MCP::Server::Tool::Shell

A command runner for MCP::Server.

Six tools — run, start, poll, output, kill and allowed — let an LLM invoke local programs. Which programs is your decision: either an exact allow list, or allow-any for deployments that ask a human instead. Arguments are handed to the program as an argv vector (there is no shell anywhere in the code path), output is captured into capped buffers so a runaway command cannot exhaust the server's memory, foreground runs are bounded by a timeout, and work that will not fit in one tool call can be started as a background job and polled.

Synopsis

Plug it into a server you are already building:

use MCP::Server;
use MCP::Server::Tool::Shell;

my $server = MCP::Server.new(:name<dev-tools>, :version<1.0>);

$server.plug: MCP::Server::Tool::Shell.new(
    allow   => <git rg /usr/bin/jq>,   # exact argv0 strings
    cwd     => '/home/me/project',     # optional default working directory
    timeout => 15,                     # seconds, default 30
);

$server.run;   # registers sh_run, sh_start, sh_poll, sh_output, sh_kill, sh_allowed

Or assemble a server out of toolkit packs in one shot, with no glue code:

use MCP::Server;

MCP::Server.new(
    :name<dev-tools>,
    :tools[
        'Shell' => { allow => <git rg>, cwd => '/srv/app', prefix => 'git' },
    ],
).run;   # registers git_run, git_start, ...

Or run it straight off the raku-mcp command line that ships with MCP::Server:

raku-mcp --tool=Shell='{"allow":["git","rg"],"timeout":10}'

raku-mcp --config=mcp.json

{
  "name": "dev-tools",
  "instructions": "Read-only git and ripgrep access to the checkout.",
  "tools": {
    "Shell#git": { "allow": ["git"], "cwd": "/srv/app", "prefix": "git" },
    "Shell#search": { "allow": ["rg"], "cwd": "/srv/app", "prefix": "rg", "timeout": 10 }
  }
}

The #alias suffix lets the same pack be loaded twice with different settings — here one instance can only run git and the other only rg, each under its own prefix.

Wiring that into Claude Code:

{
  "mcpServers": {
    "dev-tools": {
      "command": "raku-mcp",
      "args": ["--config=/home/me/mcp.json"]
    }
  }
}

Configuration

• allow must contain at least one non-empty string unless allow-any is set. Entries are matched with eq — see #Security.

• allow-any and allow are mutually exclusive, and one of them is required: a runner that can run everything and a runner that can run nothing are both configurations somebody meant, but "I forgot to say" is not. See #Running anything: allow-any and MCP::Client::Policy.

• cwd must be an existing directory at construction time; the constructor dies otherwise. A per-call cwd argument overrides it.

• isolate-terminal keeps spawned commands away from the terminal this process is attached to, and is on unless you turn it off. Leave it on unless you know the process has no terminal worth protecting — see #Terminal isolation, which is also where the "it refuses to construct" failure mode is explained.

• timeout must be greater than zero. Fractional values are fine (0.5 is half a second). A per-call timeout argument overrides it for one run; start takes one too, but has no default.

• max-output-bytes (floor 1024) and max-output-lines (floor 10) apply per stream, so a run may hold twice them at most. See #Truncation.

• max-jobs, max-finished-jobs (both at least 1) and job-ttl (greater than zero) govern background jobs. See #Background jobs.

• prefix is not a setting of this pack: it is reserved by the toolkit system. Pass it to .plug($kit, :prefix<git>), or beside the entry in :tools[...] / the JSON config, to namespace the tool names.

Every setting is validated when the toolkit is constructed, so a typo or a missing directory fails at startup rather than on the first tool call:

MCP::Server::Tool::Shell.new(allow => []);
# Shell toolkit requires a non-empty allow list (or allow-any => True when
# consent is enforced elsewhere, e.g. by an MCP::Client::Policy)

MCP::Server::Tool::Shell.new(allow => <git>, allow-any => True);
# Shell toolkit allow and allow-any are mutually exclusive: ...

