make-runner-mcp

Disclaimer — read before using. This is a personal convenience tool for running make commands in your own projects through an AI coding agent, built to narrow what that agent can do compared to raw shell or Docker socket access. It is not a hardened security boundary, it has not been independently audited, and it is not a substitute for actually reviewing what's in your Makefile. Don't use it in any context where the blast radius of a mistake matters — shared infrastructure, production systems, other people's data, or anywhere you wouldn't already trust the Makefile itself to be run unattended. See "Risks and limitations" below before relying on it for anything beyond that.

Scoped MCP server that discovers a project's Makefile targets and exposes each one as an individual MCP tool. No free-form shell/docker access is ever given to the model — only targets that literally exist in the project's own Makefile, minus anything hard-denied (deploy/destroy/prod/ publish/release) or excluded via .mcp-make-config.json.

Install

Nothing to install ahead of time for end users — each client config below runs the server straight from this GitHub repo via npx, pinned to a release tag. npx fetches and caches it on first use.

To work on the server itself (e.g. to run it locally or through the inspector below):

npm install

Two ways to run this — pick based on whether your agent is sandboxed

MCP_TRANSPORT=http is the default. This is a breaking change from earlier versions, which defaulted to stdio: a client config that spawns this as a subprocess expecting a stdio handshake (the old default) now needs MCP_TRANSPORT: "stdio" added to its env block explicitly, or it'll get an HTTP server trying to bind a port instead of speaking MCP over its own stdin/stdout.

• stdio (opt-in) — the client spawns this as a subprocess and speaks MCP over its stdin/stdout. Simplest option: no port, no token, no process to manage yourself. Use this for direct, unsandboxed use — an agent running with normal access to your machine. See Client configs (stdio) below.

• http (default) — this runs as its own persistent process, with clients connecting to it over the network instead of spawning it. Use this when the calling agent runs inside its own sandbox (e.g. Gemini CLI's --sandbox): a sandboxed agent that spawns an MCP server via stdio spawns it inside its own sandbox container, which means the server (and anything it shells out to, e.g. docker) only has whatever access that container has — typically none, and giving it more (like a mounted Docker socket) hands that same access to the agent's own native shell tool too, since the sandbox isn't scoped per-tool. Running this server outside the sandbox instead, reached only over the network, means the sandbox never needs Docker access at all — only this server does, and it only ever runs the target you named. See Running as a persistent HTTP server below — this is a real, verified pattern, not a theoretical one.

Verify it works standalone

MCP_TRANSPORT=stdio npx @modelcontextprotocol/inspector node server.js

This opens a local UI to list tools and call them directly, useful for confirming the Makefile parsing looks right before wiring it into an agent.

Client configs (stdio)

The server is the same everywhere — only the config file and its shape differ per client. Things that change between projects/teams:

• PROJECT_DIR — the target project's root.

• #v2.0.2 — bump this to whatever tag you've actually released. Always pin to a release tag, never #main: for a tool that executes commands, an unpinned branch reference means a bad push could silently change what runs on everyone's machine.

• MCP_TRANSPORT: "stdio" — required in every block below, since http is now the default and a stdio-spawned subprocess needs to opt back in explicitly.

Gemini CLI

.gemini/settings.json (project) or ~/.gemini/settings.json (user-wide):

{
  "mcpServers": {
    "makeRunner": {
      "command": "npx",
      "args": ["-y", "github:davindermahal/make-runner-mcp#v2.0.2"],
      "env": { "PROJECT_DIR": "/path/to/project", "MCP_TRANSPORT": "stdio" }
    }
  }
}

Claude Code

Project-level .mcp.json at the repo root (checked in, shared with the team):

{
  "mcpServers": {
    "makeRunner": {
      "command": "npx",
      "args": ["-y", "github:davindermahal/make-runner-mcp#v2.0.2"],
      "env": { "PROJECT_DIR": "/path/to/project", "MCP_TRANSPORT": "stdio" }
    }
  }
}

