Yep Anywhere lets you supervise AI coding agents from your phone. The last update here was the remote device control launch in early March. Since then we've shipped dozens of releases without stopping to write about any of them. This is the spring and early-summer catch-up post.

The "we" is new, too: Yep Anywhere now has a second core contributor, Jonathan Graehl. The extra pair of hands is a big part of why the spring and early summer moved this fast.

Codex grew up

Codex started as the rough sibling next to Claude Code. It's now a first-class provider with the same workflow:

  • Streaming, diffs, and approvals at parity. Live token deltas, apply_patch edits rendered as proper diffs, and tool approvals all work the way they do for Claude Code — in the stream and in replayed history.
  • Subagents render. Codex spawn_agent subagents expand inline so you can follow what a delegated agent is doing instead of staring at an opaque tool call.
  • Steering that holds up. Interrupts, queued messages, mid-flight steering, and reconnect merging were hardened so a flaky phone connection doesn't desync your session.
  • The small stuff. Collapsed Bash rows, local image previews, and rate-limit exhaustion surfaced as real errors instead of silent stalls.

We track the Codex CLI closely — protocol compatibility has moved forward through 0.142.4, including newer rollout formats. If you live in Codex, it should now feel like Yep Anywhere was built for it.

Switch any session to any model

You can change the model mid-session, from any session, to any model the provider offers — with /model or the model selector. Start a task on a fast, cheap model, hit something hard, and switch up to a stronger one without losing context or starting over. It works across providers, so the same control applies whether you're driving Claude Code, Codex, or anything else.

Share a session with a link

You can publish a read-only view of any session as a public link — with share controls, live viewer counts, and origin-aware gating so you decide exactly what's exposed. Good for showing a teammate what an agent did, filing a bug report with full context, or saving a transcript you're proud of.

Talk to your agents

Voice input grew from a single browser-only button into a real pipeline. Speech now routes through the server, so you can pick a backend that fits your setup:

  • Browser-native — zero setup, works everywhere.
  • Deepgram and xAI (Grok) — bring your own key for faster, more accurate transcription.
  • Local Whisper, Parakeet, and NeMo — fully on-device, nothing leaves your machine.

The speech options moved into the microphone menu so switching backends is one tap, and provider keys are scoped to transcription so they don't leak into your agents' environments or get billed against the wrong thing.

Now in six languages

The client UI is fully localized into English, Chinese, Spanish, French, German, and Japanese, with a language selector in Appearance settings. Locale bundles load lazily, so the languages you don't use cost you nothing.

Find it, recap it, stay oriented

Long-running, multi-session work needed better navigation tools, so we added a batch of them:

  • Session search — jump straight to the session you mean instead of scrolling the sidebar.
  • Session recaps — come back to a session that ran while you were away and get a summary of what happened.
  • Heartbeat turns — keep an agent moving on long tasks, with queued "when done" prompts.
  • Richer transcripts — KaTeX math, ANSI color output, expandable tool rows, in-transcript reverse search, and one-tap markdown copy.
  • Explicit thinking controls — off / auto / on, plus per-provider effort, so you decide how hard the model works.

Queue work at the project level

Session queueing is useful when you know what the current agent should do after its turn ends. Project Queue handles the bigger case: "start this follow-up only after the whole project is quiet."

Queue a message or a new session for a project, refresh the browser, close your phone, even restart the server — the queue is server-owned and durable. The Projects page shows the backlog, lets you edit, cancel, retry, or move an item to the top, and shows what is blocking dispatch. The session view and sidebar also surface targeted Project Queue items so queued work is not hidden.

It's deliberately conservative. The composer affordance is hidden by default, queued text is delivered verbatim, and Project Queue promotes one item per project-idle boundary instead of dumping a full backlog into an agent at once. New hosted clients also gate the UI on server compatibility so older servers don't expose controls they can't honor.

Read any file — and choose which ones

There's a shared file viewer now: click a file path anywhere — in a session, a public share, or a remote link — and get a syntax-highlighted view, including specific line ranges and Markdown media.

Opening files remotely is exactly the kind of capability that needs a leash, so it ships locked down. Settings → File access lets you scope which folders the file viewer may read — project folders, uploads, temp, your home directory, or a custom allow-list. By default, the file route refuses to serve arbitrary absolute paths until you explicitly grant the folder. Secure by default, opened up only where you decide.

Safer source control from the phone

The Source Control page is still intentionally narrow — not a full Git client — but it now covers the operations that matter when an agent has done useful work and you're away from the desktop.

  • Check remote fetches remote-tracking state only when you ask.
  • Pull attempts a safe fast-forward update and stops with clear guidance when the branch has diverged or local changes would be overwritten.
  • Push publishes completed work, including simple branch publishing to origin.
  • Better review shows recent commits and a split diff preview on wide screens, while keeping the modal flow on phones.

More models, more providers

Beyond Claude Code and Codex, the release window added:

  • Grok Build via ACP, with prompt suggestions and effort controls.
  • Ollama for fully local models, with a customizable system prompt.
  • OpenCode direct session reading, resumable ses_* URLs, permission prompts, thinking/tool rendering, image input, graceful interrupts, and vLLM response rendering.

And the Claude side stayed current — the app now targets Claude Agent SDK 0.3.199 and tracks the current Claude Code runtime assumptions.

Windows keeps getting better

Windows has come a long way since March. We fixed drive-letter and backslash path handling, spawn failures, Codex session discovery, temp-directory resolution, and a string of mixed-slash deduplication bugs. If you tried Yep Anywhere on Windows months ago and bounced off, it's worth another look.

Quieter, safer, faster

A lot of the release window went into work you won't see directly:

  • Security hardening — XSS-hardened markdown rendering, path containment across file/upload/share routes, a relay origin allowlist, approval audit logging, unsafe-Unicode visibility in approval prompts, and stronger SRP resume verification (with a compatibility grace window so older clients aren't cut off abruptly).
  • Leaner connections — API responses are gzip/deflate compressed, and idle agent sessions get reaped so a closed tab doesn't leave a provider process running forever.
  • Less friction — saved new-session defaults (provider, model, permission mode), output appearance controls, and lifecycle webhooks.

Coming soon: a one-click desktop app

Claude and Codex both have excellent desktop apps, and Yep Anywhere already works alongside them. We're building a desktop app of our own — not to replace them, but to make Yep itself easier to run. No fiddling with Node and npm: download it, click install, and it keeps itself up to date automatically. It's coming soon.

Get the update

Everything above (except the desktop app) is in Yep Anywhere 0.6.0. If you're on the npm install:

npm update -g yepanywhere

Then restart yepanywhere. New here?

Install Yep Anywhere