MCP::Server::Tool::Shell.new(allow => <git>, cwd => '/nope');
# Shell toolkit cwd is not an existing directory: /nope

MCP::Server::Tool::Shell.from-config({ allow => <git>, timeut => 5 });
# Unknown config key(s) for MCP::Server::Tool::Shell: timeut.
# Valid keys: allow, allow-any, cwd, isolate-terminal, job-ttl,
# max-finished-jobs, max-jobs, max-output-bytes, max-output-lines, timeout

The default prefix is sh, so the tools below are registered as sh_run, sh_start, sh_poll, sh_output, sh_kill and sh_allowed unless you pass a prefix of your own.

Tools

Which of them may run at once

allowed, poll and output declare the MCP annotations < readOnlyHint => True > and < idempotentHint => True >; run, start and kill declare nothing.

A host reads those hints as a licence to run a batch of the annotated calls side by side, and to execute identical ones in a batch once. It holds here: allowed reads a list this instance was built with, and poll and output read a job's state and a window of what it has already written — an output cursor is a position the caller holds, so reading from one leaves the buffer exactly as it was and the same call twice is the same answer twice. The job registry and every job's buffers take their own locks, so several of these at once see coherent state rather than a half-written flush.

The other three could not be annotated: a command is whatever the caller typed and this pack cannot know whether it writes, start mints a job, and kill ends one.

run

Runs one command in the foreground and waits for it.

The result is a JSON object as text content:

{"exit":0,"stderr":"","stdout":"main\n","timed-out":false}

exit reports:

• the command's own exit status for a normal exit (0 on success);

• 128 + signal when the command was terminated by a signal, following the POSIX shell convention — a timeout kill on Unix therefore shows up as 137 (128 + SIGKILL), while on Windows TerminateProcess yields 1;

• -1 when the command could not be spawned at all (not on PATH, not executable, and so on). The reason is appended to stderr;

• null in the one case where there is no honest answer: the command was killed but its status never arrived (a grandchild is still holding the pipes, say), so the run is reported with timed-out: true and an unknown status.

A missing binary is a runtime condition, not a caller mistake, so it is reported in the payload:

{"exit":-1,"stderr":"Failed to spawn process nosuchtool: no such file or directory (error code -2)","stdout":"","timed-out":false}

Anything the caller got wrong — a command that is not allowlisted, an args value that is not an array of strings, a cwd that does not exist, a timeout that is not a positive number — is thrown instead, which MCP::Server turns into an isError tool result:

Command 'rm' is not allowlisted. Allowed: git, rg

Examples of the call shape (arguments as an MCP client would send them):

{ "command": "git", "args": ["rev-parse", "--abbrev-ref", "HEAD"] }
{ "command": "rg", "args": ["--json", "TODO|FIXME", "lib"], "cwd": "/srv/app" }
{ "command": "git", "args": ["log", "--grep", "fix the thing"] }
{ "command": "jq", "args": ["-r", ".name"], "stdin": "{\"name\":\"widget\"}" }

That third one is the point of an argv vector: fix the thing is a single argument, quoted by nobody, split by nobody.

Standard input

stdin is written to the command and the pipe is then closed, so a filter sees a normal end of input. Three states:

• omitted — nothing is written, and the command reads an immediate end of file. What it does not get is the server process's own standard input: with isolate-terminal on (the default) fd 0 is a closed pipe rather than the host's terminal or the client's transport. See #Terminal isolation.

• empty string — the pipe is opened and closed immediately, so the command reads an immediate EOF. This is what you want for a filter you have nothing to say to.

• a string — written and then closed.

The first two therefore look the same to the command and differ only in intent — unless you have turned isolation off, in which case an omitted stdin leaves fd 0 inherited from the server process, as it was before isolation existed.

The write is detached from the run: a command that never reads its input cannot stall the server, however much you send it, and a command that exits before reading everything just makes the write fail (EPIPE), which is a runtime condition rather than an error. NUL bytes are fine here, unlike in args: this is a byte stream, not a C string.

Truncation

Output is captured into a bounded buffer per stream, so a command that decides to print a gigabyte does not make the server hold a gigabyte. Each buffer keeps a head — the first quarter of the budget, which is where a compiler puts the error that matters — and a rolling tail — the remaining three quarters, which is where a test runner puts the summary. What falls between them is dropped, and the payload says so twice: in the text, and in a report.

