MCP Modal Server

MCP Modal Server

An MCP server for managing Modal — apps, containers, volumes, and secrets — and for deploying & running Modal apps directly from Claude Code and other MCP clients.

Every tool shells out to your local modal CLI, so it operates against whatever Modal profile and credentials are configured on your machine. There are no extra tokens to manage.

Installation

The server is published on PyPI as mcp-modal. No manual install is needed — the recommended way to run it is with uvx, which fetches and launches it on demand. Just point your MCP client at the command below (see Configuration).

Every version is also tagged and published on the Releases page, with release notes and the same .whl / .tar.gz that PyPI serves attached — useful for pinning, air-gapped installs, or reading what changed between two versions.

Logging in to Modal

This server uses your local Modal credentials. If you haven't authenticated yet, run:

modal setup

This opens a browser to log in and stores a token in ~/.modal.toml. Already logged in elsewhere? Check with modal profile current.

Configuration

Add the server to Claude Code with the claude mcp CLI:

claude mcp add mcp-modal -- uvx mcp-modal@latest

Or add it to a .mcp.json file in your project root, which is the better option for a team — everyone who opens the repo gets the same configuration:

{

"mcpServers": { "mcp-modal": { "command": "uvx", "args": ["mcp-modal@latest"] } } }

Why @latest, and when to pin instead

uvx caches the environment it builds on the first run and does not check PyPI again:

"uvx will use the latest available version of the requested tool on the first invocation.

After that, uvx will use the cached version of the tool unless a different version is

requested, the cache is pruned, or the cache is refreshed."

— uv docs

So a plain uvx mcp-modal means *latest at install time, frozen forever after* — restarting the client or rebooting changes nothing, because the cache lives on disk. Different people end up on different versions depending on when they first ran it, with no warning.

• `mcp-modal@latest` re-resolves on every launch, so a restart picks up new releases. Costs one network round-trip at startup. Use it while the tool surface is still moving.
• `mcp-modal@0.4.0` (an explicit version) is reproducible and upgrades become a deliberate one-line change. Use it once you want stability, or for a wider audience.

To move a machine that is already stuck on an old cached build, switching it to either form above is enough — requesting a version invalidates the cache. Otherwise uv cache clean mcp-modal forces a refresh.

Requirements

• Python 3.11 or higher
• uv (provides uvx)
• Modal CLI 1.5 or newer, configured with valid credentials (modal setup) — 1.5 is where modal billing summary/rates landed and where the billing report switched to snake_case columns; the cost tool reads both spellings but needs 1.5 for those two views
• For Modal deploy and run support:
• The project being deployed/run must use uv for dependency management
• modal must be installed in that project's virtual environment

Security

This server shells out to your local modal CLI using whatever credentials are in ~/.modal.toml. A few tools are powerful by design — if the MCP client driving the server is ever prompt-injected (for example by malicious text inside logs it fetched), these are the escalation paths and should stay behind your client's tool-approval prompts rather than being auto-approved:

• `deploy_modal_app` / `run_modal_app` — execute arbitrary local Python on the host (modal deploy imports the app file; uv run resolves and installs the target project's dependencies).
• `modal_volume_files` with action="put" — can read any local file (e.g. ~/.ssh/id_rsa, ~/.modal.toml) and upload it to a cloud volume (a data-exfiltration primitive).
• `modal_volume_files` with action="get" and force=True — can overwrite any local path (e.g. ~/.zshrc or a shell profile, a persistence primitive).
• `manage_modal_container` with action="exec" — runs arbitrary commands inside a container, by design.

Every tool declares MCP tool annotations, so a client can distinguish the four read-only tools (list_modal_resources, get_modal_logs, search_modal_logs, analyze_modal_costs — all readOnlyHint: true) from the eight that change remote state or start compute. Six of those eight are destructiveHint: true; the exceptions are run_modal_app and inspect_modal_secret, which start compute without removing or overwriting anything. Auto-approve the reads; keep the rest behind a prompt.

Optional local-path allowlist

To contain the two filesystem-touching volume tools, set the MCP_MODAL_ALLOWED_LOCAL_PATHS environment variable to an os.pathsep-separated list of directories (: on macOS/Linux). When it is set, modal_volume_files is refused for any local path — local_path on action="put", the destination on action="get" — unless the resolved path, after expanding ~ and collapsing ../symlinks, falls inside one of those roots. The download target "-" (return contents instead of writing a file) is exempt because nothing is written to disk.

