# @agentguard-run/mcp

Local spend caps, kill-switch blocking and Ed25519-signed receipts for AI agent tool calls.

- **Type:** MCP server
- **Trust:** 60/100 (B), scored on the package rubric
- **Verification:** community-indexed — nobody has claimed this listing
- **Version:** 0.3.1
- **Author:** io.github.MerchantGuard
- **License:** SEE LICENSE IN LICENSE
- **npm:** @agentguard-run/mcp
- **Source:** https://github.com/MerchantGuard/agentguard-mcp
- **Compatible clients:** claude-code, cursor, copilot, gemini (basis: transport)

## Trust

60/100 (B), scored on the package rubric
- Publisher verified: no
- Install scripts: nothing suspicious found
- Prompt-injection scan: not run
- Obfuscation scan: not run
- Evidence age: 1 day

## Security scan

- **Status:** clean
- **Scanned:** 2026-09-22T14:49:26.519Z
- **Version scanned:** 0.3.1
- **CVEs:** none found by OSV at scan time
**Findings**
- injection-shaped content (note) in the `spend_decide` tool: Imperative addressed to the AI model

## Tools

6 declared. Statically extracted from the shipped source — a floor on the surface, not a census.
- `spend_decide` — Gate a model call BEFORE it runs. Pass metadata only (model, token counts, endpoint URL), never prompt content. Returns allow/block plus an Ed25519-signed, cont
- `verify_receipt` — Independently verify an AgentGuard signed receipt. Returns valid true/false. Tampering with any signed field makes verification fail.
- `export_receipts` — Export the tenant decision ledger, including earlier process runs, as a hash-chained array of signed receipts with a chain verification result. Memory mode expo
- `set_model_cost` — Register a cost for a model with no built-in pricing (required for self-hosted models: AgentGuard fails closed on unknown costs). Cents per 1,000 tokens.
- `provenance_preview` — Preview the provenance AgentGuard would attest for a model and endpoint WITHOUT spending: model identity, weights origin country (China-origin families flagged)
- `spend_status` — Show the active policy: caps, tenant, enforcement mode, decisions so far this session, and the signer identity.

## Install

**Verdict: review** — Installable, but 1 thing to check first: No publisher has proved control of this listing; it is indexed, not vouched for.
**Cautions** (coverage gaps and advisories — never blocking)
- No publisher has proved control of this listing; it is indexed, not vouched for.
**Config** (claude-code):
```json
"{\n  \"mcpServers\": {\n    \"mcp\": {\n      \"command\": \"npx\",\n      \"args\": [\n        \"-y\",\n        \"@agentguard-run/mcp\"\n      ],\n      \"env\": {\n        \"AGENTGUARD_LICENSE_KEY\": \"<YOUR_AGENTGUARD_LICENSE_KEY>\"\n      }\n    }\n  }\n}"
```
**Credentials it will ask for** (names only — Forge never holds a value):
- `AGENTGUARD_LICENSE_KEY` — Agentguard License Key (optional)
Placeholders only. Forge never holds, brokers, or transmits a credential value — replace each <YOUR_NAME> in your own config file. Do not send a value back to Forge; no Forge endpoint accepts one.
- This entry needs 1 credential (0 required). The generated config carries placeholders, so it will fail in the editor rather than at runtime if they are left unset.

## Blast radius

Moderate blast radius — holds an api key; runs on your machine.
- Floor 23, ceiling 23 (tier: moderate)
- This is impact, not likelihood. A high radius is not a defect: a filesystem server is supposed to write files. It is never part of the trust score.

## Machine-readable views of this entry

- Signed JSON: https://forgeregistry.com/api/v1/packages/%40agentguard-run%2Fmcp
- Install plan: https://forgeregistry.com/api/v1/packages/%40agentguard-run%2Fmcp/install-plan
- Alternatives: https://forgeregistry.com/api/v1/alternatives/%40agentguard-run%2Fmcp
- HTML page: https://forgeregistry.com/registry/%40agentguard-run%2Fmcp
- MCP: POST https://forgeregistry.com/api/mcp → `forge_get_package` / `forge_install_plan`

## About this document

Generated by Forge (https://forgeregistry.com) — a compact rendering of the same record served, signed, at the JSON URL above. Trust and scan facts are the registry's own measurements; anything Forge did not measure is named as unmeasured rather than omitted.