Or via the CLI: claude mcp add makeRunner -e PROJECT_DIR=/path/to/project -e MCP_TRANSPORT=stdio -- npx -y github:davindermahal/make-runner-mcp#v2.0.2

Claude Desktop

claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "makeRunner": {
      "command": "npx",
      "args": ["-y", "github:davindermahal/make-runner-mcp#v2.0.2"],
      "env": { "PROJECT_DIR": "/path/to/project", "MCP_TRANSPORT": "stdio" }
    }
  }
}

Cursor

.cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "makeRunner": {
      "command": "npx",
      "args": ["-y", "github:davindermahal/make-runner-mcp#v2.0.2"],
      "env": { "PROJECT_DIR": "/path/to/project", "MCP_TRANSPORT": "stdio" }
    }
  }
}

Any other MCP-compatible client

All of these follow the same command / args / env shape because it's part of the MCP spec's stdio transport — if a client supports MCP at all, this same block (adjusted to that client's config file location, with MCP_TRANSPORT: "stdio" added to env) will work without touching the server itself.

Running as a persistent HTTP server (for a sandboxed agent)

Verified end to end against a real sandboxed gemini --sandbox session driving a real Docker-based project (PHPUnit and Composer both ran inside the project's container, through this server, with the sandbox itself having no docker CLI or socket at all — the model's own native shell tool inside that same session had zero Docker access, confirmed with which docker returning nothing).

1. Run the server as its own process, outside any sandbox, with normal access to whatever the project's targets need (Docker included — nothing special, just however you'd normally run docker/docker compose on this machine):

PROJECT_DIR=/path/to/project \
MCP_HTTP_TOKEN=$(openssl rand -hex 24) \
MCP_HTTP_PORT=8791 \
npx -y github:davindermahal/make-runner-mcp#v2.0.2

Generate a real random token (openssl rand -hex 24 or equivalent) and keep it — the server refuses to start without one (fails closed, not silently unauthenticated), and every client needs it to connect. Keep this process running for as long as you want the tool available; it's not spawned per-client the way stdio mode is.

2. Point the client at it over the network, not a subprocess spawn. For a client whose agent itself runs sandboxed (Gemini's --sandbox, which auto-maps host.docker.internal to the host — verified live, zero extra network config needed), use that hostname so the sandboxed process can reach a server running on the host:

{
  "mcpServers": {
    "makeRunner": {
      "url": "http://host.docker.internal:8791/mcp",
      "type": "http",
      "headers": { "Authorization": "Bearer <the token from step 1>" }
    }
  }
}

For an unsandboxed client on the same machine as the server, 127.0.0.1 works the same as any other local service.

Security notes specific to this mode (also see Risks and limitations):

• MCP_HTTP_HOST defaults to 0.0.0.0 — reachable from your local network, not just the sandbox, unless your host firewall restricts it. The bearer token is the actual access control; treat it like any other credential (don't commit it, don't reuse it across machines/projects).

• Every project you run this way needs its own token and, generally, its own port — there's no per-project isolation beyond that; anyone who has the token for a given running instance can call any tool it exposes.

Publishing to npm instead

Shipping via GitHub tags (above) is the default — it's a public repo, so npx github:... works with zero auth setup on any teammate's machine, and tagging releases keeps updates a deliberate, visible choice rather than whatever main happens to be. Publishing to npm is a possible later step (marginally faster resolution, no GitHub dependency) but isn't required to ship this:

npm publish            # or: npm publish --registry <your-private-registry>

Once published, every config above can drop the github:...#tag args entry in favor of ["-y", "make-runner-mcp"] (add @<version> to pin, same reasoning as pinning the GitHub tag).

Passing arguments through to a target

Extra args on a tool call may only be make flags (from a small safe allowlist — -n, -k, -s, etc.) or VAR=value assignments. Bare positional arguments (e.g. an extra target name, or a package name/flag with no VAR= in front of it) are always rejected — make treats a bare trailing word as an additional build goal, not data, so allowing them would let a call to one target silently smuggle a second, possibly-denied target onto the same invocation.

