mountable.io

Overview

Persistent files for temporary agent jobs, with filesystem access delegated through one-time tickets.

Five ways to work with the same persistent filesAPI, mounted filesystem operations, MCP, CLI and SDK connect workflows to one shared filesystem. Activity from the different entry points saves files to the same persistent storage.MANY WAYS INOne place for every file.APIFilesystemMCPCLISDKYour filesbrief.pdfresearch.csvassets/Saved to mountable.io5 ENTRY POINTSTHE SAME PERSISTENT FILES
Different tools. The same files.

Temporary agent jobs often need the previous job's results, datasets or checkpoints. mountable.io gives a compatible Linux sandbox a mounted filesystem: ordinary file operations work against persistent storage rather than the sandbox's disposable disk.

For example, one job writes a dataset and unmounts cleanly. A later job mounts the same filesystem and reads that dataset without downloading and rebuilding an object-store folder itself. The quickstart is the human entry point; an agent can follow Agents or the skill.

On Linux, mounting needs usable FUSE, not just an installed CLI. A Mac needs nothing else (Mount on a Mac). Check the verified platforms and filesystem contract: this early-stage service is not a promise of high availability or immediate consistency between simultaneous mounts.

How it fits together

  • An organisation owns filesystems, members, groups, and API keys.
  • A grant lets a member, group, or API key mount one whole filesystem, read-only (ro) or read-write (rw).
  • A mount session is created by your backend (with an API key) or by a person (with mountable login). Creating it returns a one-time ticket.
  • The sandbox exchanges the ticket for a short-lived session certificate. The mount then authenticates with its session certificate and private key, rather than the backend's reusable key.

What stays out of the sandbox

In the ticket-based flow, keep the reusable API key or login on your trusted backend. Give an untrusted sandbox only the one-time ticket; its resulting session is scoped to one filesystem, one mode and an expiry. A trusted coding agent may instead use an API key or its own CLI login, with that credential's broader authority. Do not bake reusable credentials into sandbox images.

Revocation is bounded rather than instantaneous: no new operation is admitted more than 15 seconds after a session is revoked. See Permissions.

Next steps

  • Quickstart: sign in, create a filesystem, mount it.
  • Mount from a sandbox: the backend → ticket → sandbox flow.
  • Mount on a Mac: no macFUSE, no kernel extension, no sudo.
  • Sandbox comparison: which runtimes have a production mount, and which are only desk research.
  • Agents: the CLI workflow for agents, JSON output, errors and retries; also as a skill and an MCP server.
  • LLM provider starters: OpenAI, Anthropic, Gemini and Grok orchestrators that pass the sandbox a one-time ticket on stdin.
  • Workflow connectors: n8n, Make.com and Zapier call the control-plane API and hand a sandbox only a one-time ticket.
  • Workspace agents: OpenClaw and Hermes Agent use the mountable.io MCP server for a persistent workspace.
  • Python and TypeScript SDKs: read, write and follow files from a backend without mounting, and create mount tickets.
  • Permissions: sign-in, teams, invitations, keys, grants and modes.
  • Billing: card setup, usage charges and nonpayment recovery.
  • Contract and limits: what a mount guarantees, and what it does not.