Where to Run Your Coding Agent: 20 Tools Compared

TL;DR

Every coding agent product answers one question: which machine runs the agent. Cloud platforms like Devin and Google Jules run it on theirs, which is fast to start and cannot reach your VPN, your staging database or your private package registry. Desktop orchestrators like Sculptor and Traycer run it on yours, which fixes the access problem and stops the moment your laptop sleeps. A control plane on a machine you own does both, and it is the only option nobody can switch off. Seven products in this category shut down, were acquired, or abandoned the positioning in the last eighteen months, so pick with that in mind.

---

Compare by where the agent runs, not by feature list

Feature grids in this category are mostly noise. Every product plans, edits, runs tests and opens a pull request. They differ on something more basic: the machine the agent is actually sitting on when it does that work.

That one decision sets everything else. It decides whether the agent can read your .env. It decides whether it can curl your staging API. It decides whether the work survives you closing your laptop, and whether the whole thing keeps existing if the vendor gets acquired.

So here is the field sorted that way.

Runs in the vendor’s cloud. Fastest to start, cannot see your machine.

Product	Open source	Reaches your VPN	Entry price
Devin	No	Not on default tier	Free, then $20/mo
Google Jules	No	Not documented	Free with Google AI plans
OpenAI Codex Cloud	No	Not documented	Included with ChatGPT plans
Charlie Labs	No	Not documented	Free, then $50/mo
Blitzy	No	Inbound-only by design	$50k pilot
Tembo	Control plane only	Yes, via Tailscale	Free, then $60/mo

Runs on your desktop. Full access to your environment, stops when the machine sleeps.

Product	Open source	Mobile	Entry price
Sculptor (Imbue)	MIT, ~200 stars	No	Free, bring a subscription
Traycer	MIT, ~900 stars	No	Free
Conductor	No	No	Free, Mac only
Zencoder	No	Slack only	$45/user/mo

Runs on your machine, driven from your phone. The segment this site is in.

Product	Open source	Mobile	Entry price
Anthropic Remote Control	No	Claude iOS and Android	Included with Pro
Orca	MIT, ~36,800 stars	iOS, Android (APK)	Free
Happy Coder	MIT, ~23,000 stars	iOS, Android, web	Free
T3 Code	MIT, ~16,500 stars	iOS, Android	Free
Paseo	AGPLv3, ~12,000 stars	iOS, Android, web	Free
VibeTunnel	MIT, ~4,600 stars	Phone web only	Free, donations
CC Pocket	Not published	iOS, iPadOS, macOS	Free
Cursor	No	iOS	Free, then $20/mo
OpenHands	MIT, ~83,000 stars	No	Free
LFG	MIT	Installable web app	Free

Prices and star counts checked on 4 August 2026. They move often, so open the vendor’s own pricing page before you commit.

Vendor cloud: fast to start, walled off from your work

Devin is the biggest name here, and Cognition is now a very large company. It raised over $1B at a $26B valuation in May 2026 and reported $492M run-rate revenue. It absorbed Windsurf along the way. Visit windsurf.com today and you land on devin.ai/desktop.

Devin is also the clearest illustration of the cloud tradeoff, because its own documentation is unusually blunt about it. The enterprise deployment page states that MFA VPNs are not compatible with the standard cloud tier. It states that private network access needs a single-tenant VPC with PrivateLink or an IPSec tunnel. And it states that Devin is not available for true on-premises or self-hosted deployment at any price.

Read that as a technical fact rather than a criticism. An agent in a vendor’s VM can only reach things that are publicly routable. Your internal staging environment is not.

Google Jules has the same shape and less flexibility. Each task gets a short-lived Google Cloud VM. The model is Gemini, with no option to bring your own. Limits are counted in tasks per day: 15 on the free tier, 100 on Pro, 300 on Ultra. Paid access currently requires a personal Google account ending in gmail.com, which rules out most corporate users.