For a target whose underlying command takes its own arguments (composer install <pkg> --no-dev, npm run <script> -- <flags>, etc.), have its recipe read a single variable and pass one VAR=value pair at call time:

composer: ## Run composer inside the app container, e.g. ARGS="install symfony/console --no-dev"
	docker compose exec app composer $(ARGS)

Called with args: ["ARGS=install symfony/console --no-dev"]. The value half of a VAR=value pair is checked against a more permissive (but still shell-metacharacter-free) pattern than other arguments specifically to support this — letters, digits, spaces, and typical package-name/version/ flag punctuation are allowed; shell control characters (; & | $ \ ' " < > ( ) # \ * ? [ ], newlines) are not, since make` ultimately splices this value into a recipe line that runs through a real shell.

This keeps the "one call always executes exactly the target it names, in addition to whatever its own vetted recipe does with $(ARGS)" guarantee intact — the passthrough content is only ever data to the recipe you already wrote, never a way to select a different target.

If a project's Makefile currently relies on $(MAKECMDGOALS) or a similar bare-word passthrough trick (common for wrapping tools like composer/npm), see MAKEFILE-GUIDE.md — a self-contained guide written to be handed directly to an AI coding agent, with instructions to find and convert those targets to the ARGS pattern above.

Per-project overrides

Drop an optional .mcp-make-config.json in a project's root to narrow which targets are exposed beyond the built-in denylist, and/or restrict which environment variables make runs with:

{
  "deny": ["clean-volumes", "seed-prod"],
  "allow": null,
  "envAllowlist": null
}

allow, if set, becomes an explicit whitelist — only those exact target names are ever exposed, regardless of what else is in the Makefile.

envAllowlist, if set, restricts the environment make (and everything its recipes run) sees to just those variable names — useful since by default the child process inherits this MCP server's entire environment, which in an agent sandbox often includes API keys or tokens that would otherwise be readable, and potentially echoable back into the tool result, by any recipe. Leave it unset (the default) for unrestricted passthrough, matching prior behavior.

A config file that fails to parse, or has an allow/envAllowlist entry that isn't a plain array of strings, fails closed — every target is denied (or every env var withheld) rather than silently falling back to no restriction at all.

Risks and limitations

This tool reduces risk compared to giving an agent raw shell or Docker access. It does not eliminate it, and it is not a security product. Read this before pointing it at anything you'd be upset to lose or break.

• HTTP mode opens a network port that can execute commands. It's authenticated (a required bearer token, fails closed if unset) but bound to 0.0.0.0 by default — reachable from anything that can route to that port, not just your intended sandboxed agent, unless your host firewall restricts it. On a shared network, either firewall the port or bind MCP_HTTP_HOST more narrowly. The token is the real access control; don't commit it, don't log it, don't reuse one across machines/projects.

• It's only as safe as your Makefile. The built-in denylist blocks five words (deploy, destroy, prod, publish, release) and a rm- prefix — it does not evaluate whether a target is actually dangerous. If your Makefile has a target that wipes a volume, drops a database, or force-pushes something, and its name doesn't happen to contain one of those words, this tool will expose it to the agent like any other target. The real protection this tool provides is narrower than "safe by default": it's "only what's already in a file you wrote and can read," nothing more. Review your Makefile with that in mind — and use .mcp-make-config.json's deny/allow to explicitly exclude anything you don't want an agent invoking, don't rely on the built-in denylist alone.

• The ARGS="..." passthrough convention blocks shell injection, not dangerous values. The character set allowed in a VAR=value pair stops shell metacharacters (;, |, `, etc.) from reaching a recipe's shell, but it does not know what your recipe does with that value. A recipe like rm -rf $(ARGS) would still accept a value like -R 777 / — every one of those characters is allowed. If you write a target that consumes $(ARGS), treat that value as agent-controlled input and keep the underlying command it's handed to narrow and low-risk (a package manager, a test runner) rather than anything destructive.

• The full environment is passed through by default. Unless you set envAllowlist in .mcp-make-config.json, every recipe runs with this server's entire environment — including any API keys or tokens present in the agent's sandbox. Those become readable, and potentially echoable back into the tool's output, by any recipe. Set envAllowlist per project if that matters to you.

