sweep-mcp

An MCP server that deletes directories — and the guard rails that make handing that capability to a language model something other than reckless.

It exposes sweep over the Model Context Protocol, so an agent can find and reclaim node_modules, .venv, target, __pycache__ and friends. The interesting part is not the deleting. It is everything that has to be true before a delete is allowed to happen.

Research published in 2026 found security issues in 66% of 1,808 scanned MCP servers, with 30+ CVEs filed against MCP implementations in a single quarter. This repository is one server written the other way round: the refusals first, the feature second.

---

The threat model

An MCP server that can delete directories is a loaded gun pointed at whatever the process can reach. Five things could go wrong, and each has a control and a test that makes it fire.

Risk	Control	Test
Agent scans / or C:\ and deletes the machine	Root allowlist, fixed at server start. The agent cannot set, extend or read past it. No roots configured → every request refused	test_empty_allowlist_denies_every_path
../../ or a symlink used to escape the allowlist	Every path is resolve()d before the containment check, so it is judged on where it really points	test_dotdot_traversal_is_denied, test_symlink_pointing_outside_root_is_denied
Agent asks to delete a path it never scanned	reclaim accepts ids, not paths. There is no path-shaped route to deletion anywhere in the API	test_there_is_no_way_to_delete_by_path
The directory changes between scan and delete	Every find is re-checked against the live filesystem at deletion time — still contained, still a directory, still not a link, marker file still present	test_revalidate_refuses_a_directory_swapped_for_a_symlink
Accidental destruction	Dry run is the default. Deletion requires confirm="delete" exactly; anything else is treated as no	test_a_wrong_confirmation_string_stays_a_dry_run

Two smaller ones, for the same reason:

• Ticket ids are random, not sequential. f-3a91c02b77de, not 1. A counting id invites an agent to iterate until something deletes.
• Tool descriptions are string literals. They are never assembled from anything read off disk, so a directory named ignore-previous-instructions reaches the model as data rather than as a sentence (test_tool_descriptions_are_static).

What it does not protect against

The controls above are worth exactly what they cover, and no more.

• A misconfigured root. Start it with --root / and it will happily work through your filesystem. The allowlist is only as good as what you put in it, and nothing here second-guesses that choice.
• A malicious client. The guard constrains what may be deleted, never who is asking. A client that scans legitimately and then confirms legitimately gets exactly what it asked for.
• An agent that has been talked into it. Static tool descriptions keep filesystem content out of the model's instructions, but nothing here can stop an agent that decides, for its own bad reasons, to delete a real node_modules it genuinely found.
• The last few milliseconds. revalidate() narrows the window between check and delete; it does not close it. Closing it properly needs directory file descriptors held across the operation, which shutil.rmtree does not offer portably. This is a real race, and it is smaller, not absent.
• Windows junctions on Python 3.11. os.path.isjunction arrived in 3.12. Below that, junction detection degrades to symlinks only — the same compromise sweep itself makes, stated here rather than buried.
• Privilege. This is not a sandbox. It runs as whoever started it.

Scanner results

Run against agent-audit 0.19.2 on 18 August 2026.

Static scan — 0 findings, risk score 0.0 (LOW), 7 files.

One low finding appeared on the first pass: AGENT-110, source-map artifacts not excluded from distribution. This package ships no JavaScript, so there were no source maps to leak, but the exclusion is now declared in pyproject.toml anyway — it costs nothing, and arguing with a supply-chain checklist is a poor use of anyone's afternoon.

Tool inspection — 0 tool findings. No tool-poisoning, cross-origin or rug-pull patterns detected in any of the three definitions.

The risk ratings it assigned are worth reproducing, because they are wrong in an instructive way:

Tool	Rated	Inferred permissions
list_targets	MEDIUM	SHELL_EXEC, FILE_READ
scan	HIGH	FILE_DELETE, SHELL_EXEC, FILE_READ
reclaim	HIGH	FILE_DELETE, SHELL_EXEC, FILE_READ

scan cannot delete anything. It is read-only, declared read_only_hint=True, and there is a test asserting it stays that way. It was rated HIGH with a FILE_DELETE permission because the scanner matches keywords against the description, and this server's scan description contains the sentence:

"Read-only: this never deletes anything."

