Mobile control plane for a fleet of AI agents.
Direct agents from your phone. A steward turns your goal into a plan, a fleet executes across your own hardware, you ratify and review.
Android, iOS, iPadOS — open source, self-hosted, Apache 2.0.
Every other "Claude Code on your phone" tool is a 1:1 bridge to a single local session. TermiPod is a 1:N control plane over a fleet:
- Multi-agent — coordinate Claude Code, Codex, Aider, or any CLI, each with its own pane, profile, and budget. One inbox for all of them.
- Multi-host — a steward on a $5/mo VPS can delegate work to a GPU box at home behind NAT, via A2A over a reverse-tunnel relay. Nothing else spans hosts cleanly.
- Director, not operator — write a natural-language goal. The steward decomposes it into a plan. You ratify. You don't author DAGs, you don't babysit terminals.
- Governance built-in — budget caps, policy overrides, per-agent usage, immutable audit log, team roles. No other mobile agent tool has these.
- Offline-first — SQLite snapshot cache shows last-known-good lists on every screen when the hub is unreachable. Subway-safe.
- Vendor-agnostic, self-hosted — Apache 2.0 Go hub you run yourself. No account gating, no cloud relay, no vendor lock-in.
30-second demo. Type ablation sweep on nanoGPT, tell me which optimizer scales better on your phone. Your VPS steward returns a 6-step plan; you tap Approve. Six training runs spawn on your home GPU box. Three hours later, a briefing document with loss curves lands on your Me tab. You ratify it on the train home.
See docs/discussions/positioning.md for the full thesis, ICP, and competitive analysis. The full doc index lives at docs/README.md.
⚠️ Outdated — pre-IA-redesign, tmux-era screens. These predate the v1.0.175–182 IA redesign (Projects · Activity · Me · Hosts · Settings) and the v1.0.281+ steward chat work. They still show the SSH/tmux side accurately, but don't reflect the hub workflow which is now the primary surface. Refresh deferred until after the Candidate-A hardware demo lands — capturing screenshots before the demo's final UX polish would just need re-doing. The refresh itself is planned to be CI-automated; design captured in docs/discussions/screenshot-automation.md.
| Dashboard | Agent Commands | Key Palette |
 |
 |
 |
| Terminal | Vault (Keys & Snippets) | Insert Menu |
 |
 |
 |
AI agents produce ten times more output than any human can review. The moment you run more than one agent, or span more than one machine, the existing mobile tools fall apart — session-bridge apps (Claude Code Remote Control, Happy, Tactic Remote) give you one session in your pocket; messenger-bridge agents (OpenClaw, Hermes, Claude Code Channels) give you one chat-style assistant across your chat apps. Neither is a cockpit for a fleet. TermiPod is built on a different axiom: the human is a director, not an operator.
| Dimension | Remote Control / Happy / Tactic | TermiPod |
|---|---|---|
| Topology | 1 phone ↔ 1 session ↔ 1 host | 1 director ↔ N agents ↔ M hosts |
| Agent count | One active at a time | Fleet; steward spawns more |
| Host span | Single local machine, keep it awake | VPS + GPU + laptop, coordinated via A2A |
| Agent vendor | Claude Code (+ Codex) | Agent-agnostic — any CLI that takes a pty |
| Authoring model | You type messages | You write a goal; steward decomposes it |
| Governance | None | Policies, budgets, audit log, team roles |
| Data ownership | Cloud relay or laptop-only | Hub holds names/events; hosts hold bytes |
| Offline | Needs live relay | SQLite snapshot cache — last-known-good on every list |
| Open source | Happy: yes; others: no | Apache 2.0, self-hosted Go hub |
| Solo ML researchers | Nightly sweeps across a VPS + home GPU box — kick off from the phone, review briefings in the morning |
| Indie AI hackers | Multiple agent CLIs across projects — unified inbox, attention queue, per-pane profiles |
| Autonomy-focused startups (1–5 eng) | Budget caps, policy overrides, audit log, team roles — no other mobile agent tool has these |
| Open-source maintainers | Run triage / review agents overnight; approve agent PRs from bed |
| Homelab / self-hosters | Pull work off the laptop onto the phone without exposing raw SSH to the internet |
Be honest: if you run exactly one Claude Code session on one machine and your laptop is always on, use Anthropic's Remote Control or Happy. TermiPod's governance, multi-host, and self-hosting features have a setup cost. You shouldn't pay it unless you need what they buy.
