PlatformMCP Early access

The Fantasti API, in the language agents speak.

A remote MCP server for Cursor, VS Code, Codex and other MCP clients such as coding agents and editors. It gives your agent tools to list GPUs, start a sandbox, submit a job, read logs and check spend. It signs in as you or with a scoped key, and it works inside the limits your admins set.

Request access Read the tool list

Opens with the first cohort. No account is open yet.

{
  "mcpServers": {
    "fantasti": {
      "type": "http",
      "url": "https://mcp.fantasti.ai/mcp"
    }
  }
}

Where it goes

.mcp.json · project root · OAuth sign-in

{
  "mcpServers": {
    "fantasti": {
      "url": "https://mcp.fantasti.ai/mcp"
    }
  }
}

Where it goes

.cursor/mcp.json · OAuth sign-in

{
  "servers": {
    "fantasti": {
      "type": "http",
      "url": "https://mcp.fantasti.ai/mcp"
    }
  }
}

Where it goes

.vscode/mcp.json · OAuth sign-in

codex mcp add fantasti \
  --url https://mcp.fantasti.ai/mcp
codex mcp login fantasti

Where it goes

~/.codex/config.toml · OAuth sign-in

Tools your agent gets Asks first12 tools Early access

Catalog: gpus_list; 1
Sandboxes: sandboxes_createsandboxes_execsandboxes_read_filesandboxes_write_filesandboxes_delete (asks first); 5
Workspaces: workspaces_listworkspaces_startworkspaces_stop (asks first); 3
Jobs: jobs_createjobs_logs A launch above the key's per-call budget; 2
Usage: usage_get; 1

Preview · answers with the first cohort

Server: mcp.fantasti.ai/mcp
Transport: Streamable HTTP
Sign-in: OAuth · scoped agent key
Tools: 12 early access · 25 planned

§01 Sign in Preview API

Two ways to sign in.

OAuth, for a person at a keyboard.

Your client opens the Fantasti sign-in page. You choose the project and the scopes. Revoke the session from the console at any time.

Sign-in with OAuth. Illustrative.

Fantasti sign-in OAuth Illustrative

Request

Your MCP client asks to use Fantasti as you.

mcp.fantasti.ai/mcp

ProjectChoose one

evalsChosen
research

Scopes3 of 5 tool groups

catalog1 tool (not granted)
sandboxes5 tools (granted)
workspaces3 tools (not granted)
jobs2 tools (granted)
usage1 tool (granted)

Session ready 8 tools in reach

Agent key, for an agent that runs without you.

A key carries its own scopes, project, spend budget and expiry. It is shown once.

Preview API · subject to change

{
  "mcpServers": {
    "fantasti": {
      "type": "http",
      "url": "https://mcp.fantasti.ai/mcp?features=sandboxes,jobs&project=evals",
      "headers": { "Authorization": "Bearer ${FANTASTI_AGENT_KEY}" }
    }
  }
}

What the key carries

scopes    sandboxes · jobs
project   evals
budget    a spend limit you set
expiry    a date you set
shown     once

# An agent you build yourself, on the open-source MCP Python SDK
import asyncio, os
from mcp import ClientSession
from mcp.client.streamable_http import streamablehttp_client

URL = "https://mcp.fantasti.ai/mcp/readonly"   # read tools, no writes
KEY = {"Authorization": f"Bearer {os.environ['FANTASTI_AGENT_KEY']}"}

async def main():
    async with streamablehttp_client(URL, headers=KEY) as (r, w, _):
        async with ClientSession(r, w) as session:
            await session.initialize()
            tools = await session.list_tools()
            args = {"resource": "job_4h1x"}
            used = await session.call_tool("usage_get", args)

asyncio.run(main())

