Moshi vs Paseo

Of the new wave of agent development environments, Paseo is the one closest in spirit to Moshi. It's built by one person — Mohamed Boudra, who says he turned down funding to keep it independent. It runs on your machines, not in someone's cloud. It ships native mobile apps, a voice mode, and a desktop app, and it takes privacy seriously: no telemetry, no forced log-in, and a relay that's end-to-end encrypted and off until you turn it on. If you're reading this page, you probably like the same things we do.

The difference is architectural, and it's the same fork we wrote about in Moshi vs. ADEs. Paseo is a control plane: its own daemon starts your agents, holds their sessions, and serves its own clients. Moshi is a layer: your agents live in tmux or Herdr, and Moshi draws chat, diffs, approvals and a phone terminal on top — over your own SSH.

Neither bet is wrong. They optimize for different things. Let's go through it honestly.

Both rosters cover the agents most people actually run. Paseo drives Claude Code, Codex, OpenCode, Pi and OMP with dedicated providers, GitHub Copilot through the Agent Client Protocol, and dozens more through ACP and its plugin system. Moshi works with anything that runs in a terminal; moshi-hook understands 19 agents well enough to feed the inbox, and 17 of them also get Chat View, Copilot included. Paseo renders every agent it drives as chat. The bigger difference is how each reaches the agent — and that's the rest of this page.

The short version

Read it as two clusters. The rows about ownership and weight — who holds the agent, whether a plain terminal can attach, how big the download is — favour Moshi. The rows about orchestration and openness — agents that spawn agents, schedules, an SDK, an open-source license — favour Paseo.

Who holds your agent?

This is the question that decides most of the trade, so it goes first.

Paseo runs a local server it calls the daemon. The desktop app starts it automatically; on a server you run paseo from the CLI or the Docker image. The daemon launches each agent through a provider — a dedicated adapter for Claude (built on the Claude Agent SDK), Codex, OpenCode, Pi and OMP, or the Agent Client Protocol for others like Copilot — and persists the agent record, timeline and workspace to disk under $PASEO_HOME. Every Paseo client, from the phone to the CLI, talks to that daemon over WebSocket.

Paseo does this well. An agent whose runtime is closed stays resumable: prompt it again and the daemon reloads the provider session under the same ID, with its history intact. But the live session is Paseo's. It lives in Paseo's process, in Paseo's format, and you reach it through Paseo's clients. If you stop using Paseo, you don't take a running agent with you — you start a new session somewhere else.

Moshi never owns the process. You start claude or codex in a tmux window or a Herdr pane exactly as you would without Moshi. moshi-hook reads what those sessions are doing and exposes it — transcripts, approvals, a live PTY, dev-server ports — as small web APIs bound to 127.0.0.1. Close Moshi, uninstall it, and the agent keeps running. Open Ghostty at your desk, tmux attach, and you're in the same session your phone was just looking at. Herdr works the same way; Rex, Mitchell Hashimoto's upcoming multiplexer, is next on our list once it ships.

One session, desk to phone

Because the session is a normal multiplexer session, continuity comes for free. Start a long Codex run in tmux on your Mac, walk away, and pick it up from Moshi on your phone: the same pane, the same scrollback, the same running turn. Come back and tmux attach — nothing re-synced, no handoff, one session several clients can hold at once.

Paseo gets cross-device continuity a different way, and it's genuinely good at it: every client subscribes to the daemon, so the same agent shows up on desktop, phone, web and CLI at once. The difference is only in which clients. With Paseo, it's Paseo's. With Moshi, it's any terminal on earth, plus Moshi.

Two kinds of chat

Both apps show your agent as a chat rather than a wall of terminal output, but the chat comes from different places.

Paseo's chat is its native interface. The daemon drives each agent through a provider adapter and records a structured timeline of messages, tool calls and results, which every Paseo client renders. It's polished, and because the daemon owns the session, the chat is the session.

Moshi's Chat View shipped in June 2026, with moshi-hook v0.2.27 adding transcript streaming on June 14 and Moshi 3.6.0 for iOS. It's a lens, not a session. The agent keeps running in its tmux or Herdr pane; Chat View renders that live session's transcript as native chat, and the real terminal is one tap away. The same session is still there for tmux attach from your laptop. Chat View is still experimental and part of Pro, with free uses to try it. It covers 10 agents: Claude Code, Codex, OpenCode, Pi, OMP, Kimi, Grok, Hermes, Cursor and Antigravity.

