whats-allowed-mcp
{
"answer": "The whats-allowed-mcp server audits and reports on the merged permission configuration in effect for a directory — it analyzes and explains, but never edits or simulates decisions.\n\n- whats_allowed — One-call summary of all permissions in force: which settings files contribute, rule counts by type (allow/ask/deny), the winning defaultMode, blanket allows, hooks wired to tool use, and how many rules don't do what they appear to.\n- permission_sources — Lists every settings file that can contribute rules (managed policy, project, local, user, session) in precedence order, reporting existence, parse status, rule counts, and how a leading / anchors path rules in each.\n- rule_findings — Surfaces rules whose documented behavior differs from apparent intent: inert rules (accepted but never consulted, e.g. Write, NotebookEdit, Glob), /path rules in user settings anchoring to the config directory instead of the project, wildcards without word boundaries, allow rules on command-runners (npx, docker exec, bash -c), and allow rules shadowed by a deny or ask rule that fires first. Filter by kind: inert, misreads, wider, shadowed, duplicate.\n- unattended_surface — Shows what lets a tool call proceed without human approval: the winning defaultMode, blanket allows, extra directories granted, whether bypass/auto modes are disabled, and hook commands wired to agent events.\n\nScope and limitations:\n- Read-only — only node:fs, node:path, node:os; no child processes, network, or writes. Nothing in a settings file is executed.\n- Never reads env values (only names) and never reads or runs apiKeyHelper scripts, avoiding plaintext secret leaks.\n- Does not simulate client decisions (no "would this command be allowed" verdict) — use /permissions in your client for the live view.\n- Command-line flags (--allowedTools, --permission-mode, etc.) are invisible since they're not on disk.\n- Every finding is conservative and links to documented Anthropic behavior — no risk scores or intent judgements.\n- Each tool takes an optional dir argument (absolute path) defaulting to the server's working directory."
}
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@whats-allowed-mcp What permission rules are being silently ignored?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
whats-allowed-mcp
You wrote a deny rule. Did it do anything? An MCP server that reads every settings file feeding your agent's permissions, shows which one wins, and lists the rules your client accepts and then ignores.
Why
Permission rules for a coding agent are not in one file. They are in up to five at once: a machine-wide managed policy, the project file your team commits, a git-ignored local file, your user file, and whatever a session wrote. They are merged, and then evaluated deny → ask → allow, where the first match wins and specificity does not break the tie.
That merge is hard enough. The part that actually bites is that several rule shapes are accepted and then quietly do nothing:
Write(docs/**)— file permissions are checked againstEdit(path)andRead(path)rules only. AWrite,NotebookEdit,MultiEditorGlobpath rule is accepted and never consulted.Bash(command:rm *)— a rule can't match a tool's primary content field with theparam:valueform. It is ignored, because a compound command would bypass it."mcp__*"inallow— an unanchored allow glob is skipped and auto-approves nothing.Read(/.env)in user settings — a single leading slash anchors at the settings source, so that rule protects~/.claude/.env, not the.envin each of your projects.Bash(git push origin main)inallowunder aBash(git push *)indeny— a deny rule cannot carry allowlist exceptions, so the allow rule is unreachable.
Claude Code warns about some of these at startup, in a scrollback you have already lost by the time you care. Nothing collects them, nothing tells the agent itself, and nothing shows you the merged picture across all five files at once. This does.
Related MCP server: mcp-permaudit
What it looks like
Given a project .claude/settings.json:
{
"permissions": {
"deny": ["Bash(git push *)", "Write(.env*)"],
"allow": [
"Bash(npm run test:*)",
"Bash(git push origin main)",
"Bash(npx *)",
"Bash(ls*)"
]
}
}…and a user ~/.claude/settings.json with "deny": ["Read(/.env)"] and "defaultMode": "acceptEdits", whats_allowed returns (paths shortened here):
# Permissions in force
| | |
|---|---|
| Settings files contributing | 2 of 5 possible |
| Permission rules | 3 deny, 0 ask, 4 allow |
| Blanket allow rules (whole tool) | 0 |
| Hook commands wired to agent events | 0 |
| Extra directories granted | 0 |
| Rules that do not do what they look like | 5, of which 1 is ignored outright |
## Default mode
`defaultMode` = **`acceptEdits`**, set in User (`~/.claude/settings.json`).
> Automatically accepts file edits and common filesystem commands (`mkdir`,
> `touch`, `mv`, `cp`) for paths in the working directory or additionalDirectories.
## First things to look at
- **Path rule on Write** — `Write(.env*)` in Project settings
- **Allow rule on a command runner (`npx`)** — `Bash(npx *)` in Project settings
- **Trailing wildcard with no word boundary** — `Bash(ls*)` in Project settingsSeven rules. Five of them do something other than what they read as — and the one everybody would have bet on, Write(.env*), is never consulted at all.
