read-only work or booting fresh machines for writes, and run in parallel or as stacked pull requests
- Opens pull requests from its own capy/ branches, then wakes on failed CI checks, human review comments, merge conflicts, and merges, without you forwarding anything
- Review agent that reviews pull requests in a real checkout, posts findings inline on GitHub by category and severity, and follows the rules in your AGENTS.md files
- More than 50 models on Capy credits, including Claude Opus 5.5 (the default), GPT-6 Astra, Gemini 3.8 Flash, Grok 4.7, Kimi K3, GLM-5.3, DeepSeek V4 Pro, and Qwen3.8 Max, with a model chosen per thread and per subagent
- Connect a ChatGPT (Codex), GitHub Copilot, or SuperGrok subscription, provider API keys, or gateways such as OpenRouter, Vercel AI Gateway, and Cloudflare AI Gateway, and Capy uses them in the order you set before falling back to credits
- Automations triggered by cron schedules, GitHub, Slack, and Linear events, incoming webhooks, or on demand, with an optional plain-language filter that decides whether each event should run
- Start work from the web app, the desktop app, a Slack mention, a delegated Linear issue, an @capy comment on a pull request, or the REST API with streaming
- Desktop app for macOS, Windows, and Linux that can run threads in a local folder on your Mac or Windows PC, move a thread from a Mac to the cloud and back, and import Claude Code and Codex session history
- Setup scripts and VM snapshots that restore a prepared machine in about one to two seconds, plus encrypted project secrets, MCP servers, and Tailscale access to private services
- GitHub credentials never sit on the machine: git and gh calls go through Capy's proxy using short-lived grants, and Capy never merges unless you ask in the conversation
Limitations
- There is no free cloud tier: cloud machines and Capy-hosted models are paid from credits, and the cheapest plan is Capy Lite at $20 a month after a $1 first week
- Usage is billed by model tokens and hourly machine time, so monthly cost depends on which models and machine sizes you use and for how long
- A Claude subscription cannot be connected, because Anthropic's policy does not allow subscription sign-in in third-party apps, so Claude models run on Capy credits, an Anthropic API key, or, for some older Claude models, a GitHub Copilot subscription
- API keys and gateways require a paid plan, starting with Capy Lite
- No self-hosted or on-premises deployment is documented: the agent loop runs on Capy's servers and machines run in Capy's cloud
- Running processes do not survive sleep, which starts about two minutes after a thread stops, and unpushed work is lost if a machine is deleted after 30 days asleep
- GitHub is the only documented code host, so teams on GitLab or Bitbucket cannot use the pull request workflow
- SSO/SAML and audit logging are limited to the custom-priced Enterprise plan
Use cases
- Delegating a Linear issue to Capy and getting back a pull request that was built and tested on a cloud VM
- Splitting a framework migration into parallel subagents, one per directory, each shipping its own pull request
- Turning on automatic review for a repository so every push gets inline findings, with Capy's own pull requests fixing high-severity findings before they reach you
- Running a nightly automation that audits dependencies for security advisories and opens upgrade pull requests
- Mentioning @Capy on a bug report in Slack so it triages the issue, writes the fix, and reports back in the same Slack thread
- Using a ChatGPT or Copilot subscription the team already pays for to run GPT models in Capy without spending Capy credits
Our take
Capy is for teams that want background coding agents but also want to choose and mix models themselves. Pricing is credits with no per-seat fees, and you can run models on subscriptions you already pay for, so a large team can try it without a seat negotiation. The flip side is that spend follows tokens and machine hours, so set auto-reload limits and watch which models your threads default to. The workflow details are mature: subagents on separate machines, stacked pull requests, a review loop that triages its own findings, and a git proxy that keeps GitHub tokens off the VM. The constraints that rule it out are mostly structural: GitHub only, no self-hosting, and no way to use a Claude subscription.
Who should use it
Engineering teams on GitHub that want to hand tickets, migrations, reviews, and scheduled maintenance to cloud agents, especially teams that switch between Claude, GPT, Gemini, and open-weight models or already pay for ChatGPT, Copilot, or SuperGrok, and teams that want to add many members without per-seat costs.
Who should skip it
Teams whose code lives on GitLab or Bitbucket, organizations that need self-hosted execution, developers who want to run Claude models on a Claude subscription, and teams that need a fixed monthly cost rather than usage-based credits.