11 MB vs. 179 MB

We measured both macOS downloads on September 28, 2026. Moshi Desktop 0.4.6 is an 11.4 MB DMG on Apple Silicon. Paseo 0.9.2 is 178.6 MB — about 16 times larger. Installed, the gap widens: roughly 24 MB for Moshi Desktop against roughly 504 MB for Paseo.

The gap has a simple cause. Paseo's desktop app is built on Electron, so it ships its own copy of Chromium and Node, and it bundles the daemon that runs your agents. Moshi Desktop is a Tauri app that uses the webview macOS already has, and it doesn't need a daemon of its own, because tmux or Herdr is already doing that job. On Linux and Windows, Moshi Desktop is the same UI served in your browser, with nothing to install.

Disk space isn't really the point. Size is a rough measure of how much of your workflow an app owns, and a daemon-plus-clients app is, by design, going to own more. See the full table across T3 Code, Orca, Paseo, Codex and Claude in Moshi vs. ADEs.

Paseo's real strengths

These are the reasons Paseo is worth comparing to at all.

• It's free and open source (Apache-2.0). No subscription for the core product and no account. You can read every line of the daemon, the relay and the apps.
• The relay is end-to-end encrypted and optional. Pairing uses Curve25519 key exchange and NaCl-sealed payloads, so the relay only routes bytes it can't read. It's disabled until you pair a device, and you can skip it entirely by connecting directly over TCP, Tailscale or another VPN, or by tunnelling to the daemon over SSH from the desktop app. The public relay is run by Paseo at relay.paseo.sh from open-source Elixir code, and per Paseo's security docs it sees only IP addresses, timing, message sizes and session IDs. You can also run your own.
• Voice mode that talks back. By default speech never leaves your own hardware: the phone streams audio to the daemon, which runs Parakeet for speech-to-text and Kokoro for text-to-speech on your computer's CPU. OpenAI's speech models are an opt-in. You can talk a problem through with the agent hands-free, not just dictate.
• A capable mobile app. A real terminal with text selection and copy-paste, image and file attachments, a diff pane and push notifications, on iOS and on Android (Google Play or a direct APK).
• Real orchestration. Agents can spawn subagents through Paseo's MCP tools, work in isolated git worktrees, run on schedules, and be driven from a CLI (paseo run, paseo attach, paseo send) or a TypeScript SDK. There's also a built-in browser for agents to use.
• Every surface. Desktop, iOS, Android, a self-hostable web UI, a CLI and Docker, all against one daemon.
• Privacy by default. Paseo says it has no telemetry and no tracking, and it doesn't make you log in.
• Team features if you need them. Paseo Hub coordinates agents across several daemons, with triggers from GitHub, Slack and Discord. Self-hosted Hub is free and open source; hosted Hub is €15 per seat per month, with a free trial.

If orchestration, openness and talk-back voice are what you're optimizing for, Paseo is an excellent choice. The rest of this page is for people weighing that against owning the session layer.

Direct SSH vs. a relay

Paseo gives you a choice of transport, which is to its credit. A phone reaches the daemon either through Paseo's relay — opt-in, encrypted end-to-end, and blind to content — or directly over Tailscale or a VPN. The phone app doesn't speak SSH; SSH tunnelling is for Paseo's desktop app and CLI, which can reach a daemon on another machine that way.

Moshi has one transport: your own SSH or Mosh connection to your machine, the same one you'd make from a laptop. There's no relay in the session path. The terminal, the transcripts, the diffs and the dev-server previews all travel phone-to-host. The one thing that does cross Moshi's servers is the push notification: when an agent needs you, moshi-hook sends an alert through Apple's or Google's push service, because phones only accept push from a server.

Mosh also gives Moshi resilience a relay doesn't have to solve the same way. Sleep the phone, switch from Wi-Fi to cellular, and Mosh catches the session up when you come back, without a persistent socket draining the battery.

One feed, every agent and host

Paseo organizes work into workspaces and agent tabs, with subagents tracked under their parent — a model that shines when you're orchestrating. Moshi's inbox is built for triage instead: one feed across every host and project, where a finished turn floats up, a pending approval pins to the top, and stale items archive themselves. The same agent on your Mac and on a Linux box stay separate rows.