When the variable is unset (the default) there is no restriction, so existing setups are unaffected. Configure it in your MCP client, e.g.:

{

"mcpServers": { "mcp-modal": { "command": "uvx", "args": ["mcp-modal"], "env": { "MCP_MODAL_ALLOWED_LOCAL_PATHS": "/Users/me/modal-workspace:/tmp/modal" } } } }

All tools also pass user-supplied names/paths after a -- end-of-options separator, so a value beginning with - is always treated as data, never as a modal CLI flag. Secret values handed to manage_modal_secret are redacted from the echoed command, logs, and any error output.

Supported Tools

12 tools. Related operations are grouped behind an action/resource argument rather than split one-per-CLI-subcommand: every tool schema is loaded into the model's context for the whole session, so a smaller surface leaves more room for your actual work (and gives the model fewer near-identical tools to choose between).

Tools that talk to environment-scoped resources take an optional env argument to target a specific Modal environment; if omitted, they use the profile's default (or MODAL_ENVIRONMENT). The exception is manage_modal_container and container logs — a container ID is globally unique and the CLI accepts no environment there.

Read-only

1. List Modal Resources (list_modal_resources) — one lookup for the whole account.

• Parameters: resource (required), name, path (default /), env
• resource values:

• volume_files sets empty: true with a message when a listing genuinely returns nothing, so an empty directory is distinguishable from a wrong path.
• Listings over 200 entries are capped, with omitted_items giving the number dropped.

1. Get Modal Logs (get_modal_logs) — fetch or stream logs for an app *or* a container.

• Parameters: identifier (required), target (auto/app/container, default auto — anything starting ta- is a container), timeout_seconds (default 30), env, since, until, tail, source (stdout/stderr/system), timestamps, follow
• since without tail fetches *every* entry in the range; pass until as well (max range 35 days, tail max 20,000) to keep a busy app's output bounded.
• With follow=True, logs stream until the app/container stops or timeout_seconds is reached, returning a snapshot with truncated: true.
• Only covers the stdout/stderr/system streams; some failures (e.g. a crash reported as "... exited with ...") are Modal dashboard events, not log lines, and won't appear here.

1. Search Modal Logs (search_modal_logs) — grep logs and get each hit with the surrounding lines, built for "where did it go wrong?" debugging. Logs are fetched once and searched locally, so you get context, regex, case control, and exact match counts.

• Parameters: identifier (required), pattern (required), target (default auto), regex, case_sensitive, context_lines (default 3), max_matches (default 50), since, until, tail (defaults to the last 1000 entries), source, exclude (drop noise lines before searching, e.g. "queue put failed"), prefilter, timestamps (default true), timeout_seconds, env
• Bound the window on a busy app. since on its own fetches everything from then until now — hundreds of KB per hour on a chatty app, which the 30s fetch cuts off (logs_truncated: true) and the output budget trims. since *and* until around the minute you care about is the fix, and is usually kilobytes.
• prefilter=True pushes pattern down to Modal as a server-side substring filter (modal app logs --search), so non-matching lines are never fetched — the lever for logs too large to drain. Requires regex=False, and context lines then show only other matches, so use it to locate the window and re-query it with prefilter=False.
• Returns match_count and matches: timestamped, line-numbered context blocks where matched lines are prefixed with >, e.g. > 8: 2026-06-04T... ValueError: bad input. The whole fetched log is always searched, so match_count stays exact even when fewer blocks are returned. returned is how many matches came back (adjacent matches merge into one block, counted by returned_blocks). Reports excluded_lines when exclude is used.
• A window the CLI rejects (reversed range, over 35 days, tail over 20,000) comes back as success: false with Modal's own message, not a bare exit code.
• Same stdout/stderr/system-only caveat as get_modal_logs.

Deploy & run

1. Deploy Modal App (deploy_modal_app)

• Deploys a Modal app (modal deploy). Deployed web endpoints persist, so any links in the output are live and shareable (returned in urls).
• Parameters: absolute_path_to_app (required), env, name, tag, strategy (rolling/recreate), stream_logs
• The app's directory must use uv with modal installed in its virtualenv.

1. Run Modal App (run_modal_app)

• Runs a function or local entrypoint once and collects its output (modal run).