Tools
Tool | What it answers |
| The headline: which files contribute, how many deny/ask/allow rules, the winning |
| Which file decides, in precedence order — existence, parse state, rule counts, and what a leading |
| Every rule whose documented behaviour differs from its apparent intent, with the reason and the documented alternative |
| What proceeds with nobody watching: default mode, blanket allows, extra directories, mode guards, and hook commands |
Install
Claude Desktop (one-click, no terminal): download the latest whats-allowed-mcp-<version>.dxt from Releases and open it with Claude Desktop (double-click, or Settings → Extensions → Install Extension…). The server and its dependencies ship inside the bundle — no npm, no Node install.
Register with Claude Code (available in every session):
claude mcp add --scope user whats-allowed -- npx -y whats-allowed-mcpOr in any MCP client config:
{
"mcpServers": {
"whats-allowed": {
"command": "npx",
"args": ["-y", "whats-allowed-mcp"]
}
}
}git clone https://github.com/stcmain/whats-allowed-mcp
cd whats-allowed-mcp && npm install && npm run build
# then point your client at node /path/to/whats-allowed-mcp/dist/index.jsPublished as whats-allowed-mcp on npm and as io.github.stcmain/whats-allowed-mcp in the MCP Registry.
No configuration. No environment variables. Respects CLAUDE_CONFIG_DIR when you have moved your user config.
What it checks, and what it refuses to guess
Every finding corresponds to behaviour Anthropic documents, and every one links to the paragraph it comes from. There is no risk score, no severity ranking and no "suspicious rule" heuristic — those produce confident nonsense on ordinary configurations.
Reported:
Path rules on
Write/NotebookEdit/MultiEdit/Glob, which file permission checks never consultTool(param:value)rules aimed at a tool's primary content field, which are ignored outrightUnanchored tool-name globs in
allow, which auto-approve nothing:*used anywhere but the end of a shell pattern, where the colon is literal/pathrules in user settings, which anchor at the config directory rather than at your projectTrailing
*with no word boundary —Bash(ls*)also matcheslsofAllow rules on commands that run other commands (
npx,docker exec,bash -c,xargs,devbox run…), which approve whatever follows themBash rules that try to constrain a URL, which the docs themselves call fragile
Allow rules a deny or ask rule reaches first
Rules declared in more than one file
Deliberately not reported:
Whether your rules express what you want. That is a judgement about intent and this server does not have one.
Shadowing that needs a wildcard solver. An allow rule is only called unreachable when the blocking rule is a prefix pattern that provably covers it. Anything subtler is left alone rather than guessed at, so this under-reports on purpose.
A verdict for a hypothetical command. See below.
Honest limitations
It does not simulate your client's decision. There is no "would
git pushbe allowed" tool, and that is a design choice: Claude Code's built-in read-only command set, wrapper stripping (timeout,nice,xargs…), compound-command splitting, sandbox state and PreToolUse hook results all participate in the real answer. A static verdict would be confidently wrong often enough to be worth less than no answer. Use/permissionsin your client for the live view; use this to understand why it says what it says.Command-line flags are invisible.
--allowedTools,--disallowedTools,--permission-modeand--settingssit between managed policy and local settings, and they are not on disk.Session rules added through
/permissionsland in files this server reads, but rules added for one session only do not.It reports; it does not fix. Nothing is edited, and no finding is ever a recommendation to delete a rule. Where the docs give a corrected form, it is quoted; where they don't, the finding stops at the observation.
~/.claude/settings.local.jsonis reported but not ranked. It exists and Claude Code writes to it; the published precedence table lists four scopes and does not include it. This server refuses to invent its position.Claude Code's schema is the model. Other clients with their own permission systems are not parsed.
It never reads git history, and it does not tell you who added a rule or when.
git log -p .claude/settings.jsondoes that better.
Design notes / threat model
No child processes. No shell. No network. No writes. The only Node APIs used are
node:fsreads,node:pathandnode:os. There is nochild_processimport anywhere in the source, so nothing in a settings file can be executed by reading it — including the hook commands andapiKeyHelperscripts it reports on.envvalues are never read — only names. Settings files are one of the most common places for a plaintext API key, and a tool that pasted one into a context window would be worse than the problem it solves. The same applies toapiKeyHelper: presence is reported, the script is neither read nor run.Settings-authored strings are fenced and labelled. Rule text and hook commands have to be shown to be useful. They are emitted inside inline code spans with backticks neutralised, pipes escaped and newlines flattened, so a crafted string cannot break out of the span or out of a markdown table, and each block carries a standing note that the quoted text is data from a file, not instructions.
diris the one model-controlled path, and it is bounded by construction: resolved, real-pathed, and required to be an existing directory. From there only fixed filenames are read —.claude/settings.json,.claude/settings.local.json, the user config file, the platform's managed-policy path. No globbing, no traversal, no arbitrary file reads.Bounded work: a 2 MB ceiling per file, a capped walk when looking for the project root, and files that fail to parse are reported as unparsed rather than partially guessed at.
Findings are conservative by construction. Every check is anchored to documented behaviour; anything that would need a judgement call about intent is not reported at all. For a tool people use to decide whether a guard rail is real, under-reporting is the correct failure mode.
Who makes this
Built by Shift The Culture — we run a one-person company on AI agents and ship the tooling we needed ourselves. This server is free and MIT-licensed, no strings.
