fleet
Creates and manages agent pods on Fly.io Sprites, supporting command execution, file reads/writes, pod listing, and stopping, with automatic provider selection by price.
Creates and manages agent pods on Upstash Box, supporting command execution, file reads/writes, pod listing, and stopping, with automatic provider selection by price.
Creates and manages Vercel Sandbox pods, supporting command execution, file reads/writes, pod listing, and stopping; Vercel Hobby sandboxes are limited to 45 minutes.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
fleet
One API for every agent pod. A pod is a Linux machine your agent can run commands in.
Fleet starts each pod on the cheapest sandbox provider you have a key for, and moves to the next one when a provider fails. You bring your own provider accounts: connect each one once in the cabinet (fleet keeps the key encrypted), or send its key with each call.
Status: beta, live at https://fleet.flisk.si. Sign in with GitHub to reach the cabinet: up to 10 named fleet keys, your pods, an overview, providers and docs. Daytona, Blaxel, Upstash Box, Vercel Sandbox, E2B, Freestyle and Tensorlake have been run end to end on a real account. The other ten adapters are written from the providers' public API docs and tested against fake replies only.
Providers
Provider | Status | Header for your key | List price, 2 vCPU / 4 GB, per hour | Source |
Daytona | Working: run end to end on a real account on 2026-11-12 |
| $0.166 | |
Upstash Box | Working: run end to end on a real account on 2026-11-12 |
| $0.100 (active CPU) | |
Blaxel | Working: run end to end on a real account on 2026-11-12 |
| $0.166 | |
boat | Blocked: needs a card (7-day trial, then $20 a month), help wanted |
| $0.018 | |
Isorun | Blocked: provider down (sign-in failed and its EU runner answered 521, seen 2026-11-13), help wanted |
| $0.110 | |
E2B | Working: run end to end on a real account on 2026-11-13 (free team: 2 vCPU base template, 1 h sessions) |
| $0.166 | |
Hopx | Blocked: provider down (API error 526 since 2026-11-12), help wanted |
| $0.166 | |
Runloop | Blocked: invitation only (waitlist: support@runloop.ai), and winding down: transition support ends 2026-11-13, help wanted |
| $0.317 | |
Vercel Sandbox | Working: run end to end on a real account on 2026-11-12 (Hobby: 45-minute sandboxes) |
| $0.341 at most (CPU billed only while busy) | |
Fly Sprites | Blocked: needs a card on accounts whose Fly trial ended, help wanted |
| $0.164 at most (billed on use) | |
Morph | Blocked: needs billing (a new account got 402 payment required, seen 2026-11-13), help wanted |
| $0.100 | |
Tensorlake | Working: run end to end on a real account on 2026-11-13 (free plan) |
| $0.200 (active CPU; free plan caps a sandbox at 1 vCPU / 1 GB) | |
Freestyle | Working: run end to end on a real account on 2026-11-13 (free plan) |
| $0.132 | |
LangSmith Sandboxes | Untested, help wanted |
| $0.189 |
Most providers need one API key. Three need more, joined with colons in one header value (the cabinet's Providers page has one input per part): Blaxel workspace:key, Vercel token:projectId[:teamId], boat key[:organization].
Working means fleet has started, run commands and files in, and stopped pods on a real account of that provider. Untested providers are built from public docs and pass tests against fake replies, so they should work; they wait for a test on a real account (bun scripts/real-check.mjs). Blocked providers cannot be tested yet: the provider is down, asks for a card, or is invitation only. Have an account at one of them, or want another provider (#4)? Help is welcome.
Prices were read on 2026-11-12. You pay each provider directly on your own account; fleet adds nothing.
Not added, because fleet is a Worker that only speaks plain HTTP: Modal (no public HTTP API, its SDKs use gRPC), Northflank (commands only over an HTTP/2 or WebSocket stream), CodeSandbox (its HTTP exec is undocumented) and Cloudflare Sandbox (each user would deploy their own Worker).
Related MCP server: Gemini-Cloud-Agent-MCP
Browser pod: a pod with no key at all
The browser pod, a small real Linux in the visitor's own tab, moved to its own repo: flisk-sh/linux2wasm. Fleet still serves a pinned release of it at /browserpod/, so sites that load it from here keep working. Run sh scripts/fetch-browserpod.sh to fetch it locally.
API
Send your fleet key as Authorization: Bearer fleet_..., plus one header per provider you hold an account at.
curl -X POST https://fleet.flisk.si/v1/pods \
-H "Authorization: Bearer fleet_..." \
-H "X-E2B-Key: ..." -H "X-Daytona-Key: ..." \
-d '{"cpu":2,"ram":4}'Call | Body | Returns |
|
|
|
|
|
|
|
|
|
|
| |
|
| |
|
| |
|
| |
| price table, no key needed |
Without
providerororder, fleet sorts your providers by list price for the size you ask for.failedlists the providers that were tried before the one that started the pod.cwd,env,timeoutand the file calls are shell commands fleet builds on top of run, so they work the same on every provider with no extra adapter code. Files move at most 256 KB a call. A timed-out command ends with exit code 124 andtimedOut: true.GET /v1/podsneeds only the fleet key. E2B ids in the list lack the access token, so keep the idPOST /v1/podsgave you to run commands there.A pod id holds the provider name and the provider's own id. For E2B it also holds that pod's access token, so treat pod ids as private.