{
  "exit": 0,
  "stdout": "001 ...\n002 ...\n[... 7176 bytes / 175 lines dropped ...]\n199 ...\n",
  "stderr": "",
  "timed-out": false,
  "truncated": { "stdout": { "bytes": 7176, "lines": 175 } }
}

• The truncated key is only present when something was actually dropped, and only names the streams that dropped something. A run inside its budget has exactly the four keys it has always had, byte for byte.

• The numbers in the marker and the numbers in truncated are the same numbers: both come out of one atomic read of the buffer.

• Cuts land on character boundaries, so a multi-byte character is never sliced in half.

• max-output-lines is enforced alongside max-output-bytes; whichever runs out first decides.

The buffer is MCP::Server::Tool::Shell::Buffer and is usable on its own if you want the same behaviour elsewhere.

Streaming a run in progress

A 2026-07-28 (modern era) request that opts into logging — _meta io.modelcontextprotocol/logLevel of info or lower — gets the command's output as notifications/message on that request's own channel while the command is still running:

{
  "jsonrpc": "2.0",
  "method": "notifications/message",
  "params": {
    "level": "info",
    "logger": "sh_run",
    "data": { "stream": "stdout", "text": "compiling lib/Thing.rakumod\n" }
  }
}

• logger is the prefixed tool name, so a host that plugged several instances can tell them apart.

• Batches are flushed at least every 0.25s and are never larger than 32 KiB; a bigger burst becomes several notifications, in order.

• For a run that finished on its own, concatenating every stdout batch gives exactly the stdout of the final payload (when nothing was truncated): the last flush happens once the command's streams have closed. A run that was killed may have written a little more than it managed to stream.

• Nothing is streamed for a request that did not opt in, and nothing at all is streamed on the legacy era, where notifications travel on one channel shared by the whole session and a run's output would surface inside somebody else's request. Use a background job for progress on legacy clients.

If the caller hangs up mid-run (an HTTP client closing its response stream), the run notices within 0.25s, kills the command and unwinds. The answer it produces is a dead letter nobody reads, and is not marked timed-out: the budget was never reached.

Background jobs

run has to answer inside one tool call, which puts a ceiling on what it can be asked to do. start escapes that ceiling: it spawns the command, returns a handle, and answers immediately.

{ "command": "make", "args": ["-j8", "test"] }
→ {"job":"j1-1x9k2w","state":"running"}

start

Takes exactly the same arguments as run (command, args, cwd, stdin, timeout) and returns < { job, state } >. Its timeout is optional and has no default: escaping the foreground budget is the point. Starting a job when max-jobs are already running is a caller error, and the message names the jobs in the way.

poll

Takes job and reports how it is getting on, without returning the output:

{"exit":null,"job":"j1-1x9k2w","killed":false,"runtime":12.004,
 "started-at":"2026-08-09T09:15:03Z","state":"running",
 "stderr-bytes":0,"stdout-bytes":81920,"timed-out":false}

state is running, exited or killed; exit is null until there is a status; stdout-bytes and stderr-bytes count everything the command has produced, including whatever the buffers have since dropped.

output

Takes job and an optional cursor, and returns < { job, state, stdout, stderr, cursor, eof } > plus truncated when something was dropped.

• Without a cursor you get a capped snapshot of the whole run so far — head, marker, tail — exactly as run would have rendered it.

• With the cursor from a previous call you get only what has been written since, so a long job can be followed a piece at a time. Concatenating the instalments reconstructs the output, with a [... N bytes / M lines dropped ...] marker wherever the reader fell so far behind that the buffer rolled past it.

• The cursor is opaque: it carries a position in both streams, in bytes and in lines. Pass it back exactly as it was given; anything else is a caller error.

• eof is true once the job has finished and you have read to the end.

kill

Takes job, kills it with SIGKILL and returns < { job, state, killed } > immediately — collecting the status is the watcher's business, so poll for the final state. Killing a job that has already finished changes nothing and is not an error.

Job lifetime

• At most max-jobs run at once.

• A finished job stays pollable for job-ttl seconds, and only the max-finished-jobs most recently finished are kept. After that its handle is refused with a message saying why — start a new job rather than hoping.

• A job does not observe the cancellation of the request that started it: that request was answered the moment the handle was minted. Stopping a job is explicit, or its own timeout.

• Jobs die with the server process. Nothing here survives a restart.