You do not have to take any vendor’s word for how often this bites. GitHub’s cloud agent posts a specific line into a pull request when its firewall blocks an outbound connection: “Firewall rules blocked me from connecting to one or more addresses”. Searching GitHub for that exact phrase returns over 246,000 public issues and pull requests, and the count is still rising year on year. That is a quarter of a million public records of a cloud agent failing because it could not reach the internet.

Cursor’s cloud VMs have a sharper version of the problem. They run on a kernel that does not expose /dev/net/tun, so no TUN-based VPN can run inside them. Not WireGuard, not OpenVPN, not Tailscale. The feature request has sat open since March 2026.

Worth separating two questions that sound alike here. “Do I need to set up a VPN to reach my own machine” is a setup-friction question, and the answer for most of the tools below is no, because they ship a relay or a tunnel. “Can the agent reach the services on my VPN” is a capability question, and it is the one that decides whether the tool is usable on a real codebase. A cloud sandbox can answer the first and never the second.

Charlie Labs is worth reading because it argues the opposite case well. Its position is that running agent code on your development machine, with access to your credentials and filesystem, is a security posture that will look reckless in hindsight. That is a real argument. The cost of taking it is that every internal service you want the agent to touch needs its credential copied into a vendor dashboard, and anything reachable only over your VPN is out of reach entirely.

Blitzy sits at the far end. It sells to enterprises at a $50,000 pilot and bills per line of code. Its VPC deployment is explicitly inbound-only and never makes outbound requests. Excellent security posture. Also means the agent cannot call your internal API.

Your desktop: full access, until it sleeps

The open-source desktop orchestrators fix the access problem completely, because there is nothing to fix. The agent is on your machine, so it has your dotfiles, your keys, your running Postgres and your VPN routes by default.

Sculptor from Imbue gives each agent its own git worktree, branch, terminal and diff view, in isolated containers. It is MIT licensed and runs Claude Code. Imbue is explicit that it is an experimental research preview, and it supports Mac and Linux only.

Traycer takes a wider view. It orchestrates 16 different agents including Claude Code, Codex, OpenCode, Cursor, Copilot and Devin, using subscriptions you already own. The base tier is free with no credit meter.

Both share one limitation, and it is the reason this whole category exists. They are desktop apps. Shut the lid and the agents stop. There is no server-side scheduler and no way to trigger a run from your phone. Traycer sells a $10 per month Sync tier specifically so you can move between devices, which is a problem an always-on machine does not have.

Conductor is in the same camp and is worth a note, because it is often listed as a way to reach your agents from a phone and it is not one yet. Its pricing page still says the mobile app is “coming very soon”, and there is no Conductor app on the App Store as of 3 August 2026. It is also Mac only. Its cloud tier, at $50 per month, moves your agents onto Vercel sandboxes, which trades your always-on box for a vendor’s without adding a phone client.

Your machine, your phone: the segment that just consolidated

This is where the real competition is, and it had a violent 2026.

Anthropic shipped a version of this themselves, and it is the right place to start if Claude Code is the only agent you run. Remote Control landed in February 2026 and drives a session on your own machine from claude.ai or the Claude mobile app. It is included with Pro and Max, pairs by QR code, and sends push notifications.

It is a remote control, though, not a control plane, and that distinction is mechanical rather than marketing. The session still belongs to the claude process on your laptop. Your phone is a window onto it. Anthropic’s own limitations page says exactly that:

If you close the terminal, quit VS Code, or otherwise stop the claude process, the session ends. To keep a session running on a remote machine after you disconnect from SSH, start it inside tmux or screen.

A remote control shows you a session that is still tied to a machine that sleeps. A control plane puts the session on a machine that does not, and the window comes free with it. That is the whole difference, and Anthropic wrote the workaround into its own docs.