- Ed25519 / RSA keys — generate on-device (RSA 2048 / 3072 / 4096) or import, stored in Android Keystore / iOS Keychain with optional passphrase, one-tap public-key copy
- SSH ProxyJump — connect through bastion/jump hosts to internal machines
- SOCKS5 proxy — route through corporate proxies, VPNs, or Shadowsocks/Clash
- Raw PTY mode — direct shell access for servers without tmux, with a one-tap shortcut from any tmux connection card
- Connection testing — verify SSH + tmux before saving
- Auto-reconnect with exponential backoff — up to 5 retries; commands you type while disconnected are queued and flushed automatically once the link is back
- Latency indicator — live ping in the header (color-coded: green < 100 ms, red > 500 ms) so you know whether the lag is your fingers or the network
- Adaptive polling — refresh rate ramps from 50 ms (active) down to 500 ms (idle) to save battery
- Background connection service — Android foreground service keeps SSH alive while the app is backgrounded; optional keep-screen-on for long sessions
- Dashboard — recent sessions sorted by last access with relative timestamps ("Just now", "5 min ago"); one tap reconnects and restores the last window + pane
- Visual navigation — breadcrumb header: tap Session > Window > Pane to switch
- Pane layout view — accurate proportional split visualization, tap any pane to focus
- Two-finger swipe between panes
- Pinch to zoom the terminal (50%–500%) for quick readability bumps
- Copy / scroll mode — toggle to select text without the screen jumping; updates buffer until you exit and copy lands in the system clipboard
- Create / rename / close sessions and windows
- Bell / Activity / Silence alerts — tmux window flags monitored across all connections; tap any alert to jump straight to that window and pane (alert auto-clears)
- 256-color ANSI terminal rendering with auto-extend scrollback
| Component | What it does |
|---|---|
| Action bar | Swipeable button groups per profile — ESC, Tab, Ctrl+C, arrows, one tap away |
| Compose bar | Multi-line text field with send button. Multi-line input ships as one bracketed paste so AI agents and shells see the block intact, not N separate commands. Long-press send to omit Enter |
| Direct Input mode | Real-time keystroke streaming with a live indicator — every tap goes straight to the pty, ideal for vim, less, htop, REPLs |
| Custom keyboard | Flutter-native QWERTY with Ctrl/Alt/Esc/arrows. Built-in live key strip (Home / End / PgUp / PgDn / Del + pulse indicator) replaces the wasted compose-row gap. Arrow row auto-hides when nav pad / joystick is on. Toggle off entirely for CJK / voice input |
| Navigation pad | D-pad, joystick, or gesture surface for arrow keys + action buttons |
| Snippets | Slash commands with dropdowns for enums, text fields for free-form args. Long-press the bolt key to stash the current compose text as a draft snippet |
| Modifier keys | Ctrl / Alt as toggle buttons — tap to arm, double-tap to lock |
4 built-in profiles — Claude Code, Codex, General Terminal, tmux — each with optimized button groups. Create custom profiles for any CLI. Each pane remembers its profile and auto-detects from pane_current_command.
- SFTP upload/download with progress tracking and remote directory browser
- Image transfer with format conversion, resize presets, and path injection
Opt-in coordination layer for teams running multiple AI agents across machines. Paste a hub URL + bearer token in Settings → Hub and the five-tab IA — Projects · Activity · Me · Hosts · Settings — comes alive.
Research-demo workflow: write a project directive on the phone → a steward agent decomposes it into a plan → workers on GPU hosts execute runs in parallel via cross-host A2A → a briefing agent summarizes overnight into a reviewable document. Every step surfaces on the phone; you ratify, not operate. The MVP demo target is the locked ablation-sweep template (nanoGPT-Shakespeare optimizer × size; see docs/decisions/001-locked-candidate-a.md); paper-reproduction and benchmark-comparison templates are deferred until the demo lands.
Me (center, default) — your personal triage: pending approvals, urgent tasks, recent activity digest, and a Vault shortcut. Tap a pending approval to Approve / Reject inline.