It has three siblings, all also free and MIT:
whats-running-mcp — what is actually running on the box right now, instead of what an old transcript claims.
whats-loaded-mcp — what is eating your context window before you type: skill descriptions, memory files and their imports.
whats-inherited-mcp — what a checkout you did not write tells your agent to do: instruction files, hooks, and the MCP servers it declares.
The rest of that tooling is paid:
Agent Fleet Ops Kit ($29) — the other failure modes of running three or four agents on one box: two sessions editing the same checkout, a dev server nobody owns (so the agent tests a different app than it edits), and MCP servers leaked from crashed sessions that hold ports and RAM for weeks. Prefer PayPal? Same kit on Payhip.
Agent Reliability Kit ($29) — a Stop hook and two CLIs that block a turn when an agent claims "done" against a repo, URL, or build that was never actually checked. Prefer PayPal? Same kit on Payhip.
The server above stays free and MIT either way — it has no upsell in it, no telemetry, and no dependency on the paid kits.
Sponsors
This server is MIT and stays MIT. There is no pro edition, no telemetry, and nothing held back from the free build. Sponsorship is how the maintenance gets paid for without any of that changing.
No sponsors yet — the first slot is open. Company sponsors get their name or logo in this section, in the three sibling servers, and on the sponsor page. Tiers, exactly what the placement is, and what it explicitly does not buy: https://shifttheculture.media/sponsor
Individuals: https://paypal.me/ShiftTheCultureLLC — any amount, no perks, no tier.
License
MIT © Zachary Pampu
Available Tools
4 toolspermission_sourcesWhich files decide, in what orderA
Every settings file that can contribute permission rules, in documented precedence order, with whether it exists, whether it parsed, how many rules it carries, and what a leading / in its path rules anchors to. Use when a rule is not taking effect, when you cannot tell which file granted something, or to confirm a managed policy is or is not present.
Inputs: dir.
rule_findingsRules that do not do what they look likeA
Permission rules whose documented behaviour differs from their apparent intent: rules Claude Code accepts and never consults, /path rules in user settings that anchor at the config directory rather than your project, wildcards without a word boundary, allow rules on commands that run other commands, and allow rules a deny or ask rule reaches first. Each finding cites the documented behaviour. Use before trusting a guard rail you wrote a while ago.
Inputs: dir, kind.
unattended_surfaceWhat proceeds with nobody watchingA
The parts of the configuration that let a tool call go through without a human: the winning defaultMode, blanket allow rules, extra directories granted, whether bypass and auto modes are disabled, and hook commands — which run on agent events without a permission prompt of their own. Use before leaving an agent running unattended, or when reviewing what a repo's settings would do on your machine.
Inputs: dir.
whats_allowedWhat this agent can do without askingA
Start here. One-call summary of the permission configuration in force for a directory: which settings files contribute, how many allow/ask/deny rules each carries, the winning defaultMode, blanket allows, hooks wired to tool use, and how many rules do not do what they look like they do. Use at session start, before an unattended run, or when a prompt appeared that you did not expect.
Inputs: dir.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool addresses a distinct aspect of permission configuration: overall summary, file sources, behavioral anomalies, and unattended execution pathways. There is no functional overlap between the four tools; their purposes are clearly separated and easy to distinguish.
All four tool names follow the same snake_case convention and are descriptive noun phrases that clearly reflect their output (e.g., permission_sources, rule_findings). The naming is consistent across the entire set, with no mixed conventions or ambiguous verbs.
With only four tools, the server is tightly focused on its purpose of analyzing permission configuration. Each tool covers a necessary dimension of the domain, and the count is well within the ideal range for a specialized MCP server.
The tool set provides comprehensive coverage for inspecting and understanding permission configuration: a starting summary, source file inventory, anomaly detection, and unattended execution risk. For its diagnostic scope, there are no obvious missing operations or dead ends.
Maintenance
Related MCP Connectors
Lets coding agents check their own code for leaked secrets, risky dependencies and AI-code mistakes
Deterministic allow/require_approval/deny verdicts for agent actions, before they happen.
Find your AI agent's likely failure mode, get runtime settings, and clarify ambiguous prompts.
Permission boundary receipts for ChatGPT agents.
Related MCP Servers
- AlicenseAqualityDmaintenanceOpen-source permission control plane for AI agents — scan, enforce, and audit every tool call with code-level policies that prompt injection can't bypass.14103 PyPI19Business Source 1.1
- FlicenseNot gradedqualityDmaintenanceAudits Claude Code permission settings (.claude/settings.json) via an MCP tool, detecting conflicts and misconfigurations.-
- AlicenseAqualityDmaintenanceA blast-radius auditor for AI agents that computes combined effective permissions across all connected MCP tools, detects dangerous capability combinations deterministically, and provides a live risk graph and one-click policy fixes.4MIT
- AlicenseNot gradedqualityBmaintenanceEnables analysis of historical agent tool-call audit logs to detect over-permissioned wildcards, synthesize strict least-privilege JSON security boundary policies, and quantify attack surface and blast-radius reduction.8MIT