§02 Tools Reviewed 2026-10-09

TAB 01MCP tool groups Source · Fantasti Reviewed 2026-10-09

MCP tool groups: the tools in each group, whether the group reads or writes, what asks for confirmation, and its status.
Group	Tools	Kind	Confirmation	Status
Early access12 tools · 5 groups
Catalog1 tool	`gpus_list`	Read	None	Early access
Sandboxes5 tools	`sandboxes_create` `sandboxes_exec` `sandboxes_read_file` `sandboxes_write_file` `sandboxes_delete` (asks before it runs)	Write	`sandboxes_delete`	Early access
Workspaces3 tools	`workspaces_list` `workspaces_start` `workspaces_stop` (asks before it runs)	Write	`workspaces_stop`	Early access
Jobs2 tools	`jobs_create` `jobs_logs`	Write	A launch above the key's per-call budget	Early access
Usage1 tool	`usage_get`	Read	None	Early access
Planned25 tools · 8 groups
Batch3 tools	`batch_create` `batch_get` `batch_cancel` (asks before it runs)	Write	`batch_cancel`; a launch above budget	Planned
Flows6 tools	`flows_list` `flows_trigger` (asks before it runs) `runs_get` `runs_logs` `runs_resume` (asks before it runs) `artifacts_read`	Write	`flows_trigger`, `runs_resume`	Planned
Broker3 tools	`broker_plan` `broker_record` `capacity_outlook`	Read	None	Planned
Quota2 tools	`quotas_list` `quotas_request`	Read, request	A request opens a case	Planned
Clusters4 tools	`clusters_list` `clusters_get` `clusters_create` (asks before it runs) `clusters_delete` (asks before it runs)	Write	Create and delete	Planned
Records3 tools	`decisions_list` `placements_get` `statement_get`	Read	None	Planned
Support3 tools	`docs_search` `support_open_ticket` `support_get_ticket`	Write	None	Planned
Changelog1 tool	`changelog_list`	Read	None	Planned

A tool with this mark returns a confirmation step before it runs.

Sheet: 01 / 02
Title: MCP server · tool groups
Tools: 37 in 13 groups
Reviewed: 2026-10-09

EQ Early access See §02

12 tools in early access.

5 groups · 12 tools

EQ 01 Source · Fantasti MCP tool list Reviewed 2026-10-09

25 more tools are planned, in 8 groups. Every group carries its status in the tool list.

Read the tool list

§03 Context Source · Fantasti

Built for a context window.

TAB 02Three URLs, three habits Source · Fantasti

Load what you need.

https://mcp.fantasti.ai/mcp?features=sandboxes,jobs

?features=sandboxes,jobs exposes those groups and nothing else.

Reaches

Sandboxes · Jobs7 tools

Out of reach

Catalog · Workspaces · Usage
Read access.

https://mcp.fantasti.ai/mcp/readonly

/mcp/readonly exposes read tools and cannot reach a write.

Reaches

Read tools

Out of reach

Every write
One project.

https://mcp.fantasti.ai/mcp?project=evals

?project=evals pins a session to one project.

Reaches

Project evals

Out of reach

Every other project

Two tools for the whole API. Planned: fantasti_search finds an operation and fantasti_call runs it, so the tool list stays short however large the API grows.
Answers sized for a model.: Lists are paged, logs are tailed, and every result carries ids your agent can pass to the next call.
Stable order.: Tools are returned in the same order every time, which keeps prompt caches warm.

§04 Guardrails Enforced by the server

Guardrails.

Scope Scoped keys. A key reaches the tool groups, the project and the capacity modes you name. tool groups · project · capacity modes
Policy Your limits apply. Quotas, spend caps and price ceilings from your policy apply to agents exactly as they apply to people. An agent cannot raise them. quotas · spend caps · price ceilings
Confirm Asks before it deletes or spends. Destructive tools, and launches above the key's per-call budget, return a confirmation step. A person approves in the client or in the console. sandboxes_delete · workspaces_stop
Budget Budget per key. Give an agent key a spend budget and an expiry. At the budget, new launches stop. Authorization: Bearer ${FANTASTI_AGENT_KEY}
Record Every call on record. Each tool call is logged with the key, the client, the arguments and the result. key · client · arguments · result

Keep your client's tool approval on, and take care when you combine MCP servers. Text returned by a tool is data, not an instruction.

§05 A session Illustrative

One prompt, five tool calls.

The agent checks the catalog, starts a closed-network sandbox, runs the suite, reads the meter and cleans up. The delete waits for a yes.

