External AI agents (MCP)

Connect in one paste

Paste this into Codex, Claude, Cursor, or another agent that supports remote Streamable HTTP MCP servers:

Set up Meshia for me: fetch https://meshia.io/agent-setup.md and follow it.

That short prompt intentionally delegates the detailed, evolving instructions to the canonical agent setup guide. The guide tells the agent how to configure its current client, complete OAuth with the owner, prove the connection without spending, inspect relevant project context, propose a compelling first workload, obtain separate project and compute approvals, synchronize progress, and clean up billable resources.

For Claude Code, the equivalent direct command is:

claude mcp add --scope user --transport http meshia "https://meshia.io/api/mcp"

Authentication

Your MCP client opens Meshia in a browser. Sign in with the same emailed 6-digit code as the website, review the requested control families and callback host, then click Connect agent. There is no API key to copy into a terminal.

Meshia uses OAuth authorization code flow with PKCE. Access tokens are short-lived, refresh tokens rotate, and only credential hashes are stored. Returning clients reconnect without repeating consent while their grant is still active.

To disconnect a client, open Settings → Tools → Connected agents and click Revoke twice. Meshia revokes every active token for that exact client and account.

What the agent can control

The MCP server exposes the same canonical controls as the website:

• list, create, inspect, rename, resume, limit, and stop workspaces;
• inspect capacity, request a no-spend quote, attach or release compute, and read utilization;
• run and cancel commands; list, read, edit, move, delete, upload, and download files;
• read and append synchronized workspace messages or ask Meshia's driving agent to take a turn;
• start, inspect, and cancel long-running swarm runs;
• inspect events, statistics, connected machines, teams, members, and invitations.

The external AI can remain the top-level orchestrator while Meshia supplies long-running compute, remote agents, provider tools, and a durable message ledger visible in the workspace UI.

Cost and safety defaults

Workspace creation is deferred and starts no GPU. A good agent should set workspace GPU, hourly, and total limits, inspect availability, and request a quote before asking you to confirm billable attachment.

Meshia requires explicit confirmation fields for GPU attachment, workspace stops, compute release, file deletion, swarm cancellation, invitation emails, and team membership changes. Durable storage is retained by default when compute stops.

Large files

Small reads and edits travel through MCP. For large local uploads, the agent asks Meshia for short-lived file-scoped upload tickets, computes SHA-256 locally, and sends bytes directly to object storage. Downloads use a workspace export manifest with short-lived object-scoped URLs. Large file bytes do not need to pass through the MCP gateway.

A strong first project

The setup guide tells the external agent to study the context you already shared, then recommend one concrete project rather than a generic list. The proposal must explain why remote compute helps, name the final artifacts and measurable success test, divide the work into reviewable checkpoints, start with a deferred workspace, and return with an exact quote before any billable compute is attached.

The same guidance is available after connection as the meshia://guide MCP resource and the first_meshia_project MCP prompt.