Projects — project inventory; tap any row for the Project Detail surface with Overview, Tasks, Plans, Runs, Reviews, Outputs (run-produced checkpoints / curves / reports), Documents, Blobs, Channels, Schedules, and an Agents section walking agent_spawns for a parent→child org chart. FAB spawns workers via YAML with template picker, host picker, and saved presets; the Steward gets its own spawn flow. Metrics digests from trackio / wandb / TensorBoard auto-surface as inline sparklines on Run Detail. Project name / goal / template / docs root / budget are all editable inline.
Activity — team-wide audit feed (audit_events promoted): policy changes, template edits, agent lifecycle, channel posts, run state transitions — all in one filterable timeline.
Hosts — host-runner check-ins with last-seen timestamps; hosts behind NAT publish agent-cards to the hub directory and accept peer A2A calls via a reverse-tunnel relay so steward agents on a VPS can invoke workers on GPU machines end-to-end.
Team screen (header icon on Projects) — Members, Policies, team-scope channels (including the #hub-meta steward room, reachable from the AppBar chip), and Team Settings with cron Schedules, per-agent Usage / budget rollups, an Audit log, and the Templates browser for team-wide agent / prompt / policy YAML. Everything that drives behaviour — project templates, agent skills, launcher commands — is data you can edit on disk; no code changes required to add a new agent kind.
The hub itself ships as a separate Go daemon under hub/ — install with go install or run from source. See docs/how-to/install-hub-server.md for hub setup, docs/how-to/install-host-runner.md for adding a worker host, and docs/how-to/run-the-demo.md for the no-GPU dress rehearsal.
- Data export / import — full JSON backup of connections, keys, snippets, history, and settings; restore on a new device or migrate from the legacy MuxPod app
- Built-in file browser — manage SFTP downloads and app storage from Settings, share or delete files in place
- Update checker — Settings → Check for updates queries GitHub releases and links to the latest APK
- Help & onboarding — cheat sheet for action bar + tmux keybindings, 4-card walkthrough
- Deep linking —
termipod://connect?server=<id>&session=<n>&window=<n>&pane=<i>opens a specific server / session / window / pane from external apps. Each connection has a stable Deep Link ID (set in Edit) so URLs survive renames; pairs with claude-telegram-notify to tap a Telegram alert straight into the right pane. Legacymuxpod://URLs still resolve - Tablet & foldable adaptive layout
- i18n — English and Chinese, follows system locale
Against mobile agent control tools (session-bridge category):
| Feature | TermiPod | Claude Code Remote Control | Happy Coder | Tactic Remote |
|---|---|---|---|---|
| Multi-agent fleet | Yes | One session | One per CLI wrapper | One session |
| Multi-host (A2A) | Yes — VPS + GPU + NAT via relay | No | No | No |
| Agent-agnostic | Yes — any CLI | Claude Code only | Claude Code + Codex | Claude Code + Codex |
| Director / steward model | Yes — steward decomposes goals | No | No | No |
| Governance (budget, audit, policy) | Yes | No | No | No |
| Self-hosted hub | Yes (Go daemon) | Cloud relay (Anthropic API) | Relay server | Cloudflare Tunnel |
| Offline snapshot cache | Yes (SQLite) | No | No | No |
| Platform | Android + iOS + iPad | iOS + Android + Web | iOS + Android + Web | iOS only |
| Pricing | Free, Apache 2.0 | Claude Max ($100+) | Free, open source | Paid trial |
Against messenger-bridge agents (different thesis, overlapping user question):
| Feature | TermiPod | OpenClaw | Hermes Agent | Claude Code Channels |
|---|---|---|---|---|
| UI surface | Purpose-built mobile app | Existing messengers (WhatsApp, Telegram, Slack, Signal, iMessage, Discord, 15+) | Telegram / Discord / Slack / WhatsApp / Signal | Telegram + Discord |
| Primary use case | Fleet cockpit + research ops | Personal assistant across messengers | Personal assistant + self-improving skills | Remote Claude Code via chat |
| Agent topology | Director → steward → fleet | One agent, cross-platform memory | One self-improving agent | One local Claude Code session |
| Multi-host / A2A | Yes | No | No | No |
| Plan ratification | Structured plan + Approve | Conversational | Conversational | Conversational |
| Governance UI | Budgets, policy, audit | None (chat only) | None (chat only) | None |
| Rich UI (plots, kanban, runs) | Native | Text only | Text only | Text only |
| Offline snapshot | Yes | No | No | No |
| Open source | Apache 2.0 | MIT | Open | Plugin repo |
| Vendor lock-in | None | None | None | Requires Claude Pro/Max |
When to pick which: Use a messenger bridge when you want a personal agent answering in the apps you already use 20 hours/day. Use TermiPod when you're directing a fleet across your own infrastructure and need purpose-built screens for plans, runs, reviews, governance. The two coexist happily on the same phone.