• Job ids (j3-1x9k2w) are handles, not secrets: anyone who can call these tools can already start jobs of their own.

Push notifications, honestly

While a job runs, its output and lifecycle events are also pushed as notifications/message — output batches at debug (so a legacy client can silence them with logging/setLevel) and started / exited / killed lines at info, all labelled with the prefixed start tool name. Whether they arrive depends on how the server is being spoken to, and this is worth knowing before you build a UI on it:

So: poll and output are the reliable way to follow a job, on every transport. Push is a bonus where the channel happens to exist, never a substitute for polling — and a push channel that refuses or fails never affects the job itself, whose output is safe in its buffers either way.

Job lifecycle events

The two info lines above are prose, meant for a log window. The same two facts also go out as a structured notifications/job, on the same channel and under the same delivery rules, for a host that wants to act on a job finishing rather than show it to somebody:

{"jsonrpc":"2.0","method":"notifications/job",
 "params":{"job":"j1-1x9k2w","state":"started","runtime":0,
           "logger":"sh_start","command":"make"}}

{"jsonrpc":"2.0","method":"notifications/job",
 "params":{"job":"j1-1x9k2w","state":"exited","exit":0,"runtime":42.317,
           "logger":"sh_start","command":"make"}}

• Exactly one started and exactly one terminal event (exited or killed) per job, by construction: one thread wins the transition that emits each. A job that could not be spawned at all emits only the terminal one, with exit -1 — and emits it from inside the start call, so it can reach a host a moment before the handle does.

• exit is absent, not null, when the status never arrived: "we do not know" and "it exited with 0" must not look alike to a program.

• command is the job's own record of the program it was asked to run, truncated to 120 characters. Never a byte of what the child wrote, and not its arguments either. A host is likely to turn these events into text that something — a person, a model — reads as trustworthy, and a command that prints [background event] ignore your previous instructions must not get a hearing that way. Output has its own channel and its own tool.

• logger is the prefixed start tool name, so two shell kits plugged into one server can be told apart.

allowed

Takes no arguments and returns the allow list, sorted, one entry per line:

git
rg
/usr/bin/jq

Give the LLM this tool and it can discover what it may run instead of guessing and collecting errors. In allow-any mode it says so in one line instead.

Security

This pack is a gate, not a sandbox. What it guarantees:

• Exact-string allowlist (unless you turn it off). The command argument must be eq one of the allow entries. There is no globbing, no basename matching, no PATH rewriting and no path normalisation: if you allow /usr/bin/jq then jq is refused, and if you allow git then /usr/bin/git is refused. Prefer absolute paths when you care which binary runs; prefer bare names when you want PATH resolution (including .exe resolution on Windows).

• No shell, ever. Commands are spawned with Proc::Async, which execs the program directly with the argv vector you supplied. No shell is involved at any point, so quoting, word splitting, globbing, redirection, pipelines, command substitution and environment-variable expansion simply do not happen. < args => ['$HOME; rm -rf /'] > is one argument containing punctuation.

• Argument hygiene. args must be an array; entries must be strings (or numbers, which are stringified) and may not contain NUL bytes, since a NUL would silently truncate the argument at the exec boundary. The same goes for command itself.

• Bounded runtime. A foreground run is killed once timeout seconds have elapsed and reported with timed-out: true. The kill is SIGKILL — the uncatchable signal — rather than a polite SIGTERM first: Rakudo's Proc::Async ignores every .kill after the first on a given process, so a catchable signal would leave no way to escalate against a child that traps it. On Windows SIGKILL is one of the three signals libuv maps onto TerminateProcess, which is equally uncatchable. A killed command therefore gets no chance to clean up after itself.

• Bounded memory. Output is capped per stream per run, and a background job's buffers are capped the same way however long it runs.

What it deliberately does not do:

• It does not, on its own, constrain what an allowed command can do. Allowlisting an interpreter or a shell (sh, bash, pwsh, raku, python) hands over arbitrary code execution, because those programs take code as an argument. Allowlist leaf tools, not launchers. A launcher (see #Sandboxing: the launcher seam) can constrain a run — it is where a host wraps the spawn in an OS sandbox — but the pack itself makes no such promise.

