AgentPhone
Give an AI agent its own US or Canada number to call and text.
Give your AI agent a persistent, hardware-isolated Linux sandbox it can create, run commands in, checkpoint and destroy.
Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.
Paste it into any MCP client. Setup docs
The Sprites MCP server connects Fly.io's code sandboxes to Claude, ChatGPT and any MCP-compatible agent. It exposes 18 tools that create a persistent Linux microVM, run shell commands inside it, snapshot and roll back its filesystem, manage background services, and rewrite its outbound network policy. Sign-in is OAuth, and the default token only reaches Sprites the connector itself created.
Verified connector
Listed by Anthropic as a partner connector in its Connectors Directory.
Connection checked by Agentman on .
Anthropic states this reflects the level of review a connector received, not a security audit.
Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.
No. The exec tool runs commands inside a Sprite, which Fly.io describes as a hardware-isolated Firecracker microVM in its cloud, not on your computer. Nothing executes locally, and the connector reaches only Sprites owned by the Fly.io organization you pick during the OAuth flow.
Anything. Sprites documents outbound access as unrestricted by default, so a fresh Sprite can resolve and reach any domain. Tightening it is opt-in through a network policy, a DNS-based allowlist applied from outside the Sprite. The policy_network_update tool is how the connector applies that restriction.
Not by itself. Every Sprite has an HTTPS URL, but Sprites documents it as private by default and reachable only by members of your organization. Switching a URL to public is a separate sprite config update command, and no tool in the connector's 18 published names performs that switch.
No. Sprites states plainly that destroy_sprite permanently deletes a Sprite and its data with no undo. Checkpoints protect against a bad command inside a surviving Sprite, not against the Sprite being removed. Restrict the connector token to a name prefix so it can only reach Sprites you created for it.
The writable filesystem overlay. Sprites lists files, directories, installed packages, config files and on-disk databases as captured, and running processes, in-memory state and open network connections as not captured. Restoring replaces your current filesystem with the saved one, and the state you overwrite is not backed up for you.
Yes, and the default already does. Sprites issues a restricted token that can only create Sprites whose names begin with mcp- and caps how many. The consent screen lets you change the prefix, raise or lower the cap, or switch to full access across the organization. Sprites recommends staying restricted.
Yes, while they run. Sprites bills compute per second and only while a Sprite is active, so an idle Sprite costs no compute. Storage is billed on bytes actually written against a 100 GB per-Sprite ceiling, and kept checkpoint blocks bill separately. Fly.io publishes no pricing page in the Sprites documentation.
The Sprite was cold and is still waking. Sprites documents cold wakes as taking one to two seconds, and its troubleshooting tells you to wait a few seconds and run the same tool again. A warm Sprite resumes in 100 to 500 milliseconds instead, with running processes preserved.
mcp- name prefix and count cap, the organization-tool table with its read-only and destructive labels, the Sprite-tool categories, the safety-model notes on destroy, exec, restore and network policy, and the troubleshooting rows used in the FAQ (retrieved 2026-08-18 as Markdown via the .md suffix) · retrieved 2026-08-18defaults preset bundle, raw-IP and private-range blocking, live policy reload, the read-only in-Sprite policy file, the per-Sprite HTTPS URL, and the org-private-by-default URL auth with its public switch (retrieved 2026-08-18) · retrieved 2026-08-18sprite auth setup --token CI path and the account page as the token source (retrieved 2026-08-18) · retrieved 2026-08-18https://sprites.dev/mcp — an anonymous GET returned HTTP 401 with body {"error":"authentication failed"} and no WWW-Authenticate header. The RFC 9728 descriptor at https://sprites.dev/.well-known/oauth-protected-resource returned HTTP 200 naming the endpoint as its resource, header as the only bearer method, https://sprites.dev as its own authorization server, and sprites:read plus sprites:write as the two supported scopes. The RFC 9728 path-insert and path-append forms both returned 404. RFC 8414 metadata at https://sprites.dev/.well-known/oauth-authorization-server returned HTTP 200 with authorization_code and refresh_token grants, PKCE S256, a dynamic client registration endpoint and the same two scopes; its /mcp-suffixed form returned 404 (2026-08-18) · retrieved 2026-08-18docs.sprites.dev/robots.txt allows all crawlers, names sitemap-index.xml, and carries no Content-Signal header on any axis. sprites.dev/robots.txt is entirely commented out. The documentation host serves llms.txt and llms-full.txt, and appending .md to any docs URL returns clean Markdown. The sitemap lists 39 URLs and contains no pricing page; sprites.dev/pricing returns 404 (all checked 2026-08-18) · retrieved 2026-08-18prompt_names array, the Fly.io author attribution and the prose permissions paragraph (snapshot 2026-08-16) · retrieved 2026-08-16Connect Sprites once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.