[Cloud]: Agent37

[vishnukool]

Link

https://www.agent37.com/cloud

Why?

Every cloud in the matrix today gives you a VM that costs the same whether you are using it or not. Agent37 sandboxes sleep when idle, keep their disk, and wake on the next request. Parked, an instance costs $0.54/mo instead of $4.94.

That fits how people actually use spawn. Launch a box for a task, walk away, come back tomorrow and the work is still on disk, and you paid cents for the gap. No other provider in the table has an "off but keeps its state" mode.

Two things specific to us:

• We host Hermes and OpenClaw as products, baked into our images. spawn hermes agent37 is one API call and about 12 seconds with no install step, where every other cloud installs the agent from scratch on every spawn.
• OpenClaw is already an agent in your matrix. We are the company that runs it as a managed product, so this is less a new provider showing up than an existing part of the ecosystem asking to be reachable from the CLI.

Sandboxes are gVisor isolated with a hard per-instance disk quota. SSH runs over an authenticated WebSocket through Cloudflare, so there is no public host and no inbound port 22.

On the integration

I read the repo history, including the Fly.io removal in #1979 and the trail of transport fixes that led to it. I would rather not hand you another bespoke transport to carry, so the module I have written is deliberately boring:

• No infrastructure ask. labs/spawn/{cloud}/*.sh is a passthrough of sh/, the bundle release is cut by cli-release.yml, and nothing needs a secret, a Packer template, a Docker image, or a workflow edit.
• No new npm dependencies. Transport is stock OpenSSH. The only deviation is a ProxyCommand, about 300 lines of node:tls, no SDK.
• Every command path branched in the same PR: delete (both arms), status, list, connect, fix, export. bug: spawn delete broken for Fly.io VMs (sources non-existent fly/lib/common.sh) #1815 was a missed delete arm and I would rather not repeat it.
• We maintain it. If it ever stops earning its place in the matrix, we will send the removal PR ourselves.

Current state of the branch, run locally against main:

gate	result
biome check packages/cli/src/	clean, 211 files
bun test	2301 pass, 0 fail, 123 files
sh/test/macos-compat.sh	0 errors, 0 warnings, 107 files
build-clouds.ts	9 built, 0 failed, emits agent37.js
bash -n sh/agent37/*.sh	clean

It claims agent37/hermes and agent37/openclaw as implemented and marks the other eight missing, per the rule in .claude/rules/discovery.md. I will post the end-to-end run of both cells in this thread.

Two small things I found while working in the repo, offered either way:

1. packages/cli/sandbox.js is emitted by build-clouds.ts but missing from packages/cli/.gitignore, so it shows up as an untracked artifact after any build. Same class as fe137e4. One line.
2. The "Request it here" link under the cloud picker on openrouter.ai/spawn/{agent} is a <button type="button" class="spawn-cloud-feedback"> with no click handler, so clicking it does nothing. That is why this is an issue rather than a form submission.

How would you like this delivered?

gate.yml auto-closes PRs from non-members, so I have not opened one. Happy to take collaborator access, or to hand the branch to whoever wants to open it, whichever is less work for you. I can also provide a funded sk_live_ key so you can run every cell yourself before merging anything.

Vishnu, founder of Agent37.

[vishnukool]

End-to-end run done, against production, with a real OpenRouter key. Results below, including the part that did not work.

agent37/hermes — works

Creating Agent37 instance 'spawn-e2e-hermes2' (agent37-hermes, 2 vCPU · 4 GB RAM · 6 GB disk)...
Instance created: legsmi9xep (running, model anthropic/claude-sonnet-5)
Waiting for SSH over the Agent37 edge...
Instance reachable over SSH
Agent is wired to your OpenRouter key
Snapshot boot — skipping agent install

Then a real turn through the instance:

status: completed | text: 'SPAWN_E2E_OK' | usage: {input_tokens: 2, output_tokens: 14, cost_usd: 0.055664}

And the part that actually matters for you, the billing attribution. OpenRouter usage_daily on the key immediately before and after that turn:

before: 1.889084203
after:  1.945224203
delta:  0.056140   (turn reported cost_usd 0.055664)

The spend landed on the user's OpenRouter account, not on Agent37 managed credits. That is enforced, not incidental: the module reads the boot-time provider config off the box and refuses the instance if it does not point at https://openrouter.ai/api/v1, rather than silently handing over a box billing somewhere else.

Interactive session verified separately: spawn hermes agent37 drops into the Hermes TUI over the same SSH transport, delete works, and the instance is gone from the account afterwards.

Three bugs the run found in my own module, all fixed:

1. The provider probe read $HOME/.hermes/config.yaml, but SSH logs in as root while the image runs the agent as node. It missed every time and rejected a correctly wired instance.
2. Remote PATH was built from root's home for the same reason.
3. I had defaulted the model to nousresearch/hermes-4-405b. No nousresearch/* model on OpenRouter advertises tool support, and Hermes sends tools on every turn, so the first real request came back 404 No endpoints found that support tool use. Default is now anthropic/claude-sonnet-5, overridable with --model.

agent37/openclaw — dropped to missing

It provisions and wires up correctly, but OpenClaw 2026.7.1 validates a provider's model catalog against text | image | video | audio, and 100 models in OpenRouter's catalog declare input modality file. Config validation fails, the gateway crashloops three times and stays down.

models.providers.openrouter.models.330.input.3: Invalid input (allowed: "text", "image", "video", "audio")
OpenClaw gateway crashed 3 times consecutively.

That is our image to fix, not spawn's, and I would rather mark the cell missing than ship a combination that does not survive its first request. It comes back in a follow-up once the image moves. Worth flagging regardless of Agent37: any OpenClaw on that version pointed at OpenRouter will hit this.

So the PR claims exactly one cell, agent37/hermes, verified end to end. Matrix headline goes to 80 working combinations.

Gates, re-run after the fixes

gate	result
biome check packages/cli/src/	clean, 211 files
bun test	2301 pass, 0 fail, 123 files
sh/test/macos-compat.sh	0 errors, 0 warnings
build-clouds.ts	9 built, 0 failed, emits agent37.js
bash -n sh/agent37/*.sh	clean

Still happy to hand the branch over or take collaborator access, whichever you prefer. And I can provide a funded key so you can reproduce the run above yourselves.

[vishnukool]

@AhmedTMM @louisgv @scott sorry to tag directly, there is no obvious channel for this one.

Short version: the Agent37 cloud module is written, every gate you run is green locally, and agent37/hermes is verified end to end against production with the turn billing to the user's own OpenRouter key (usage delta in the comment above). It needs no infrastructure from you: no secrets, no workflow edits, no Packer, no Docker, no new npm dependencies.

The only thing blocking a PR is gate.yml, which auto-closes PRs from non-members before anyone reads them. So I need one of three things, whichever you prefer:

1. Collaborator access, and I open the PR and maintain the module.
2. Tell me where to push the branch and one of you opens it.
3. Tell me you are not taking new clouds right now, and I will stop. A no is genuinely useful.

Happy to send a funded sk_live_ key so you can run the cell yourselves before deciding anything.