Paseo

Detailed report

Paseo provides laptop and phone access to agents running on machines you control. Its server-owned background service suits an always-on development host. Reconnects, long histories and provider integration remain the main reliability concerns to test.

Research checked November 13, 2026. Most official documentation is undated. Dates below identify releases or reports where available.

Product and workflow

The product is Paseo at paseo.sh, backed by getpaseo/paseo on GitHub. The repository also contains its official logo asset.

Paseo lets you supervise agents on your existing development machines through one interface. It uses your existing subscriptions, credentials and tools. It combines agent conversations, files, terminals, diffs, and pull requests. Its daemon, the background service managing agents, owns the environment while clients connect to it. That is useful when your development environment lives on an always-on server. Official product rationale.

Agents, models, and payment

The technical documentation lists native integrations for Claude Code, Codex, OpenCode, Pi, Antigravity, and Muse Code. Its ACP catalog adds agents including Copilot, Cursor, Gemini CLI, Amp, Factory Droid, Qwen Code, and Kimi Code. ACP is a protocol for connecting agent interfaces. Catalog availability does not establish equal feature support: Antigravity, for example, lacks Paseo steering/MCP tools and runs with approval prompts bypassed. Models depend on each configured agent and account. Supported providers.

Paseo supports existing subscriptions as well as bring-your-own API keys, subject to the underlying agent. Its docs say Claude Code uses normal Claude plan limits. Codex supports ChatGPT subscription authentication or separately billed OpenAI API usage. Paseo adds no Codex usage charge. These are documented integration arrangements, not a blanket guarantee about every provider’s account terms. Claude integration, Codex integration.

Your machines, remote access, and persistence

Desktop downloads cover macOS 13+, Windows, and Linux. Mobile clients cover iOS and Android, alongside web and CLI access. Headless operation is explicitly documented: install the daemon on a server, VM, or development box. Docker can run the daemon and self-hosted web interface, but its image requires you to add agent CLIs and credentials. These are your execution environments, rather than mandatory vendor cloud sandboxes. Downloads, server setup.

Desktop and CLI support SSH through your local OpenSSH configuration. SSH connects to an already-running remote daemon. It does not install or start one. Authentication is non-interactive and key-based. Phones use the optional encrypted relay or direct access through Tailscale/VPN. Multiple clients can therefore reach the same host without copying its development environment between laptops. Connectivity documentation.

Closing a client and reconnecting is architecturally different from stopping the daemon. Provider session restoration also varies: the provider API supports persisted session identifiers and history replay, but Antigravity retains its conversation without replaying the old Paseo transcript after restart. Treat “reconnect,” “resume after daemon restart,” and “survive host reboot” as separate acceptance tests. I found no universal guarantee of uninterrupted execution across all three. Provider persistence contract.

Workflow and orchestration

A workspace can use an existing directory or a managed Git worktree, a separate checkout and branch. Multiple agent sessions may share one workspace, so isolation is a choice, not an automatic property of every agent. Workspace model.

The documented workflow covers parallel tasks, diff inspection, editing, terminals, tests, browser checks, and opening PRs. Worktrees separate files, but processes still share host network ports. The desktop browser can support agent-driven verification. Passing tests remains something you or an agent must actually run. Parallel-development guide.

This goes beyond displaying several chats: agents can launch other agents across providers, exchange prompts, inspect progress, collect results, create worktrees, and operate across reachable hosts. Schedules and periodic continuation prompts are documented. Built-in orchestration tools require enabling MCP injection. The CLI offers another route. Orchestration documentation.

Setup, data, and security

The desktop app bundles its daemon. On servers, install Paseo plus authenticated agent CLIs. Install and authenticate gh for PR-related features. Local state defaults to ~/.paseo. The project is Apache-2.0 licensed and says it has no telemetry, tracking, or mandatory Paseo login. Plugins execute trusted code with access to the daemon machine and connected clients. Setup, repository.

The relay is optional and documented as end-to-end encrypted, although it can observe connection metadata. Direct access defaults to localhost. Network exposure needs deliberate authentication and transport protection. “Code stays on your machine” describes Paseo’s execution architecture: cloud model providers can still receive material through their agents. Self-hosting therefore does not imply offline inference. Security architecture.

Pricing and team implications

The core application is free and open source. Optional Hub connects GitHub, Slack, and Discord triggers to agents on your machines:

• Self-hosted Hub: free software that you operate.
• Hosted Free: $0/month, one seat, 50 agent runs monthly.
• Hosted Pro: $15 per seat/month, unlimited agent runs. Seats include members and pending invitations. Trigger-only users do not count.

“Unlimited runs” does not include unlimited model tokens or free compute. Your agent subscriptions/API bills and machine costs remain separate. Official Hub pricing.

Hub offers useful team automation, but the site still labels daemon provisioning and shared settings/skills as future work. Its security documentation is candid: Hub does not sandbox agents or neutralize malicious external requests. Execution permission can reach existing workspaces and configure provider permission modes. Teams should evaluate permissions and host isolation before connecting shared infrastructure. I did not establish enterprise support guarantees or comprehensive administrative controls. Hub security.

Development and user reports

This is active, pre-1.0 software. The download page showed stable 0.15.4. 0.16.0-beta.2, dated November 11, 2026, includes faster diff loading, rebuilt mobile timelines, and background-task controls. The repository says one person leads the project, funded by users. Open source reduces dependence on continued vendor access, but does not remove maintenance concentration. Releases, project description.

Independent evidence is encouraging but sparse. In a maintainer-started Reddit thread, a separate user praises custom Pi compatibility, Android remote access, and relay reliability. The retrieved page labels that comment “1mo ago”. It did not expose an exact date. This is one useful experience, not consensus. User report.

Concrete negative reports deserve equal weight:

• June 22–July 12, 2026: broken older chats/imports, later closed after the reporter believed it fixed. Issue #1659.
• September 18: a maintainer response describes a reconnect improvement and remaining detection gaps. Issue #3842.
• October 11: a user reports image-heavy OpenCode sessions repeatedly reconnecting through the official relay. Issue #2220.

These establish failure modes, not their prevalence. Homepage testimonials are vendor-selected evidence.

Fit and practical trial

Choose Paseo if cross-device supervision of your own persistent machines is central, and you accept maintaining the environment. Avoid making it critical infrastructure immediately if reconnect reliability or guaranteed support is essential.

Trial it on one disposable Linux development account: connect two laptops and a phone, run Claude and Codex in separate worktrees, switch Wi-Fi/cellular, disconnect clients, then restart the daemon. Check history, permissions, running commands, diffs, tests, PR creation, and model charges. This tests continuity against the reported reliability concerns.