Orca

Detailed report

Orca combines coding agents, your own compute, parallel workspaces and remote control in an open-source environment. Its server-owned runtime suits switching between devices. Remote and mobile features remain beta, and orchestration is experimental.

Research checked November 13, 2026. Most technical documentation is undated. Dated releases and reports are identified below. These are documented capabilities, not results from a hands-on trial.

Product and workflow

The product is Orca by Stably AI, at orca.dev and stablyai/orca. The official website and repository link to each other. This excludes unrelated security, research, and database products named Orca.

Orca’s appeal is consolidation: run several agents, give each task a separate Git checkout, inspect changes, operate a browser, and manage remote machines without constantly switching applications. For this workflow, a useful arrangement is an always-on Linux runtime that both laptops and a phone control. The agents and repository stay put while you change devices. That architecture is explicitly documented, rather than inferred from “remote” marketing. Product overview, remote-server documentation.

Agents, models, and payment

Orca launches command-line agents. Its configured integrations include Claude Code, Codex, Cursor CLI, OpenCode, Pi, Grok, Gemini, Copilot, Amp, Goose, Cline, Kimi, and others. “Any CLI agent” means it can launch that program. It does not mean every agent gets equal chat rendering, account management, usage reporting, or orchestration support. Claude and Codex have deeper integration. Supported-agent matrix.

You bring existing agent subscriptions or configure credentials through the underlying agent. Orca is not a bundled model subscription. Model availability therefore follows the agent, provider, account, and execution host. Supervised workers support explicit model selection for several integrations, but there is no universal model catalogue guaranteed across agents. Orca also documents linking a GLM Coding Plan using its API key. Orchestration, usage and account tracking.

The published application price is $0: free, MIT-licensed software. I found no published paid seat requirement for ordinary desktop or self-hosted use. Agent subscriptions, API consumption, cloud infrastructure, and networking services remain separate costs. Enterprise rollout and support are contact-based, with no public price established in the reviewed material. Official site, license, enterprise offering.

Your laptops, devices, and headless servers

Desktop downloads cover macOS (Intel and Apple Silicon), Windows, and Linux. Mobile companions support iOS and Android. Android is distributed as an APK. Installation, mobile documentation.

Three remote arrangements matter:

• SSH worktrees: Orca on your laptop controls selected workspaces on your server. Worktrees and agent processes execute remotely. Editing and review remain in the client. Documentation promises reattachment after disconnection. SSH architecture.
• Remote Orca Server: The remote machine owns projects, credentials, terminals, and session state. Multiple desktop, browser, mobile, and automation clients can connect to that runtime. Headless Linux is explicitly supported through orca serve, including a systemd user-service installation path. Server architecture.
• Per-workspace environments: Repository recipes create disposable VMs, sandboxes, or containers using your provider account. Orca does not sell managed VPS hosting. This feature is experimental and requires lifecycle scripts and authentication setup. Execution modes.

For your use case, choose the server-owned runtime for continuity across laptops. SSH is better when one laptop should coordinate several execution hosts.

Persistence has clear limits. A background terminal daemon keeps agent processes alive across application quits and updates. A host reboot or daemon failure ends those processes, although saved layouts and scrollback return. This is persistent execution while the host survives, not durable computation across machine failure. Session restoration.

Mobile supports replies, terminal access, workspace creation, source-control review, and account switching across connected hosts. It remains a companion rather than a full editor. Direct access and Relay are distinct paths. Relay requires sign-in. Mobile capabilities and pairing.

Isolation, orchestration, review, and tests

A Git worktree is another checkout of a repository. Orca uses these to separate agents’ file changes, making parallel tasks or competing implementations practical. Its editor, browser, and diff-comment workflow help you inspect results before accepting them. Workspace model.

Structured orchestration adds task dependencies, supervised workers, coordinator messages, completion tracking, and decision gates. It is explicitly experimental. The documentation also says its experimental panel and switch are temporarily unavailable and older orchestration commands are retired. These workflows are still changing. Orchestration reference.

Tests run through the agents and your project’s tools. Orca does not establish correctness automatically. It integrates hosted reviews, CI logs, and failed-check handoffs to agents. GitHub has the deepest support, with other hosting integrations documented at varying depths. Reviews, issues, and CI.

Setup, security, and teams

Basic setup means installing Orca, opening a repository, and authenticating your agent CLIs. Remote setup adds credentials and tools on the execution server. Laptop authentication does not automatically transfer. The recommended server connection uses Tailscale or another private network. Pairing links grant runtime access, and individual client grants can be revoked. Remote setup and access controls.

The main concern is the security model. Orca documents permission-bypass defaults for supported agents, with settings to restore manual approval. Separate worktrees do not isolate processes from host files, credentials, or networks. For consequential repositories, use restricted accounts or disposable environments and review launch permissions. Permission defaults.

Orca says model traffic goes directly to configured providers. Its telemetry documentation describes anonymous product events sent to US-region PostHog, excluding prompts, source contents, and terminal output, with opt-out controls. That telemetry statement should not be mistaken for a complete audit of Relay or every integration. Telemetry policy, enterprise data claims.

Teams gain shared runtime access, repeatable environments, and familiar Git review. I did not establish comprehensive role-based permissions, tenant separation, or certified compliance. “SOC 2 Readiness” is not evidence of completed certification. Treat a shared runtime as a trusted engineering machine.

Development and user reports

The official changelog lists v1.4.257 on November 12, 2026, following releases on November 9 and 10. Development is active. Nevertheless, remote servers and mobile remain beta, and orchestration remains experimental. Release speed does not prove reliability. Changelog.

Independent evidence is encouraging but limited:

• An approximately October 2026 Reddit discussion praises Orca for keeping repositories organised across hosts. Its author also needed custom rebasing, authentication, and server-startup scripts. That supports usefulness while exposing setup work. Multi-machine account.
• An October discussion praises the sidebar, mobile access, usage display, and integrated workflow. Other commenters question sandboxing. These are individual reactions in a repository-promotion thread, not representative survey evidence. Community discussion.
• Issue #20802, opened September 15, 2026, reports reconnect banners and dropped terminal input despite a live runtime. It is now closed and linked to follow-up work. It demonstrates a relevant historical failure, not a proven current defect. Issue and investigation.

The available reports do not establish broader user satisfaction.

Fit and practical trial

Choose Orca if you want extensive integrated tools, existing subscriptions, and control over your own servers. Avoid making it critical infrastructure immediately if you require stable remote behaviour, strict access separation, or minimal maintenance.

Trial it on one non-sensitive repository: run Claude and Codex in separate worktrees on headless Linux, disconnect laptop A, reconnect from laptop B, then reply from your phone. Test host reboot recovery separately. Complete a real test-and-PR cycle, check credential exposure, and measure reconnect failures and model usage before expanding.