Native iOS depth

Both apps push a notification when an agent needs you. Paseo promises full feature parity between its mobile and desktop apps, which should include approvals, though we couldn't find a doc that shows approving from the phone. Around that loop, Moshi goes deeper into the platform:

• Live Activity & Dynamic Island. The active turn sits on your lock screen.
• Apple Watch. Approve or deny from your wrist.
• Usage rings. Claude Code's 5-hour and 7-day windows, Codex's rolling windows and OpenCode's provider limits, drawn as rings in the app and on a Watch complication. (Paseo shows provider usage too, as balance bars.)
• Image and file paste. Pick a screenshot or a PDF, and Moshi copies it to your host over SCP and hands the agent a path. (Paseo's mobile app attaches images and files too.)
• On-phone voice. Dictation with Parakeet, Apple's SpeechAnalyzer or a local Whisper, running on the phone itself, so it works even when the phone can't reach your computer. It's one-way — you speak, it types. Paseo takes the other local-first route: recognition and speech run on your computer, and its voice mode can talk back. Both keep your voice off third-party servers by default.

Review beyond the diff

Moshi's diff viewer includes a Browse tab that walks the whole repo at the working tree or any past commit, opening files as syntax-highlighted source, all through the gateway tunnelled over your SSH session. Its browser preview opens dev servers running on your host the same way. Paseo's mobile app has a diff pane of its own, so both let you review from the phone; Moshi's extra is the whole-repo browse.

On the desktop, Moshi Desktop adds more review tools on top of the same sessions:

• Git Diff and History tabs. A commit graph with ref chips, and the current branch with its ahead/behind count in the sidebar. It's for looking, not committing: there's no staging or commit UI. Paseo goes further here, with a commit graph of its own plus git actions and a branch switcher.
• Comment on code. Hover a line in a diff or file, press +, and your note travels to the agent with your next message.
• Design comments on previews. Pin numbered comments to elements of a running web page; when the framework's dev tools expose it, each comment carries the source file and line, so the agent knows exactly what to change.
• Side-by-side previews of your dev servers, reached through the same SSH tunnel.

Paseo has its own built-in browser, which the agents can also drive as a tool, and on desktop you can annotate or screenshot a page element and send it to the agent with a comment. So both tools close the loop from "this button looks wrong" to the code. Moshi's version pins numbered comments that follow the element as the page moves and carries the source file and line when the framework exposes it; Paseo's goes further in letting agents use the browser themselves.

Where Paseo wins

To keep it balanced, pick Paseo when:

• Free and open source are non-negotiable. Moshi's desktop app is free, but the product is closed source and the phone's agent layer is paid.
• You want orchestration. Subagents, schedules, worktree isolation, a CLI and an SDK for scripting agents are Paseo's home turf; Moshi deliberately doesn't orchestrate.
• You want voice you can converse with, running on your own computer, not just dictation.
• You'd rather not manage SSH. Pair with a QR code and go through the encrypted relay.
• You want team workflows — agents triggered from GitHub, Slack or Discord through Paseo Hub, self-hosted for free or hosted at €15 per seat per month.

Pricing

Paseo's core is free; it's funded by GitHub Sponsors, Open Collective and Paseo Hub. Hosted Hub costs €15 per seat per month, billed monthly, after a free trial. Seats count Hub members and pending invites, not people who only trigger agents from GitHub, Slack or Discord. Self-hosting Hub is free. Moshi charges for the phone's agent layer and Pro features, and uniquely lets you buy it once instead of subscribing. Either way, you still pay Anthropic or OpenAI for the agent itself.

As for why Moshi isn't open source: we wrote about that in Moshi vs. ADEs. The short version is that everything holding your work — tmux, Herdr, SSH, Mosh — already is.

The verdict

Paseo and Moshi are chasing the same freedom — agents on your own machines, reachable from anywhere — from opposite ends. Paseo builds the whole control plane and gives it to you as open source. Moshi leaves the control plane to tmux and Herdr and builds the best layer it can on top. If you want the full picture across T3 Code, Orca, Codex and Claude, read Moshi vs. ADEs; for the phone-first agent clients, see Moshi vs Happy.