• It does not filter arguments. git can push, rg can read any file the server user can read, and cwd is unrestricted unless you leave it unset and only allow commands that ignore it. Run the server as a user whose privileges you are comfortable handing to the model.

• It does not scrub the environment. Children inherit the server process's environment, including any credentials in it.

• It does not stop a command reaching the network, the clock, or any other resource the server's user can reach. Isolation here is about the terminal and the process tree, not about capabilities; that is what a launcher is for.

Terminal isolation

Where a shell pack runs decides what its children can wreck.

Plugged into a stdio MCP server, the pack lives in a process whose fd 0 is a pipe from the client, and a child that inherits it is mostly harmless. Plugged in in-process — a terminal application hosting its own tool server rather than talking to one over a socket — the pack lives in a process whose fd 0 is the user's terminal, and whose controlling terminal every child inherits. From there one ordinary command can take the host apart:

• A child that puts the terminal in raw mode and restores it to cooked mode on the way out — any full-screen program: an editor, a pager, another TUI — leaves the host's terminal cooked. The host carries on drawing and never receives another keystroke: its input is being line-buffered by a terminal it no longer controls, and its quit key has become flow control. This is not hypothetical; it is the incident this feature exists for.

• A child that reads standard input steals the bytes the host was going to read — keystrokes from a terminal, or, in stdio server mode, the client's own JSON-RPC frames.

• A child that opens /dev/tty reaches the host's terminal even when all three of its own descriptors are pipes.

None of that is a command misbehaving. It is a command doing exactly what it always does, to a terminal it should never have been able to see. So, unless you set < isolate-terminal => False >, the pack does two things at both spawn sites (a foreground run and a background start):

• Standard input is detached. A run given no stdin argument gets a pipe that is closed immediately, so its fd 0 reads end of file rather than the host's terminal or transport. A run that is given stdin is unchanged. This half works on every platform.

• The child gets a new session (POSIX). A fresh session has no controlling terminal, so /dev/tty cannot be opened at all: the host's terminal is not merely unwritten-to, it is unreachable. Because a new session is also a new process group, this is what makes a kill reach the whole tree.

A new session has to be created by the child itself, between fork and exec, which is not somewhere Raku code can run — so the pack execs a tiny trampoline that calls setsid(2) and then execs the real command in the same process, preserving the pid every kill and exit status depends on. perl is used by preference (it is on macOS and every mainstream Linux, and the pack supplies the trampoline text itself, so a command that cannot be exec'd still reports exit: -1 with the reason on stderr exactly as it always did); setsid(1) is the fallback for images without perl. The trampoline adds one exec — a few milliseconds — per run, and it can never reach a shell: the exec is execvp(3), never /bin/sh.

The trampoline is chosen by the pack, never by the caller: no tool argument reaches it, and the command still has to pass the allow list before anything is spawned. It is found on PATH (falling back to the usual absolute locations), on the same trust assumption the allow list already makes for a bare command name — a process whose PATH has a writable directory ahead of /usr/bin has larger problems than this.

Which mechanism is available is resolved and verified once per process, by running a one-liner through the trampoline and checking that it came back a session leader with the pid that was spawned. A toolkit that asks for isolation and cannot get it refuses to construct, naming what it tried:

# Nothing on this machine can detach a child:
MCP::Server::Tool::Shell.new(allow => <git>);
# Shell toolkit could not isolate spawned commands from the controlling
# terminal, so it refuses to start: ... Install perl (or util-linux for
# setsid), or, if this process has no terminal worth protecting, construct
# the toolkit with isolate-terminal => False.

That is deliberate: a protection that is quietly missing is worse than one that is loudly absent, because the failure it prevents costs the user their session.

Kills follow the tree. Because an isolated child leads its own process group, a timeout, a cancelled call and the kill tool all signal the group: the grandchildren a command spawned die with it instead of outliving it and holding its pipes open (which used to lose the exit status, reported as exit: null). With < isolate-terminal => False > the old behaviour returns — on Unix, only the direct child is signalled.

Windows has no sessions, no controlling terminal and no termios to corrupt, so the session half of this simply does not apply there and the pack does not pretend otherwise: isolate-terminal still detaches standard input (which is the platform's own version of the hazard — a child reading the host's console, or its stdio transport), and kills already took the whole tree via taskkill /T /F. Console-handle isolation beyond standard input is not attempted.