MCP
POST /mcp is an MCP server with six tools: pod_create, pod_run, pod_write_file, pod_read_file, pod_list, pod_stop.
The tools carry MCP annotations: pod_list and pod_read_file are readOnlyHint, pod_stop is destructiveHint, pod_create and pod_run are openWorldHint. A fleet key made as Read-only in the cabinet sees only pod_list and pod_read_file, gets an error from the other tools, and gets 403 on the HTTP routes that change anything. A database made before this needs ALTER TABLE keys ADD COLUMN scope TEXT NOT NULL DEFAULT 'full'; once (wrangler d1 execute fleet --remote) before read-only keys can be made.
{ "mcpServers": { "fleet": {
"type": "http",
"url": "https://fleet.flisk.si/mcp",
"headers": { "Authorization": "Bearer fleet_...", "X-E2B-Key": "..." }
} } }What fleet stores
In Cloudflare D1 (schema.sql), sized for the free plan: every write is one row, every read uses an index.
A SHA-256 hash of each fleet key, its name and when it was created and last used (written at most once a day).
Pod facts for the cabinet: provider, size, price, start and stop times, and which providers failed first. The pod id is cut at the first
~, so E2B's access token is never stored.A usage log with one row per pod call (start, run, write, read, stop): pod id, time, status, how long it took and, for
run, the first 200 characters of the command and its exit code, and forwriteandreadthe file path. The cabinet shows it as a pod's history (click a pod id on the Pods page), andGET /v1/pods/:id/historyreturns it. The command line is stored as sent, so keep secrets inenv, which is not recorded. A database made before this needsALTER TABLE calls ADD COLUMN detail TEXT; ALTER TABLE calls ADD COLUMN exit_code INTEGER;once, andschema.sqlre-run for the new index; until then fleet keeps logging without the detail.Provider keys you connect in the cabinet, encrypted with AES-GCM under the
PROVIDER_KEYS_SECRETsecret (src/vault.js); only the last 4 characters are kept readable, to tell keys apart. Disconnect deletes the key. Keys sent in headers are never stored.Sessions are signed cookies, not stored.
Free-plan guards
At most 120 API calls a minute per account, counted in Cloudflare's Cache API, so the limit itself writes nothing.
At most 300 pod starts a day per account.
Demo pods
Send X-Demo-Key: demo (or ttl=30 to make the pod die after 30 seconds) and fleet answers like a provider, with a simulated pod and no machine. It lets anyone try the flow without an account, and scripts/synthetic.mjs uses it to run 1 to 10 synthetic agents against a live fleet. It never shows up in the price table and proves nothing about real providers.
Limits to know
boat: fixed machine types (fleet picks the smallest that fits), needs a boat plan, machines in Europe only.
E2B: memory comes from the E2B template, not from the request.
Daytona: at most 4 vCPU and 8 GB a pod; output comes back as one stream, with no separate stderr.
Pods stop by themselves after about an hour on Isorun, E2B and Daytona if you forget them.
Icons and link previews
assets/ holds the sources (the icon from flisk-emoji, and og.html, the 1200 x 630 preview card). bun scripts/build-assets.mjs turns them into src/assets.js (needs rsvg-convert and Google Chrome), which the Worker serves as /icon.svg, /favicon.ico, /apple-touch-icon.png and /og.png. src/seo.js holds the page list behind the head tags, /sitemap.xml, /robots.txt and /llms.txt.
Run it locally
Needs Bun.
bun testbun scripts/dev.mjsThe dev server listens on http://localhost:8799 with an in-memory store. "Console" there skips GitHub and signs you in as local-dev.
bun scripts/simulate.mjs drives fleet's routing against a simulated provider (start delay, time to live) to measure 1 to 10 pods and what a caller sees when a pod dies. It needs no provider account.
Add a provider
Add one file in src/providers/ that exports label, keyHeader, priceSource, price(size), create, run and stop, register it in src/providers/index.js, and add tests with fake replies in test/providers.test.js.
License
MIT
Git hooks
bun install (or npm install) points git at .githooks/ through the prepare script. pre-commit checks your changes and pre-push runs the tests. To skip either once, add --no-verify. To set up by hand: git config core.hooksPath .githooks.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Cloud Linux VMs for your agent: run code, Docker, files and HTTPS previews in a Sandboxer microVM.
Linux sandboxes billed by the second from a prepaid balance, and GPU pods on your own RunPod key.
Linux VMs for AI agents: create VMs, run commands, edit files and open preview URLs.
On-demand GPU nodes for agents: create nodes, run commands, and submit jobs, billed by the minute.
Related MCP Servers
- AlicenseAqualityBmaintenanceProvides agents with real Linux VMs to run commands, manage files, expose ports, and track costs, all through an ephemeral sandbox service.11MIT
- FlicenseNot gradedqualityDmaintenanceEnables AI agents to operate a persistent Linux sandbox in the cloud, running commands, managing files, using Git, and publishing artifacts through a Streamable HTTP MCP endpoint.-
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to safely run bounded, sandboxed tasks on remote machines with persistent state and reviewable artifacts.Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to operate disposable Linux computers via MCP, providing tools for shell, file read/write, directory listing, and port exposure. It also supports browser automation and integration with any MCP-compatible client.Apache 2.0