Platform MCP

fal provides a second MCP server — the Platform API MCP — focused on operating and debugging your fal account. Where the Run MCP lets your assistant build with models (search, schemas, inference), this server lets it inspect your serverless apps: requests, logs, error analytics, deploy and revision history, runner state, queue depth, spend, and files. Connect it to Claude Code, Cursor, or any MCP-compatible client with your own API key, then ask things like “why did my last request to my-app fail?” — your assistant walks the same debugging steps you would: find the failing request, read its logs, check whether it’s widespread, correlate with recent deploys and runner state. The server is hosted at https://api.fal.ai/v1/mcp/platform and is read-only: it can observe your account but never change it.

You need a fal API key to connect. If you don’t have one yet, create one here.

Setup

Claude Code
Cursor
Other MCP Clients

Run this command in your terminal:

claude mcp add --transport http fal-platform \
  https://api.fal.ai/v1/mcp/platform \
  --header "Authorization: Key YOUR_FAL_KEY"

Open MCP Settings

Use Cmd+Shift+P (Ctrl+Shift+P on Windows) and search for “Open MCP settings”.

Add the fal Platform server

Add the following to your mcp.json file:

{
  "mcpServers": {
    "fal-platform": {
      "url": "https://api.fal.ai/v1/mcp/platform",
      "headers": {
        "Authorization": "Key YOUR_FAL_KEY"
      }
    }
  }
}

Restart Cursor

Save the file and restart Cursor to activate the connection.

The Platform API MCP uses the Streamable HTTP transport at:

https://api.fal.ai/v1/mcp/platform

Authentication is via the Authorization header with the platform’s Key scheme:

Authorization: Key YOUR_FAL_KEY

Any MCP client that supports Streamable HTTP transport can connect. Refer to your client’s documentation for the exact configuration format.

Note the auth scheme: this server uses Authorization: Key YOUR_FAL_KEY (the Platform API’s native scheme), not Bearer. You can connect both fal MCP servers side by side — they complement each other.

Available Tools

The server exposes 15 tools: eleven first-class serverless debugging tools, plus a four-tool discovery gateway that opens up the rest of the Platform API (compute, workflows, keys, account, organization, storage) without bloating your assistant’s context.

Serverless debugging

Tool	What it does
`list_apps`	List your apps with the identifiers every other tool accepts; add `expand=endpoints` for each app’s route-level endpoint ids
`fetch_requests`	List recent requests for your endpoint(s) with status, timings, and optional input/output payloads — the grounding step of any debugging session
`fetch_request_logs`	Search historical app logs; filter by request, level, or free text, with optional stack tracebacks
`get_analytics`	Aggregated error rates (full error taxonomy), latency percentiles, and cold starts per endpoint over a time window
`fetch_events`	Operational event history: deployments, config changes, and runner lifecycle transitions — the “what changed around this time?” step
`list_revisions`	Revision history with deployment status and who deployed — line revision boundaries up against error changes to spot a bad deploy
`get_runner_history`	Historical runner counts by state (running/idle/pending/draining) — capacity analysis and incident forensics
`get_current_state`	Real-time snapshot: runner counts by state, concurrency, and current queue size
`get_usage`	Serverless compute spend and machine-seconds by app (requires an admin-scoped key)
`list_files`	List files in your serverless persistent storage
`read_file`	Read a file from persistent storage (small text artifacts)

Discovery

Tool	What it does
`list_namespaces`	List the Platform API areas beyond serverless, with operation counts
`search_operations`	Keyword-search the full operation catalog
`describe_operation`	Get the parameters and response shape for one operation
`invoke_operation`	Execute any read-only operation from the catalog with your key

Examples

These are quick tastes. For the full guide — the debugging loop, a worked incident example, and a library of prompt recipes — see Debugging with AI.

Debug a failing endpoint

“My requests to my-app started failing about an hour ago — figure out why”

The assistant will:

Use fetch_requests to find the failing requests and their request_ids
Use fetch_request_logs with traceback: true to read the stack traces
Use get_analytics to check whether the failures are widespread or isolated
Use fetch_events to check whether a deploy or config change lines up with when the failures started
Use get_current_state to see if runners are healthy right now

Judge the blast radius

“Is my-app having an incident or was that one bad request?”

The assistant calls get_analytics with the error-count metrics over the incident window and compares error rates against request volume.

Did a deploy break it?

“Errors on my-app jumped Tuesday afternoon — did that start with a deploy?”

The assistant uses list_revisions to see when each revision went out (and who deployed it), then lines those boundaries up against the error timeline from get_analytics.

Why is my queue growing?

“Requests to my-app are taking forever — are we capacity-bound?”

The assistant checks the current queue size with get_current_state, then compares get_runner_history against request volume from get_analytics over the same window: a backed-up queue while runner counts stay flat means undercapacity; runners scaling with traffic means a spike.

Watch spend

“How much did my serverless apps cost this week, and was any of it surge pricing?”

The assistant calls get_usage with a weekly window (admin-scoped key required).

Reach beyond serverless

“List my compute instances”

The assistant uses search_operations("compute instances") → invoke_operation — no serverless-specific tool needed.

How It Works

The server runs inside the fal Platform API and is fully stateless:

Your assistant sends a request to https://api.fal.ai/v1/mcp/platform with your API key
Tool calls execute against the Platform API with that key — normal permissions and ownership scoping apply, so you can only ever see your own account
Results are returned to your assistant

Your API key is sent per-request and never stored. Everything is read-only in the current version: write operations (like flushing a queue) are listed in the catalog but refuse to execute. A few things worth knowing:

Data freshness — request, analytics, and usage data are warehouse-backed and can lag real time by a few minutes. Responses carry an as_of marker, and the server instructs your assistant not to mistake ingestion lag for “no traffic” during a live incident.
Output limits — large responses are truncated with guidance to narrow the query, so a noisy log search won’t flood your assistant’s context.
Key scopes — most tools work with any valid key; get_usage requires an admin-scoped key.

FAQ

How is this different from the Run MCP?
Can it change anything on my account?
Is my API key stored?
Does it cost extra?

The Run MCP at mcp.fal.ai/mcp is for building with models: search the catalog, run inference, upload files. The Platform API MCP is for operating your account: debugging serverless apps, checking spend, inspecting infrastructure. Connect both — they don’t overlap.

No. The server is fully stateless. Your key is sent per-request in the Authorization header and is never stored.

Setting Up

Model APIs

Serverless

Compute

Organizations

Platform MCP

Setup

Available Tools

Serverless debugging

Discovery

Examples

Debug a failing endpoint

Judge the blast radius

Did a deploy break it?

Why is my queue growing?

Watch spend

Reach beyond serverless

How It Works

FAQ

​Setup

​Available Tools

​Serverless debugging

​Discovery

​Examples

​Debug a failing endpoint

​Judge the blast radius

​Did a deploy break it?

​Why is my queue growing?

​Watch spend

​Reach beyond serverless

​How It Works

​FAQ

Setup

Available Tools

Serverless debugging

Discovery

Examples

Debug a failing endpoint

Judge the blast radius

Did a deploy break it?

Why is my queue growing?

Watch spend

Reach beyond serverless

How It Works

FAQ