Capy

Detailed report

Capy coordinates work across your machines and cloud environments. It supports headless servers and browser control from phones, but coordination and transcript storage rely on its hosted service.

Research checked November 13, 2026. Most technical docs have no publication date. Dates below refer to explicitly dated releases or articles. This is document research, not a hands-on trial.

Product and workflow

The product is Capy at capy.ai, operated by Scrapybara, Inc. Its official site uses a capybara mascot and describes orchestrating coding agents across cloud environments and personal machines. The YC company profile connects Capy to Scrapybara and founders Nalin Semwal and Justin Sun, which helps distinguish it from namesakes.

Capy provides a delegation workflow. A long-lived Captain directs a crew of task threads, reads their reports, and surfaces decisions requiring you. Each thread can have further subagents. Plans and rules live in the Captain’s Drive so work can continue across turns. This could replace substantial manual coordination across terminals, laptops, and issue trackers. “Thousands of agents” remains a vendor scale claim, not capacity I independently verified. Captain documentation

Agents, models, and accounts

Capy primarily supplies its own agent harness, the software controlling model calls and tools, with models from Anthropic, OpenAI, Google, xAI, and other providers. Its catalog includes Claude, GPT, Gemini, Grok, Kimi, GLM, DeepSeek, Qwen, and more.

It supports direct API keys, gateways including OpenRouter and Bedrock, and custom OpenAI-compatible endpoints. Subscription connections include Codex, GitHub Copilot, SuperGrok, and Claude Code. Using a Codex subscription does not establish that the native Codex CLI runs underneath. Claude Code is explicitly different: its mode uses the Claude Agent SDK and can consume Claude Pro/Max allowances. Choosing an Anthropic model in Capy’s ordinary mode can still incur API or Capy-credit charges. Connected subscriptions retain their own limits. Bring-your-own-key tokens incur no Capy token charge. Models and connections

Your devices and headless servers

Desktop clients support macOS Intel/Apple Silicon, Windows x64, and Linux x64/ARM64. More relevant to you, Dock CLI explicitly turns your own headless server into a paired execution device. Linux requires glibc. Alpine/musl is unsupported. Dock also supports macOS and Windows, including Windows ARM64. First pairing is interactive. Afterward it can run under systemd, launchd, or tmux. It executes with its operating-system user’s permissions, so a dedicated server account is sensible. Your own devices incur no Capy machine-time charge. Desktop support, Dock CLI

This is documented device pairing, not a documented SSH-host manager. You might use SSH to install Dock, but I found no evidence that Capy manages ordinary SSH sessions or resumes arbitrary existing terminal agents. Separately, its Tailscale integration lets vendor cloud machines reach private services. That should not be confused with executing directly on your server. Tailscale architecture

Browser control includes phones. Local threads wait during disconnection for up to 24 hours and resume when the device reconnects. Transfers preserve conversations, branches, commits, and eligible uncommitted files, but go through Capy’s cloud. Ignored files, untracked files over 2 MB, and running processes do not transfer. Cloud checkpoints and automatic cloud handoff can be disabled. This is useful continuity, not live process migration. Persistence and transfers

Isolation, testing, and delivery

Local execution defaults to a fresh Git worktree, a separate checkout sharing the repository, although existing worktrees and the current checkout are options. This separates branches, but does not sandbox operating-system permissions.

Cloud threads receive separate Linux VMs. Subagents can share a parent machine or receive another. Sleeping disks persist, but running processes do not, and rebuilding a machine can lose unpushed edits and installed packages. Machine lifecycle

Capy can run tests, open GitHub PRs, fix failed CI, answer review feedback, and resolve conflicts. Its review interface supports merge controls and shows GitHub blockers. Scheduled and event-driven automations add ongoing maintenance workflows. These are meaningful orchestration features, although documented autonomy does not prove reliable results on your repository. PR workflow, Background work

Setup, ownership, and security

Setup consists of signing in, connecting repositories and model accounts, then configuring environments or pairing devices. Cloud handoff requires the GitHub App to access the relevant repositories. Quickstart

Capy stores transcripts centrally, including local-device threads. Anything an agent reads or prints can enter that transcript. Local checkpoints are enabled by default. Capy says it does not train its models on non-public content without agreement, claims SOC 2 Type II compliance, and offers Enterprise deployment in your cloud and negotiated zero retention. These are vendor statements. Request the audit report and deployment terms.

Team visibility deserves attention: new work is generally organization-visible unless configured otherwise, and the Pulls page has broader repository visibility than private thread access. Security and privacy

The platform is proprietary. I found no public self-hostable platform repository. Scrapybara’s MIT-licensed CodeCapy is a separate PR-testing bot. Neither that repository nor Capy’s “Open Source Program” makes the orchestration platform open source. License classification

Pricing checked November 13, 2026

Prices are USD, per organization, with unlimited members:

¹ Annual billing charges twelve months upfront. Lite offers seven days for $1. Max increases in $100 steps. Lite/Pro/Max allow 128 GB/512 GB/1 TB concurrent cloud RAM, with maximum standard machines of 4/8/16 CPUs respectively.

Credits pay for model tokens and cloud execution. This is not unlimited inference. Capy lists zero model markup. Cloud machines cost $0.10–$1.60/hour across standard sizes, plus two minutes of wind-down. A 128 GB option is $3.20/hour on request. Pro adds cloud mobile devices at $3.60/hour for Android or $5.40/hour for iPhone/iPad. bring-your-own-key or subscription use avoids Capy token charges, but cloud compute still costs credits. Official pricing

The docs mention free-plan bring-your-own-key access, while the pricing page emphasizes paid plans. Confirm free-account limits rather than assuming Lite is mandatory.

Development and user reports

Dated activity is substantial: Captains became generally available and Dock launched on October 31, 2026. Dock installation expanded on November 1. Version 0.8.0 on November 10 added split panes and access controls. Release notes also acknowledge failures involving concurrent subagents, stalled loading, and interrupted reviews. These demonstrate active repair, not a measured failure rate. Headless support is particularly new. Changelog

Independent first-hand evidence for Capy is limited. Targeted Reddit/Hacker News searches produced no useful attributable discussion, and no relevant public platform issue tracker surfaced. Somi lists zero reviews. A September 19, 2026 Lazyweb assessment identifies parallel, reviewable work as a possible benefit and raises the concern that throughput may simply enlarge the review queue, but explicitly is not a hands-on trial. Vendor testimonials and directories should not be mistaken for consensus.

Fit and limits

For this workflow, Capy merits a trial because it explicitly joins laptops, headless Linux servers, and browser supervision. Teams gain shared work, integrations, and no seat charges. The main concern is vendor dependence: local execution still relies on Capy for coordination and transcript storage. Avoid it if you require offline operation, an auditable self-hosted platform, or guaranteed continuity with existing native CLI sessions.

Practical trial

Pair two laptops and one dedicated Linux server account. Run three bounded tasks in separate worktrees, reconnect from a phone, interrupt Dock, and transfer one task between server and cloud. Check recovery, omitted files, permissions, PR quality, and total charges. Record the changes you accept and the review time they require. Unresolved questions are transport details, export completeness, Enterprise deployment boundaries, and reliability during prolonged outages.