• Recipe output isn't filtered for secrets. Whatever a target prints to stdout/stderr is returned to the calling agent close to verbatim (truncated only by size, at 200KB). If a recipe's own logs happen to print something sensitive, this tool won't catch or redact it.

• Protected variables cover make's own built-ins, not project-specific conventions. Overriding PATH, SHELL, and similar is blocked. A project-specific convention like CC=, DOCKER=, or COMPOSE= used as an executable name inside a recipe has no equivalent protection — there's currently no way to extend that list per project.

• This has not been independently security-reviewed or audited. It's been built and tested for one person's own use across their own projects, not pentested or reviewed by a third party. Read the source (server.js is a single, deliberately short file) before trusting it with anything you care about.

• It's designed for a trusted human running their own projects, not for isolating untrusted users from each other. There's no concept of per-user permissions, auditing, or multi-tenant isolation. If you need any of that, this isn't the right tool as-is.

If any of this is a dealbreaker for your use case, the honest fix is either: don't expose the target in question at all (deny/allow it out), rewrite the target's recipe to be inherently safe regardless of what $(ARGS) contains, or don't use this tool for that project.

Gemini CLI specifically — tested, not theoretical

The following was verified against a real gemini CLI session (transcript inspection, not just reading the final answer), because it changes what "safe" means for this tool with that client:

• When make-runner-mcp fails to connect for any reason, Gemini CLI does not fail loudly. It silently falls back to its own native read_file and run_shell_command tools and accomplishes the same task via full, unrestricted shell access — with no clear signal beyond a generic "MCP issues detected" banner that also fires for unrelated, pre-existing broken servers. I triggered this twice, independently: once from an untrusted project folder, once from an unrelated npm cache permission error inside --sandbox. Both times the model read the Makefile directly and ran the command itself, completely bypassing make-runner-mcp, and the final answer looked identical either way — you cannot tell from the response alone whether the tool boundary was actually enforced. This means make-runner-mcp's guarantee only holds while the MCP connection is actually up; it does not degrade safely.

• A project folder must be in Gemini CLI's persistent trust store before it will load any project-level .gemini/settings.json MCP servers — the --skip-trust CLI flag alone is not sufficient for this, even though it looks like it should be. Without that trust, make-runner-mcp is silently never loaded at all (see the point above for what happens next).

• --sandbox mode and stdio-spawned Docker-based targets don't mix — this is exactly why http is now the default transport. Gemini's --sandbox re-execs itself (and everything it spawns via stdio, MCP servers included) inside its own container. The default sandbox image (gemini-cli/sandbox) does not include the docker CLI at all — a stdio-spawned target that shells out to docker/docker compose fails with make: docker: No such file or directory, verified verbatim. The tempting-looking fix — a custom sandbox image with docker CLI installed and the host's Docker socket mounted in — doesn't actually preserve any scoping: Gemini's sandbox is one shared container for the entire session, not scoped per-tool, and the native run_shell_command fallback (previous bullet) runs inside that same container. Mounting the socket in would hand that same unrestricted native shell tool identical Docker access, with no separation between "the vetted make targets" and "whatever command the model decides to run directly" — the exact Docker-outside-of-Docker exposure this project exists to avoid (see CLAUDE.md).

  The actual fix, verified end to end: run this server in http mode (the default) outside the sandbox — as its own process on the host, with normal Docker access — and let the sandboxed session reach it only over the network via Gemini's auto-mapped host.docker.internal. The sandbox container itself then never has Docker access at all — no CLI, no socket — so there's nothing for the native shell tool to fall back to, and Docker access exists exclusively through this server's vetted targets. Confirmed live: mcp_makeRunner_make__unit-test and mcp_makeRunner_make__composer both correctly drove a real project's Docker container (real PHPUnit and Composer output) through a sandboxed session that had zero docker CLI of its own (which docker returned nothing in that same session). See Running as a persistent HTTP server above for the exact setup.