The word deletes is on the FILE_DELETE keyword list. SHELL_EXEC comes from the word command, in "the command that regenerates it" — this server never executes a shell command at all.

The description could be reworded to score better. It has not been, because the description's audience is a language model deciding whether to call the tool, and "this never deletes anything" is the single most useful sentence in it. A keyword scanner is a smoke detector, not a judge — a clean run is worth having and worth publishing, and a rating derived from string matching on prose is not evidence about behaviour in either direction.

Two scanners were tried and one could not be run: Invariant Labs' mcp-scan has been renamed snyk-agent-scan following the Snyk acquisition and now requires a SNYK_TOKEN. Cisco's mcp-scanner is not published to PyPI. agent-audit's own stdio transport also times out against this server on Windows while a hand-driven handshake to the identical binary returns in milliseconds, so its analyser was fed the real tools/list output directly — the transport was bypassed, the analysis was not.

Install

pip install -e .

Configure

The server takes one or more --root directories. It has no default. Starting it without a root is an error rather than an invitation to scan everything:

{
  "mcpServers": {
    "sweep": {
      "command": "sweep-mcp",
      "args": ["--root", "/home/you/code", "--root", "/home/you/work"]
    }
  }
}

The tools

Tool	Destructive	What it does
list_targets	no	The kinds of directory it knows how to reclaim, and the command that regenerates each
scan	no	Finds reclaimable directories under a path inside the roots. Returns ids, sizes, and the regenerate command
reclaim	yes	Deletes finds by id. Dry run unless confirm="delete"

All three carry MCP ToolAnnotations, so a client that gates destructive tools behind a confirmation prompt has what it needs: reclaim declares destructive_hint=True, the other two declare read_only_hint=True. That is asserted in the suite too — a hint that silently regressed would be worse than none.

A session

scan(path="/home/you/code")
  → 12 finds, 3.4 GB
    f-3a91c02b77de  /home/you/code/api/node_modules   1.9 GB  npm install
    f-8e0244fd1b6c  /home/you/code/ml/.venv           842 MB  python -m venv .venv

reclaim(ids=["f-3a91c02b77de"])
  → dry_run: true
    would_delete: 1 path, 1.9 GB
    note: Nothing was deleted. Call again with confirm='delete'.

reclaim(ids=["f-3a91c02b77de"], confirm="delete")
  → deleted: 1 path, reclaimed 1.9 GB

Tests

38 tests. 34 run on Windows; all 38 run on Linux.

The four that skip locally need to create symlinks, which Windows refuses without developer mode — and one of them covers the swap-between-scan-and-delete attack, which is the single most important test here. Skipping it quietly would make the suite a decoration, so CI runs on Linux and fails the build if those tests report as skipped there.

pytest -q --cov=sweep_mcp

Coverage sits at 83%. The gap is mostly main() and the argparse wiring, which the suite drives through build_server instead — the transport is the part least worth mocking and least likely to be where the damage comes from.

Nothing in the suite mocks the filesystem. A mocked Path would pass every test in here while the server still deleted the wrong directory.

Layout

src/sweep_mcp/
  guard.py    212 lines - the allowlist, the tickets, the re-check. Imports no MCP.
  server.py   280 lines - three tools. Translation only.
tests/
  test_guard.py    23 tests - one per rule, each named for the attack it stands in for
  test_server.py   15 tests - driven through call_tool, the way a client would

guard.py deliberately knows nothing about MCP. Every decision that could lose someone their data is testable without a client, a transport or an agent in the loop. If a rule looks like it is enforced in server.py, that is a bug — it belongs one layer down, where the tests can reach it.

Licence

MIT.