Name: Guardian MCP
Author: andrebyrd-odingard

99/100

trust score

199

downloads a day

local

runs on your machine

downloads in 30d

What it does

Guardian is the human authority layer for AI agents. Put it in front of any MCP server and every tool call is decided before it runs:

ALLOW runs it.
ASK holds the exact call until a person approves it on their phone or in the console.
DENY never runs it.

The decision is deterministic, with no model in the decision path, and every outcome carries a signed receipt you can verify offline.

Install: npx -y @odingard/guardian-mcp proxy -- <your MCP server command>

What's new in 0.3.2:

guardian-mcp tools -- <server> lists any MCP server's tools with a suggested mode for each. It uses the server's own read-only and destructive annotations first. It also flags tool names Guardian would reject. No account needed.
Receipt verification accepts the key-fetch result as returned.
No automatic first-use registration by default (since 0.3.1). Register tools once in the console instead.

Also for Claude Code: Guardian Hook (@odingard/guardian-hook)

A Claude Code PreToolUse hook that sends shell commands, file writes, web fetches and MCP calls to Guardian before they run. Commands are routed by consequence:

npm test runs under Allow.
terraform apply, kubectl apply, rm -rf and git push wait for a human approval.

The hook fails closed: if Guardian can't confirm a call, the call is blocked. Every allowed call carries a signed receipt.

Proven on production: Claude Code's terraform apply was held, approved on a phone, executed, and its receipt verified offline. A read-only ls in the same session ran with no approval.

Honest limits:

The hook is Claude Code only for now.
Decision-gated: the receipt proves Guardian authorized the exact call, not the command's own outcome.

Links:

npm: https://www.npmjs.com/package/@odingard/guardian-mcp
Source: https://github.com/Odingard/guardian-sdk/tree/main/packages/mcp
MCP registry: io.github.Odingard/guardian-mcp
Hook on npm: https://www.npmjs.com/package/@odingard/guardian-hook
Hook source: https://github.com/Odingard/guardian-sdk/tree/main/packages/hook

Quickstart

# 1 — add it to Claude Code
claude mcp add --scope user --env 'GUARDIAN_BASE_URL=${GUARDIAN_BASE_URL}' --env 'GUARDIAN_AGENT_CREDENTIAL=${GUARDIAN_AGENT_CREDENTIAL}' --env 'GUARDIAN_CONNECTION_ID=${GUARDIAN_CONNECTION_ID}' --transport stdio odingard-guardian-mcp '--' npx -y @odingard/guardian-mcp

# 2 — ask your agent for something these tools do
#   guardian_invoke_tool, guardian_claim, guardian_receipt

It needs GUARDIAN_BASE_URL, GUARDIAN_AGENT_CREDENTIAL, GUARDIAN_CONNECTION_ID: this reads them from your environment, so export them first.

Guardian MCP costs nothing on mcprush: there is no plan to choose here, no cap to set and nothing mcprush can bill you for.

Publisher

andrebyrd-odingard1 server

1 category

All 1 server

Installation

Where are you running it?

Every route below installs the same thing and ends at the same approval screen. This one runs on your machine: your client starts Guardian MCP as a process under your own user, with your files and your network, so the tool surface below is what it can do to your computer rather than to a server somewhere else. Its scan result is on the Security tab — read the surface before you approve it.

Claude Codedeclared by this listing

This is a public server: you run it yourself and this marketplace is not in the path. You need Node.js — the current LTS (24) is the safe choice; npx comes with it. Claude Code registers it in one command; --scope user makes it available in every project. Claude Code reads GUARDIAN_BASE_URL (Guardian control-plane base URL), GUARDIAN_AGENT_CREDENTIAL (Agent credential (gac1.…)), GUARDIAN_CONNECTION_ID (Default tool connection id for invocations) from your environment each time it starts the server, so the value stays out of its config: export them in your shell profile (macOS, Linux), or run setx GUARDIAN_BASE_URL …, setx GUARDIAN_AGENT_CREDENTIAL …, setx GUARDIAN_CONNECTION_ID … and open a new terminal (Windows). The '--' is in quotes so that PowerShell passes it on.

claude mcp add --scope user --env 'GUARDIAN_BASE_URL=${GUARDIAN_BASE_URL}' --env 'GUARDIAN_AGENT_CREDENTIAL=${GUARDIAN_AGENT_CREDENTIAL}' --env 'GUARDIAN_CONNECTION_ID=${GUARDIAN_CONNECTION_ID}' --transport stdio odingard-guardian-mcp '--' npx -y @odingard/guardian-mcp

Reconnect, or start a new session, and the tools appear in the model’s tool list.