The other limitations follow from the same fact. A network outage of more than about ten minutes ends the session. It is disabled entirely if ANTHROPIC_BASE_URL points at a gateway or proxy, it does not work on Bedrock or Vertex, and it is incompatible with zero data retention. It is also Claude only, so your Codex and OpenCode sessions are not in that inbox.

Happy Coder is the strongest independent option. MIT licensed, roughly 23,000 GitHub stars, native iOS and Android apps, realtime voice, and end to end encryption with a relay you can self-host in about 1,300 lines of TypeScript. It costs nothing, with no usage limits and no account required. It wraps Claude Code and Codex, so you run happy claude instead of claude and your session appears on your phone.

If you want the thinnest possible way to see a Claude Code session on a phone, install it. We would rather tell you that than have you find out afterwards.

Its own marketing makes the argument this whole post is about, and it makes it better than we would: open source projects cannot rug pull you, and when a VC-funded startup pivots or shuts down, you are stuck.

Orca is the largest of these by some distance. MIT, about 36,800 stars since March 2026, 30-plus supported agent CLIs, and a desktop app that runs a fleet of agents in isolated git worktrees. It has native iOS, an Android APK from its releases page, and SSH worktrees so the agents themselves can live on a bigger remote box.

Its mobile companion has a specific limitation worth knowing before you install it, and the docs state it plainly: closing the desktop app drops the connection, because there is no relay. Your phone and desktop also have to share a network path, so LAN or Tailscale. Even when the agents already run on a remote server through SSH worktrees, the phone still reaches them through the desktop app you left running. There is an open issue asking for exactly the missing piece, a mobile client that talks to a headless server without the desktop in the middle.

Paseo is architecturally the closest thing here to running your own control plane. AGPLv3, about 12,000 stars, native iOS and Android plus a web app, and support for Claude Code, Codex, Cursor, Copilot, OpenCode and Pi. Crucially its daemon runs headless, so you can put it on a home server or a VPS rather than a laptop, and the phone reaches it through an end-to-end encrypted relay, your own tunnel, or the LAN. Nothing about your code goes to a vendor.

If you want the shape this post is arguing for and you want it AGPL, install Paseo. It is pre-1.0 and effectively a one-maintainer project, which is the tradeoff.

T3 Code is newer and moving fast. MIT, about 16,500 stars in five months, native apps on both platforms, and it drives Claude Code, Codex, Cursor, Grok Build and OpenCode on your machine. It is free with no paid tier, and its authors describe it as very early.

Then look at what happened to the startups in this exact segment.

Terragon shut down. The service ended on 9 February 2026, and the stated reason was that it could not reach the traction needed to be a sustainable business. Subscribers got a final month and refunds. The code was open sourced as an unmaintained snapshot.

Omnara was the flagship “control Claude Code from your phone” product. It has moved upmarket to an enterprise agent control plane. Its open-source repository carries a literal deprecation notice committed in November 2025, and the README says the Python version is no longer maintained. The iOS app on the store still describes the old product.

Meanwhile Cursor shipped a native iOS app that manages both its cloud agents and agents running on your own Mac in one inbox. OpenAI Codex is on iOS too. The incumbents absorbed the wedge.

VibeTunnel takes the generic route: it puts any terminal in a browser, so it works with any agent. It has about 4,600 stars, no native app, and you bring your own Tailscale or ngrok. It has also been in 1.0 beta for over a year.

The pattern in this segment is stark. The two venture-funded specialists are gone or repositioned. The free MIT projects are the ones still shipping daily.

The part nobody puts in a comparison table

Count the products in this category that stopped existing in the last eighteen months.

Codegen was acquired by ClickUp in December 2025. Its standalone product was discontinued in January 2026, roughly three weeks after the announcement. Customers got a migration document pointing them at five competitors. Its Apache-licensed SDK is still on GitHub, archived, and useless without the API behind it.

Terragon shut down on 9 February 2026, citing insufficient traction.

