# cloudrouter (Manaflow)

Official pricing: https://cloudrouter.dev  
Category: agent-sandbox · Isolation: firecracker · DISCONTINUED

## Pricing regimes (raw)


## Features

Yes: snapshots, memory snapshots, pause/resume, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, full VM (own kernel), Docker inside, root, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, SSH, HTTPS preview URLs, open source, GPU, Python, Node.js, gVisor or VM (no shared kernel), firewall inside (nftables)

No: fork/clone, persistent disk, volumes, preinstalled agents, public IPv4, custom domains, raw TCP inbound, private networking, BYOC, arm64, Windows, macOS, credential injection, live fork (no pause), memory fork, webhooks, MCP server, automatic snapshots, snapshots on demand, agent harness API, inbound access rules, extra volumes, shared volumes

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/cloudrouter.json → feature_evidence

## Caveats

- Discontinued or closed to new customers, confirmed on the live site 2026-09-29. Confirmed the existing card. Both cloudrouter.dev URLs resolve to the Manaflow homepage, which lists cmux rather than cloudrouter. No live pricing or purchasable tier was found. The official source still documents the historical E2B/Modal router, default large 8-vCPU/32-GB/80-GB box, 10 concurrent d
- No public pricing ever existed: no /pricing page, no rates in docs, HN launch thread or source; card is null-priced and all modes flagged so the engine never auto-picks it.
- Product appears discontinued: cloudrouter.dev (and /pricing) redirect to manaflow.com as of 2026-09-28; last CLI release 2026-02-16.
- cloudrouter is a router over E2B (default) and Modal (GPU); at launch 'all usage is routed through us' (co-founder on HN) - bring-your-own-key was planned, never shipped.
- Concurrency 10 default; tiers low/mid/high (50/100/500) exist only in backend source code with no price.
- Features reflect E2B (Firecracker, pause = memory snapshot) for CPU sandboxes and Modal (gVisor) for GPU sandboxes.
- Default sandbox timeout 600 s (auto-stop) and default size large (8 vCPU/32 GB) - over-provisioned vs 4/8 unless --cpu/--memory passed.
- regions null: not documented.

## How this provider charges


cloudrouter is a CLI plus agent skill from Manaflow (the team behind cmux). It lets Claude Code, Codex, Cursor and similar agents
start cloud VMs (`cloudrouter start .`) and GPUs (`--gpu H100`). It is not a compute provider itself. It routes to **E2B**
(default, Firecracker microVMs) or **Modal** (GPU sandboxes). Each VM comes with a VNC desktop, VS Code in the browser,
Jupyter Lab, Chrome with CDP, and Docker.

**Launched:** Show HN on 2026-02-13 (https://news.ycombinator.com/item?id=47006393). The npm package `@manaflow-ai/cloudrouter`
was first published on 2026-02-09.
**Status:** on 2026-09-28, https://cloudrouter.dev and https://cloudrouter.dev/pricing both redirect to https://manaflow.com/, and that
page lists only cmux. The last CLI release (0.9.30) was on 2026-02-16 and the last skill-repo commit on 2026-02-19. The product
looks dormant or discontinued.

**Pricing was never published.** There is no pricing page. The Wayback CDX (checked before the archive went offline) has no
`/pricing` capture, and the docs, HN thread and source code give no rate. On HN (2026-02-14) the co-founder wrote: "Right now, all usage is routed
through us, hence `cloudrouter login` is required for now. Plan on adding bring-your-own cloud/key".

## Regime table

| Regime | When it applies | How billed | Numbers | Source URL |
|---|---|---|---|---|
| Default account (logged in) | Every user after `cloudrouter login` | Unknown. Usage ran on Manaflow's own E2B and Modal accounts. No price, credit or free-tier terms were published | 10 concurrently running sandboxes; self-serve GPUs T4, L4, A10G | https://raw.githubusercontent.com/manaflow-ai/manaflow/main/packages/convex/convex/cloudRouterSubscription.ts, https://news.ycombinator.com/item?id=47006393 |
| Subscription tier "low" | Only in backend source (`cloudRouterSubscription` table) | Unknown fee | 50 concurrent; unlocks L40S, A100 | same .ts file |
| Subscription tier "mid" | Source only | Unknown fee | 100 concurrent; adds A100-80GB, H100 | same |
| Subscription tier "high" | Source only | Unknown fee | 500 concurrent; adds H200, B200 | same |
| GPU approval | L40S/A100/A100-80GB/H100/H200/B200 without a tier | Contact founders@manaflow.ai | n/a | https://raw.githubusercontent.com/manaflow-ai/manaflow/main/packages/cloudrouter/internal/cli/start.go |
| Size presets | `--size` (default **large**) | Unknown | small 2 vCPU/8 GB/20 GB; medium 4/16/40; **large 8/32/80 (default)**; xlarge 16/64/160; `--cpu/--memory/--disk` override | start.go, SKILL.md |
| Running | Sandbox running | Unknown. The underlying E2B cost is per-second allocated (E2B list 4 vCPU/8 GiB = $0.3312/h) | Default timeout 600 s, extendable (`extend`, default +1 h) | https://e2b.dev/pricing (reference only) |
| Stopped = paused | `cloudrouter stop`/`pause` | Unknown. The E2B backend does not bill paused sandboxes | Resumable with `resume` | SKILL.md |
| Deleted | `cloudrouter delete` | n/a | State destroyed | SKILL.md |
| Egress / storage / IPv4 | All | Not documented | null | none |

## Gotchas

1. **You cannot budget it.** No price exists anywhere, and the tiers in the code carry no fee. The Battleships card is null-priced and every mode is
   flagged `sales`/`beta`, so the engine never picks it automatically.
2. **It is probably dead.** The domain redirects to the Manaflow homepage and releases stopped in February 2026.
3. **The default size is 8 vCPU / 32 GB.** The skill tells agents never to pass `--size`, so each agent-started sandbox is 2 to 4 times bigger than a
   typical 4/8 sandbox. If it were ever billed on allocation, that multiplies the cost.
4. **The default timeout is 10 minutes**, and GPU tasks are told to use `--timeout 7200`. Runaway GPU spend was raised on HN (agent retry loops
   starting B200s). The only guardrails were concurrency caps and approval gates on large GPUs.
5. **Multi-provider "router" in name only.** Only E2B and Modal shipped. Vercel, Daytona, Morph and Freestyle were announced as "working on".
6. **Concurrency tops out at 10** on the default account. 50, 100 and 500 need tiers that were never sold publicly.

## Worked example (4 vCPU / 8 GiB, 50 concurrent × 8 h/day × 22 days = 8,800 h, 30% CPU, 50 GiB snapshots, 100 GiB egress)

**Cannot be priced.** No cloudrouter rate exists. Feasibility: 50 concurrent is above the default limit of 10 and needs the unpriced "low" tier (50).

For reference only, if usage were passed through at the default backend's (E2B) list rates: 8,800 h × $0.3312 = **$2,914.56**
(allocation billing, so 30% CPU changes nothing), plus E2B's $150/mo Pro plan needed for more than 20 concurrent sandboxes,
giving **about $3,064.56**. Paused-state storage and egress are not billed or not documented on E2B. With the default `large` size (8/32) and no `--cpu 4 --memory 8192`,
the pass-through figure would be 8,800 × (8×0.0504 + 32×0.0162) = 8,800 × $0.9216 = **$8,110.08** + $150.
None of these are cloudrouter prices. cloudrouter's markup, subsidy or free usage is unknown.