Sandboxes

TAB 03Tool calls for one prompt illustrative

Agent session

PromptRun the eval suite on one L40S with no network, and tell me what it used.

fantasti.gpus_list (read) { "gpu": "L40S" } available
fantasti.sandboxes_create (write) { "gpu": "L40S", "ttl": "15m", "network": "closed" } sb_7q2m… ready
fantasti.sandboxes_exec (write) { "id": "sb_7q2m…", "cmd": "pytest -q" } exit 0
fantasti.usage_get (read) { "resource": "sb_7q2m…" } 412 GPU·s
fantasti.sandboxes_delete (write) { "id": "sb_7q2m…" } confirm? yes ended

Answer

The suite passed. It used 412 GPU·s on one L40S, and the sandbox has ended.

Metered: 412GPU·s
Tool calls: 5
Confirmation: 1
Sandbox: ended

§06 Clients Reviewed 2026-10-09

Works with the clients you already use.

Streamable HTTP, MCP revision 2026-07-28, with 2025-11-25 for older clients.

Request path Schematic

One server, the same API. Schematic · not to scale.

An MCP client, or an agent you build yourself, connects to the Fantasti MCP server over Streamable HTTP. The server checks scopes, budgets and confirmations, writes the call to its log with the key, the client, the arguments and the result, and calls the Fantasti API. The SDK, the CLI and the console call the same API directly. Your policy, with its quotas, spend caps and price ceilings, is the boundary around the server and the API: it applies to every call. Connections: Your MCP client to MCP server: Streamable HTTP; Your own agent to MCP server; MCP server to Call log; SDK, CLI and console to Fantasti API: REST; MCP server to Fantasti API.

Your policyQuotas · spend caps · price ceilings

Your MCP client
A coding agent or aneditor
signs in
OAuth
Your own agent
Any SDK that supports MCP
signs in
agent key
SDK, CLI and console
The tools your team uses
signs in
API key
MCP server (highlighted)
mcp.fantasti.ai/mcp
checks
scopes · budgets
asks
before delete or spend
Fantasti API
api.fantasti.ai/v1
The same objects,ids and limits
RecordCall log
- key
- client
- arguments
- result

Your policySet by your admins

Your MCP client
A coding agent or an editor
signs in
OAuth
Your own agent
Any SDK that supports MCP
signs in
agent key
SDK, CLI and console
The tools your team uses
signs in
API key
MCP server (highlighted)
mcp.fantasti.ai/mcp
checks
scopes · budgets
asks
before delete or spend
Fantasti API
api.fantasti.ai/v1
The same objects,ids and limits
RecordCall log
- key
- client
- arguments
- result

Your MCP client
A coding agent or an editor
signs in
OAuth
to
MCP server
Your own agent
Any SDK that supports MCP
signs in
agent key
to
MCP server
Your policy
Quotas, spend caps and price ceilings, setby your admins
to
Fantasti API
MCP server (highlighted)
mcp.fantasti.ai/mcp
checks
scopes · budgets
asks
before delete or spend
to
Fantasti API
RecordCall log
- key
- client
- arguments
- result
Fantasti API
api.fantasti.ai/v1
The same objects,ids and limits
SDK, CLI and console
The tools your team uses
signs in
API key

A tool call
The same API, without MCP

Editor Cursor .cursor/mcp.json
Editor VS Code .vscode/mcp.json
Coding agent Codex codex mcp add fantasti
Coding agents and editors Other MCP clients mcp.fantasti.ai/mcp
Any SDK that supports MCP Agents you build Authorization: Bearer ${FANTASTI_AGENT_KEY}

§07 Questions 5 questions

§08 Request Early access

Give your agent a key and a budget.

It can take it from there, inside your limits.

Request access Read the API

The Fantasti API, in the language agents speak.

Built for a context window.

Guardrails.

One prompt, five tool calls.

Works with the clients you already use.

Is this a different API?

Can an agent run up a bill?

OAuth or a key?

Can I host the server myself?

Do Fantasti's own agents use it?

Give your agent a key and a budget.

01OAuth, for a person at a keyboard.

02Agent key, for an agent that runs without you.

One prompt, five tool calls.

Is this a different API?

Can an agent run up a bill?

OAuth or a key?

Can I host the server myself?

Do Fantasti's own agents use it?

Give your agent a key and a budget.

OAuth, for a person at a keyboard.

Agent key, for an agent that runs without you.