Strengths
- Large model catalog across many labs, chosen per thread and per subagent
- No per-seat fees: every plan has unlimited members
- Can run models on existing Codex, Copilot, or SuperGrok subscriptions or your own keys
- Pull requests follow up on CI failures, review findings, and comments on their own
- Desktop app can run threads on your own computer instead of a cloud VM
Weaknesses
- Costs depend on model tokens and machine hours, so spend is harder to predict
- No free cloud tier beyond a $1 first week of Capy Lite
- Claude subscriptions cannot be connected
- GitHub only, and no self-hosted option
Capy pricing
Capy Lite
$20
Billed monthly
- $20 of credits per month
- Unlimited members
- Every supported model, plus your own API keys and gateways
- Cloud machines up to 4 CPU / 16 GB, 128 GB concurrent RAM
- VM snapshots, $16/month billed annually
Capy Pro
$100
Billed monthly
- $105 of credits per month (+5%)
- Cloud machines up to 8 CPU / 32 GB, 512 GB concurrent RAM
- Mobile builds and devices
- $80/month billed annually
Capy Max
$200–$1,000
Billed monthly
- $100 steps, each with 110% of the price in credits
- Cloud machines up to 16 CPU / 64 GB, 1 TB concurrent RAM
- Priority support and Slack Connect
- $160–$800/month billed annually
Enterprise
Custom
- Custom credit quota and invoicing
- SSO / SAML
- Audit logging
- Dedicated support
Free tier limits: No free cloud plan. The first seven days of monthly Capy Lite cost $1 with a verified card (one intro per card and organization). Local threads on a connected subscription in a personal workspace use no Capy credits or machine time, an organization with no plan can run a zero-cost model such as MiMo-V2.6 Pro on its own Mac, and open-source projects can apply for free access.
Note: Credits are one dollar balance shared by the organization and cover model tokens at per-model rates and machine time while a VM is awake: $0.10/hour (Small, 1 vCPU) to $1.60/hour (Hyper, 16 vCPU), with Large ($0.40/hour) as the default. Unused plan credits do not roll over. Top-ups are 1:1 with a $5 minimum, and auto-reload is available. Usage on a connected Codex, Copilot, or SuperGrok subscription or on your own API key does not draw on the balance.
Technical specs
Context window
Up to 1.05M tokens, depending on the model
API pricing
Threads started through the API are billable runs like other threads: model tokens and hourly machine time from the organization's Capy credit balance, unless a connected subscription or key covers the model
Available models
Where Capy excels
Clearing backlog tickets from Linear
Delegating an issue starts a Capy thread that reports its plan and progress back into Linear, works on a prepared cloud VM, and opens a pull request that keeps responding to CI and review, so small tickets get done without an engineer setting up a local session.
Large migrations split across subagents
Capy can split a migration by directory into subagents on separate machines, each shipping its own pull request, or stack dependent changes so every pull request shows only its own diff.
Automated review on every push
With automatic review on, the review agent posts severity-ranked findings inline on GitHub and enforces conventions written in AGENTS.md, and on pull requests Capy opened, the thread fixes high-severity findings itself.
Capy vs. competitors
Capy vs. Devin
Devin is Cognition's autonomous software engineer: it takes work from Slack, Microsoft Teams, Linear, Jira, the web app, or the API and runs it in its own cloud VM. Capy covers the same delegate-to-a-cloud-agent workflow from Slack, Linear, GitHub comments, its apps, or the API. The differences are that Capy lets you pick from more than 50 models per thread or subagent, bills one shared credit balance with unlimited members, and can run models on connected Codex, Copilot, or SuperGrok subscriptions.
Capy vs. Replicas
Replicas hosts third-party harnesses such as Claude Code, Codex, and Cursor in cloud VMs, on your own credentials, and charges for VM runtime by seat and by the minute. Capy runs its own agent across many models and charges model tokens and machine hours from one credit balance, with optional subscriptions and keys. Replicas fits teams that want to keep their chosen agent CLIs, and Capy fits teams that want one agent with a wide model choice.
Capy vs. Factory Droid
Factory's Droids run the same agent across a desktop app, CLI, IDE integrations, CI, and Slack, with Missions for planned multi-step work. Capy is centered on cloud threads, each with its own VM and subagents, plus a review agent and event-driven automations. Its desktop app can also run threads locally. Factory fits teams that want the agent inside their editors and terminals, and Capy fits teams that mainly delegate work to cloud machines.
Capy vs. OpenAI Codex
Codex is OpenAI's coding agent across a CLI, IDE extensions, cloud tasks, and ChatGPT, and it runs on OpenAI models. Capy can run GPT models on a connected ChatGPT (Codex) subscription as well, and adds Claude, Gemini, Grok, and open-weight models, subagents on separate machines, Slack and Linear delegation, and automations. Codex fits teams standardized on OpenAI, and Capy fits teams that want to mix providers.
Frequently asked questions
What is Capy?
Capy is a cloud coding agent. You describe a task in a thread, from the web or desktop app, Slack, Linear, or the API. Capy starts an isolated cloud VM with your GitHub repositories, makes and tests the change, opens a pull request, and keeps working on it when CI fails or reviewers comment.
How much does Capy cost?
Capy Lite is $20 a month with $20 of credits, Capy Pro is $100 a month with $105 of credits, and Capy Max runs from $200 to $1,000 a month in $100 steps, with 110% of the price in credits. Annual billing is 20% off, and Enterprise is custom. Credits pay for model tokens and hourly machine time, from $0.10 an hour for the smallest machine to $1.60 an hour for Hyper. Every plan has unlimited members, and the first seven days of Lite cost $1.
Is there a free version of Capy?
There is no free cloud plan. The first seven days of monthly Capy Lite cost $1, then $20 a month. Free paths are narrow: in a personal workspace, local threads in the desktop app that run on a connected Codex, Copilot, or SuperGrok subscription use no Capy credits, an organization without a plan can run a zero-cost model on its own Mac, and open-source projects can apply for free access.