Refact retired its hosted cloud in 2026 and terminated active subscriptions. Its pricing page now returns a 404. The project survives as an open-source local tool, but the original repository was archived in May 2026 and the community did not follow the fork.

Solver, from Laredo Labs, was acquired by NVIDIA. Both of its marketing domains are now dark. One returns a 404 and the other serves an unconfigured certificate.

Vibe Kanban sunset on 10 April 2026 with 27,600 GitHub stars. The stated reason is worth quoting: the vast majority of users were free, and the team could not find a business model it was excited about. Subscriptions were terminated and refunded.

Omnara abandoned the positioning it was known for and deprecated the repo behind it.

Windsurf no longer exists as a brand. It is Devin Desktop now.

There is one more data point worth noting. Tembo abandoned a managed Postgres product used by more than 15,000 developers in order to become a coding agent platform. It is shipping well, and it is also proof that a company can walk away from a working product with thousands of users when the strategy changes.

None of this means cloud agent products are bad. Several of them are excellent, and if your work fits in a sandbox you should use one. It means the thing you are renting is a business decision made by somebody else, and business decisions change. A machine you own does not get acquired.

The honest comparison: OpenHands

If you take one competitor seriously, take this one.

OpenHands is MIT licensed with roughly 83,000 GitHub stars, which is two orders of magnitude more than any other open-source option here. It has also arrived at exactly the same conclusion. Its own README describes a self-hosted, always-on engineering team, and says outright that the most powerful way to run it is on a cloud server, so agents keep running when your laptop is shut.

It supports any ACP-compatible agent, so Claude Code, Codex and Gemini all run underneath it. It offers a free hosted tier that is capped at 10 conversations per day for a single user, and a self-hosted enterprise tier for your own Kubernetes.

The difference is not the deployment model, because it is the same model. The difference is what you get out of the box. OpenHands is badged beta, its no-sandbox mode gives the agent full filesystem access with a warning in the README, and hardening is left to you. LFG makes a narrower promise: it binds to loopback, never wants a public port, and omg connect pairs your box to a relay in one command so there is no VPN to configure and no inbound port to open. Tailscale is there if you would rather keep it entirely on your own network.

Pick whichever fits. Both are free, both are MIT, and neither can be switched off by a vendor.

The cost argument, briefly

Cloud pricing in this category moved in one direction and then reversed. Devin launched at $500 per month, cut to $20 in April 2025, and by 2026 had added a $200 Max tier and an $80 plus $40 per seat team plan. Google cut its top tier from $249.99 to $199.99 and added a $99.99 tier. OpenHands replaced paid cloud credits with a free bring-your-own-key tier.

The pattern is cheap entry and an expensive ceiling. That is a reasonable way to sell metered infrastructure, and it works badly for the specific workload here. Long-running agent work is mostly waiting. Tembo bills per vCPU-second on top of a subscription. Blitzy bills $0.20 per generated line, which prices the regenerate-and-rerun loop that good agent work depends on. Factory requires headroom in three separate rolling windows, 5-hour, 7-day and 30-day, before it will accept a request, so a long run can stall until a window rolls over.

On a box you already pay for, a long run is just wall-clock time.

There is a second cost question underneath that one, and it is the one to ask first: whose subscription pays for the tokens. Most people reading this already have a Claude Max or ChatGPT Pro plan they are not coming close to exhausting. Every tool in the “runs on your machine” section above uses it directly, so the control plane costs nothing and the model spend is already sunk. Buy a hosted plan and you are usually paying a second time, once for the subscription you already had and again for the vendor’s metered inference, sometimes with a markup on top. Check which of the two a product does before you compare monthly prices, because it changes the number more than the sticker does.

Which should you use

Use a vendor cloud if your work is self-contained. A public repo, a well-scoped pull request, a codebase with no private dependencies. Devin and Jules are genuinely good at this and you will be running in minutes.

Use a desktop orchestrator if you work from one machine, want full access to your environment, and do not need anything to happen while you sleep. Traycer and Sculptor are free and open source.