Running anything: allow-any and MCP::Client::Policy

< allow-any => True > removes the allow list entirely. Everything else stays: no shell, argv vectors, NUL rejection, capped output, timeouts, jobs. What is gone is the decision about what may run — this pack will run anything the server's user can run, including a shell, including rm.

That is only a sane configuration if the decision is being made somewhere else. The intended pairing is a policy layer on the client side, such as MCP::Client::Policy, whose rules are name-based and whose default for an unmatched tool is to ask a human:

# Server: no allow list, because the client is where consent lives.
$server.plug: MCP::Server::Tool::Shell.new(:allow-any);

With a policy in front, sh_run and sh_start match no allow rule, so every call is put to the user before it happens — which is a better gate than a static list for an interactive coding agent, and a much worse one for an unattended server. If nothing is asking a human, use allow.

Sandboxing: the launcher seam

The pack is a gate, not a sandbox — but it leaves a seam for a host to bolt one on. The launcher attribute is a Callable (programmatic only: a closure cannot come from a JSON config, so from-config does not offer it) given the validated plan just before the process is built:

MCP::Server::Tool::Shell.new(
    :allow-any,
    launcher => -> %plan {
        # %plan is { command, argv, dir }.  Return { command, argv }: the
        # program actually exec'd becomes the wrapper, with the original
        # command and its arguments folded onto the wrapper's own argv.  dir
        # is left alone — the working directory is still applied by the pack.
        %(
            command   => 'sandbox-exec',
            argv      => ['-p', $profile, %plan<command>, |%plan<argv>],
            sandboxed => True,
        );
    },
);

The seam is deliberately minimal and total:

• It wraps both spawn sites. A foreground run and a background start pass through the same launcher and the same wrapping — there is no path that escapes it.

• An undefined launcher is exactly the old behaviour. No launcher, no wrapping, and the result payload is byte-for-byte what it always was.

• A launcher that throws is a failed spawn, not a crash. A launcher that dies (or returns a shape the pack cannot spawn) lands in the same exit: -1 + reason-in-stderr payload a command that could not be exec'd does. It never takes the server down, and never runs the unwrapped command as a fallback — a sandbox that failed to build is a run that failed, not a run that escaped.

• It marks its own runs. A launcher may return < sandboxed => True >, which surfaces as < sandbox => { active = true } >> on the result (a foreground run, or a job's output). A host reads it to tell a sandbox denial from an ordinary non-zero exit — the hook for an "it was the sandbox; retry without it?" prompt. The key is absent otherwise, so a run with no launcher, or one the launcher declined to sandbox, keeps the payload it always had.

The launcher does not see, and cannot change, the working directory, the timeout, the output caps or the standard input: it wraps what runs, and the pack still governs how long, how much and where.

Windows notes

• Bare names resolve the way CreateProcess resolves them, so < allow => <git> > runs git.exe off PATH. Allowlist the string you want the caller to pass (git), not the file that ends up being executed (git.exe) — the match happens before any resolution.

• Arguments are still a vector: Rakudo quotes each entry for CreateProcess, so an argument containing spaces stays one argument.

• A kill becomes TerminateProcess, which the command cannot trap or delay. There are no signals, so there is no 128 + signal exit status: a killed run or job reports whatever TerminateProcess left behind (1, in practice). Test for exit != 0 rather than for 137.

• A kill takes the whole process tree: taskkill /T /F /PID is issued against the child before the TerminateProcess backstop, so grandchildren the command spawned do not outlive the kill. It is fire-and-forget — the watcher collects the exit status independently — and if taskkill is missing from a stripped image the direct-child kill still runs. (Unix reaches the tree a different way, by signalling the process group of an isolated child; see #Terminal isolation.)

• isolate-terminal detaches standard input here as it does everywhere, but there is no session to leave and no /dev/tty to close off: Windows has neither, so the pack claims neither. See #Terminal isolation.

• Everything else — stdin, capped output, streaming, jobs — behaves the same as on Unix.

Examples

• examples/shell-server.raku — a git + ripgrep server, with the invocation line in the header.

See Also

• MCP::Server — the server framework, the MCP::Server::Toolkit role this pack implements, and the raku-mcp command.

Author

Matt Doughty

License

Artistic-2.0