Which models does Capy support?
Capy's pricing docs listed 58 models in October 2026, including Claude Opus 5.5 (the default), Claude Sonnet 5.5, GPT-6 Astra, GPT-6 Sol, Gemini 3.8 Flash, Grok 4.7, Kimi K3, GLM-5.3, DeepSeek V4 Pro, Qwen3.8 Max, MiniMax M3, and Meta's Muse Spark. You can pick a model per thread and per subagent.
Can I use my ChatGPT or Claude subscription with Capy?
You can connect a ChatGPT (Codex), GitHub Copilot, or SuperGrok subscription, and the models it covers run on that plan instead of Capy credits. You cannot connect a Claude subscription, because Anthropic's policy does not allow subscription sign-in in third-party apps. Claude models run on Capy credits or on your own Anthropic API key, which needs a paid Capy plan, and a Copilot subscription covers some older Claude models.
Does Capy run on my computer or in the cloud?
Both. By default each thread runs on an isolated cloud VM. With the desktop app for macOS, Windows, or Linux installed, a thread can instead work in a folder on your own Mac or Windows PC with your git credentials, and a thread can move from your Mac to the cloud and back.
Capy vs Devin: what is the difference?
Both take tickets from Slack, Linear, or an API and work in their own cloud VMs to open pull requests. Capy puts the whole organization on one shared credit balance with unlimited members, lets you choose among more than 50 models per thread or subagent, and can run models on connected Codex, Copilot, or SuperGrok subscriptions. Devin is Cognition's own agent and also works with Microsoft Teams and Jira.
Integrations & fit
Alternatives to consider
Cognition
Engineering teams with a steady backlog of well-scoped tickets who want to delegate them asynchronously from Slack, Teams, or their issue tracker and review the resulting PRs
Paid
Replicas
Engineering teams already using Claude Code, Codex, or Cursor who want those agents running in configured cloud VMs, triggered from Slack, Linear, GitHub, GitLab, or automations, with PRs that follow up on CI failures and reviews
Paid
Factory
Engineering organizations that want one model-agnostic agent platform across local coding, CI automation, code review, and long-running missions, with enterprise deployment and governance options
Paid
OpenAI
Developers and teams already paying for ChatGPT Plus, Pro, Business, or Enterprise who want one coding agent across the terminal, IDE, and parallel cloud tasks
FreemiumAbout Capy
Capy is built around threads: durable conversations with one agent that keep running on Capy's servers after you close the laptop. Each thread gets its own cloud machine, an Ubuntu 24.04 VM with Docker, Node, Python, and a desktop with Chrome, and the agent can split large work into subagents. Subagents that write code boot their own fresh machines, so parallel changes cannot collide, and they can run side by side or stack one pull request on top of another. When the work is done, Capy pushes to a capy/ branch and opens the pull request. It then wakes up on its own when a CI check fails, when a reviewer comments, or when the pull request merges. A built-in review agent posts findings inline on GitHub, and the thread that owns the pull request sorts them and fixes the serious ones before reporting back. Automations run prompts on a cron schedule or on GitHub, Slack, Linear, and webhook events. What sets Capy apart from other cloud agents is the commercial model. It runs its own agent, not third-party CLIs, across more than 50 models from Anthropic, OpenAI, Google, xAI, Moonshot, Z.ai, DeepSeek, Qwen, MiniMax, Xiaomi, Meta, and others. All usage draws from one dollar balance shared by the whole organization, and plans have unlimited members. You can also connect a ChatGPT (Codex), GitHub Copilot, or SuperGrok subscription, or your own API keys, so those models do not use Capy credits. Capy relaunched as V2 in August 2026, and its desktop app for macOS, Windows, and Linux, released in September 2026, can also run threads in a local folder on your Mac or Windows PC, and a thread can move from a Mac to the cloud and back. Plans run from Capy Lite at $20 a month to Capy Max at up to $1,000, with cloud machines billed hourly from the same credits. The tradeoffs: cloud work always costs credits, model and machine costs vary with use, a Claude subscription cannot be connected, and there is no self-hosted option.
Updates from Capy
Capy 0.4.3 renamed a thread's child agents to subagents across the app, CLI, Slack, and docs, and GPT-6 models running on a Codex subscription now use their full 1.05M-token context window instead of 400K. The desktop app also starts faster.
Capy 0.4.0 lets threads run on your own subscriptions, API keys, coding plans, and gateways in a set order before falling back to Capy credits, adds reviewing and merging stacked pull requests inside Capy, and made Claude Opus 5.5 the default model.
Capy 0.2.0 added Windows and Linux builds of the desktop app, which first shipped for macOS on September 7. Threads can run in a folder on your Mac or Windows PC with your own git credentials, and setup can import Claude Code and Codex histories.
Capy relaunched as V2, a cloud coding agent in which each thread gets its own cloud machine with your repositories cloned, opens the pull request, and wakes on CI, review, and merge events. Work can start from the web app, Slack, Linear, or the API.
Are you the founder? Claim this listing →