Against mobile SSH clients (TermiPod also covers this layer — breakglass, not product):
| Feature | TermiPod | Termux | JuiceSSH | Termius | ConnectBot |
|---|---|---|---|---|---|
| Platform | Android + iOS + iPad | Android | Android | Multi | Android |
| tmux integration | Native visual | Manual CLI | None | None | None |
| AI agent profiles | Claude Code + Codex, per-pane | None | None | None | None |
| SSH jump host | Built-in | CLI | CLI | Built-in | None |
| SOCKS5 proxy | Built-in | CLI | None | None | None |
| File transfer | SFTP with UI | Local FS | None | SFTP | None |
| Custom keyboard | Flutter-native | None | None | None | None |
| Open source | Yes (Apache 2.0) | Yes | No | No | Yes |
Android: Download the latest APK from Releases and install.
iOS / iPadOS: Build from source with Xcode (see below). TestFlight is on the roadmap.
git clone https://github.com/physercoe/termipod.git
cd termipod
flutter pub get
# Android
flutter build apk --release
# iOS / iPadOS (requires macOS + Xcode)
flutter build ios --release- Add a host — Tap + on the Hosts tab, enter host / port / username
- Authenticate — Password or SSH key (generate via the Keys tile on the host detail screen)
- Optional — Configure jump host or SOCKS5 proxy in the connection form
- Navigate — Expand host > session > window > pane
- Interact — Action bar for quick keys, compose bar for commands, bolt button for snippets, [+] for file transfers
| Component | Requirement |
|---|---|
| Device | Android 8.0+ (API 26), iOS 13.0+, iPadOS 13.0+ |
| Server | Any SSH server (OpenSSH, Dropbear, etc.) |
| tmux | Any version (tested 2.9+) — optional with raw PTY mode |
| Network | Direct SSH, or via jump host / SOCKS5 proxy |
The MVP target is the research demo described in docs/spine/blueprint.md §9 Phase 4: user writes a directive → steward decomposes → fleet executes runs across hosts → briefing agent summarizes overnight → user reviews on phone.
Phase status (as of v1.0.318):
| Phase | Status |
|---|---|
| P0 — Hub primitives (schema) | ✅ Shipped |
| P1 — Structured wire (protocols) | ✅ Shipped |
| P2 — App UI | ✅ Shipped |
| P3 — Integrations (trackio, A2A relay) | ✅ Shipped |
| P4 — Research demo | 🟡 Backend feature-complete; hardware run remaining |
The demo path is shipped end-to-end on the dress-rehearsal harness
(seed-demo + mock-trainer, no GPU needed). The actual hardware
run of Candidate A (nanoGPT-Shakespeare optimizer × size sweep) is
the MVP milestone — gated on two consecutive walkthrough-clean
device tests.
Still open (post-demo):
- Briefing agent overnight schedule — steward schedules the briefing autonomously (today the user fires it).
- iOS TestFlight / App Store distribution — Android APK ships; iOS builds are local-only today.
- A2A peer auth — per-agent tokens on the reverse-tunnel relay so cross-team calls can be authenticated end-to-end.
- Domain packs / marketplace — content-pack extensibility (post-MVP; see docs/discussions/post-mvp-domain-packs.md).
Live trackers:
- docs/roadmap.md — Now / Next / Later
- docs/plans/research-demo-gaps.md — demo-scoped detail
- docs/changelog.md — what shipped per release
- docs/decisions/ — ADRs (numbered, append-only)
TermiPod is a derivative work of MuxPod by @moezakura (Copyright 2025 mox), licensed under the Apache License 2.0. The original MuxPod provided SSH connectivity and basic tmux session viewing for Android. TermiPod has since diverged substantially with cross-platform support, input UX redesign, agent profiles, SFTP, ProxyJump, SOCKS5, custom keyboard, and more. TermiPod is an independent project, not affiliated with the original author. See NOTICE for full attribution.
Found a bug or have a feature request? Open an issue or send feedback from Settings > About in the app.
Built with Flutter. Designed for mobile. Made for developers who live in the terminal.