Use a phone client over your own machine if the laptop is usually awake anyway and you just want visibility from a queue. Happy Coder, Orca and T3 Code are all free with native apps on both platforms. Install one and see if that is the whole of your problem. Check how each reaches your machine before you commit: Orca needs the desktop app left running and a shared network path, while Happy and Paseo ship a relay.

Use your own always-on machine if any of these are true. Your agent needs your VPN, your staging database or a private registry. You want work to continue after you close the laptop, not just to watch it from elsewhere. You want a permission prompt answered from your phone at 2am to actually unblock the run. You do not want a vendor’s roadmap to be able to end your setup.

For that last case, the software is free either way. LFG runs agents in persistent tmux sessions on any Linux box or Mac and streams them to an installable web app. omg connect pairs that box to your phone in one command, so there is no VPN to configure and no inbound port to open, and Tailscale is there if you would rather keep it entirely on your own network. The self-hosting guide walks through the whole thing on a VPS.

If you would rather not own a machine at all, omg.dev starts the same control plane on a cloud machine in one click, already running and already awake. The honest caveat: agents there work on repos you clone into that workspace, not the ones sitting on your laptop. If your work depends on local state, self-host instead.

Start with the free path. If the box becomes a chore, pay someone to run it. That order round the right way saves you finding out the hard way which parts you actually needed.

FAQ

What is a coding agent control plane?

It is the layer that starts your coding agent, keeps its session alive, and gives you a way to watch and steer it. The agent itself (Claude Code, Codex, OpenCode) does the work. The control plane decides which machine it runs on, what it can reach, and how you get to it from a phone or another laptop.

Can a cloud coding agent reach my VPN or internal services?

Usually not on the default tier. Devin’s enterprise docs state that MFA VPNs are not compatible with its standard cloud deployment, and that private network access requires a single-tenant VPC with PrivateLink or IPSec. Google Jules documents no private network mechanism at all. Tembo shipped a Tailscale integration in July 2026, which is currently the clearest answer any hosted product gives.

Which coding agent platforms are open source?

Most of this category is. OpenHands (MIT, about 83,000 stars) is the largest, then Orca (MIT, about 36,800), Happy Coder (MIT, about 23,000), T3 Code (MIT, about 16,500) and Paseo (AGPLv3, about 12,000). VibeTunnel, Traycer, Sculptor by Imbue and LFG are also open source. Devin, Jules, Cursor, Conductor, Factory, Charlie, Blitzy and Zencoder are proprietary.

Which coding agent products shut down recently?

Codegen was acquired by ClickUp in December 2025 and discontinued in January 2026. Terragon shut down on 9 February 2026, citing insufficient traction. Vibe Kanban sunset in April 2026 with 27,600 GitHub stars, because almost all of its users were free. Refact retired its hosted cloud and terminated active subscriptions. Solver was acquired by NVIDIA and both of its domains are now dark. Omnara abandoned its mobile Claude Code positioning and deprecated the repo behind it. Windsurf was absorbed into Devin Desktop.

Do I need a paid plan to run an agent on my own server?

No. LFG, OpenHands, Orca, Happy Coder, T3 Code and Paseo are all free and self-hostable, and all of them run against the Claude or Codex subscription you already pay for. A hosted plan buys you a machine somebody else keeps running, not the software. Check whether a product uses your existing subscription or meters its own inference before comparing monthly prices, because that changes the real number more than the sticker does.

Why does my coding agent stop when I close my laptop?

Because the agent process is a child of your terminal session. Closing the lid suspends it. The fix is to run the agent inside a persistent session, usually tmux, on a machine that stays awake. That is the entire job of a control plane.

Is a desktop orchestrator enough, or do I need a server?

A desktop app is enough if you only work at one machine and are happy for agents to stop when it sleeps. You need a server the moment you want work to continue overnight, to trigger a run from your phone, or to pick up the same session from a different computer.