Free on mcprush, and it runs from its publisher’s own package: this marketplace is not in the path, so there is no key of ours to hold and no card to add.

What it asks of you

It reaches a system of yours, so it needs your own credential rather than the publisher’s: GUARDIAN_BASE_URL — Guardian control-plane base URL; GUARDIAN_AGENT_CREDENTIAL — Agent credential (gac1.…); GUARDIAN_CONNECTION_ID — Default tool connection id for invocations. You set them in your own client’s config, on your machine — this marketplace never holds it.

What it writes

One config entry your client uses to start the process locally. A local server runs with your file system and your network, which is why it is priced without metering.

What you approve

3 tools, with what each one reads, writes and reaches shown before you agree — the same list on every route above. Read the tool surface.

Tool surface

What the model actually sees. Descriptions are diffed on every release — see version history.

Only tools that write or reach the network

Read from its source code. A server that builds its tool list when it starts can list more tools than this.

Propose a tool invocation to Guardian. ALLOW executes immediately and returns the execution receipt; ASK is held for human approval (call guardian_claim with the returned requestId after the human approves); DENY refuses the action with a reason code.

Parameter	Type	Required	Notes
toolName	string	yes	Registered tool name.
arguments	object	no	Arguments the tool would run with.
connectionId	string	no	Tool connection id. Defaults to GUARDIAN_CONNECTION_ID.
purpose	string	no	Human-readable reason for this action.

Collect the capability for a request that was held as ASK and approved by the human, then execute it and return the receipt.

Parameter	Type	Required	Notes
requestId	string	yes	—

Fetch the machine-readable evidence receipts for a request id.

Parameter	Type	Required	Notes
requestId	string	yes	—

Free

$0 forever

Every tool, no call limit
No card, and mcprush charges nothing for it
Source published under Apache-2.0
Runs on your machine — nothing of it reaches our gateway
Nothing to cap, because mcprush bills nothing

Nothing here is billable: mcprush charges nothing to install Guardian MCP or to call it, at any volume.

ATrust score 99 / 100

Release 0.3.2 · scanned 5 Nov 2026

Two independent axes, because powerful and malicious are different questions. The grade is threat only. The capability level is blast radius, and it is never a penalty on the grade — it is priced as one subtract-only term in the score, where you can see it.

Threat score100

Client score99

Findings0 threat · 0 capability

Capabilityminimal

Coveragesource

Verificationrepo

Surface digest015cac60e4c53ffaf8f92e43ceb730a45e172d04d8faf1e5dd83c89860398f22

Score vectorthreat 100 → client 99

Term	Level	What it prices	Points
capability-exposure	minimal	capability blast radius (minimal) — client exposure if the model is manipulated	−0
verification-discount	repo	publisher verification (public source) — no provenance, but the source is public and inspectable	−1
coverage-honesty	source	inspection depth (source) — how much of the target the scan could see	−0

score = clamp(round(threatScore − E_cap − E_ver − E_cov), 0, 100). No model, no opinion, no network call at scoring time: the same methodology version against the same bytes always produces this number, and every line above is a field of the report rather than a summary of it.

What the scan could actually read

A grade is only as meaningful as its coverage, so the scanner publishes its own depth before it publishes its result.

Tool surfaceread

Implementation sourceread

Package metadataread

Live transportnot available

Tools were statically extracted from the published source (3 recovered), not enumerated from a running server. Tool-poisoning, Unicode-smuggling, capability and toxic-flow analysis ran on this inferred surface, but a mis-parsed registration could be missed or mis-attributed, so tool-derived findings are capped below “confirmed”. To grade the real runtime surface, scan the running server: --command "npx -y <package>".

Capability — what it could do if the model were manipulated

Tags derived from each tool’s schema and the implementation, not from what the tool calls itself. minimal is the level these add up to.

Tool	Capability tags	Why the tag was assigned
guardian_invoke_tool	no tags
guardian_claim	no tags
guardian_receipt	untrusted-input	untrusted-input: keyword "fetch"

Toxic-flow graph

The lethal trifecta, checked as a graph rather than as a checklist: untrusted input, a sensitive source and an external sink have to meet before there is a path worth worrying about.

The public result for this release does not print the flow graph, so there is nothing to show here. That is not the same as "no paths were found": what the scan did read is above, under coverage.

Supply chain and provenance

Integrityfirst-seen

Surface digest015cac60e4c53ffa

Scanned2026-11-05

This is the first scan of this surface here, so there is nothing yet to compare it against.

Every result on this tab comes from one deterministic pass over the published package — offline, rule by rule, and auditable line by line above. Same methodology version, same bytes, same score.

Guardian MCP