Back to search
getvda
MarketplaceWorks with Claude CodeGitHub
getvda.ai plugins for safer agents: MCP Drift Guard and the MCP Drift Observatory.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/getvda-ai--mcp-drift-guard)HTML
<a href="/listings/getvda-ai--mcp-drift-guard"><img src="/api/listings/getvda-ai--mcp-drift-guard/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
npx mcp-drift-guard lock # probe your servers, write mcp-tools.lock.json, then commit it npx mcp-drift-guard check # exit 1 if anything changed npx mcp-drift-guard update # accept the change after you've reviewed it (then check the git diff) npx mcp-drift-guard observe https://mcp.example.com/mcp # any public server, no config needed
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 20 lines
- {
- "name": "getvda",
- "owner": {
- "name": "getvda.ai",
- "url": "https://getvda.ai"
- },
- "plugins": [
- {
- "name": "mcp-drift-guard",
- "source": "./plugin",
- "description": "Lockfile for MCP tool definitions: warns when a server's tools change and flags tool-poisoning signals."
- },
- {
- "name": "mcp-drift-observatory",
- "source": "./directory",
- "description": "Read-only: ask Claude whether a public MCP server changed its declared tools, and see what changed."
- }
- ],
- "description": "getvda.ai plugins for safer agents: MCP Drift Guard and the MCP Drift Observatory."
- }
README.md · README · 127 lines · 1 finding marked
- # MCP Drift Guard
- **A lockfile for the tools your MCP servers declare.** It fails CI, or warns you when a Claude Code session starts, as soon
- as a server you depend on changes its tool names, descriptions or schemas.
- Tool definitions go straight into your agent's prompt. A server that rewrites a description (a "rug pull") changes
- what your agent does, and nothing in your repo changes. `npm` has `package-lock.json`. Your MCP tools now have
- `mcp-tools.lock.json`.
- ```text
- ⚠️ local stdio 3 tools drifted
- - added send_notes 🚩 exfil_wording
- - changed get_weather: description 🚩 instruction_tags, conceal_from_user, sensitive_paths
- ```
- - Works with **every** server you can reach: remote (Streamable HTTP, legacy SSE) and local (stdio), including
- servers behind auth. Your secrets stay in your CI.
- - **And the public ones you can't.** When this run can't read a remote server (no secret in this job, blocked from the
- runner, down right now), it falls back to the MCP Drift Observatory's last nightly observation and compares hashes.
- You can even pin a server you have never connected to. See [Observatory fallback](#observatory-fallback).
- - Reads `.mcp.json` (Claude Code), `.cursor/mcp.json` (Cursor) and `.vscode/mcp.json` (VS Code), with `${VAR}` expansion.
- - Shows a field-level diff (`description`, `inputSchema`, `annotations`…) and flags known tool-poisoning tells:
- hidden instructions, `<IMPORTANT>` tags, "don't tell the user", sensitive paths, exfiltration wording, invisible
- Warning: Contains a <important> block, a common wrapper for instructions hidden in a tool description
- Line 23 on GitHub
- Unicode, cross-tool overrides, descriptions that more than double in length. These are signals, not verdicts.
- - Hashes are byte-identical to the [MCP Drift Observatory](https://drift.getvda.ai)'s `drift_hash` (`canon_version` 1), so you can compare your
- pins with a public, independent record.
- ## Quick start
- ```bash
- npx mcp-drift-guard lock # probe your servers, write mcp-tools.lock.json, then commit it
- npx mcp-drift-guard check # exit 1 if anything changed
- npx mcp-drift-guard update # accept the change after you've reviewed it (then check the git diff)
- npx mcp-drift-guard observe https://mcp.example.com/mcp # any public server, no config needed
- ```
- ## GitHub Action
- ```yaml
- - uses: actions/checkout@v6
- - uses: getvda-ai/mcp-drift-guard@v0
- env:
- EXAMPLE_API_KEY: ${{ secrets.EXAMPLE_API_KEY }} # anything your .mcp.json references as ${VAR}
- ```
- Run it on pull requests **and on a daily schedule**: servers change without a commit on your side. See
- [`examples/mcp-drift-guard.yml`](examples/mcp-drift-guard.yml), which also opens an issue when the nightly run finds drift.
- | Input | Default | |
- |---|---|---|
- | `mode` | `check` | `check`, `update` or `lock` |
- | `lockfile` | `mcp-tools.lock.json` | |
- | `config` | the three files above | comma- or newline-separated |
- | `fail-on` | `any` | `any`, `signals` (only changes that carry a poisoning signal) or `never` |
- | `stdio` | `true` | `false` checks remote servers only |
- | `timeout-ms` | `30000` | per server |
- | `observatory` | `true` | fall back to the observatory for remote servers this run can't read |
- | `public-feed` | `true` | the public history cross-check below |
- | `reader-card` | | your A2A Agent Card URL, if you have one |
- Outputs: `drift`, `signals`, `report` (Markdown, also written to the job summary) and `json`.
- `stdio: true` runs the commands in your MCP config, as Claude Code or Cursor would. A pull request that edits
- `.mcp.json` can therefore run code in CI, as it already can through your test scripts. Use `stdio: false` on
- untrusted pull requests if that matters for your repo.
- ## Claude Code plugin
- ```text
- /plugin marketplace add getvda-ai/mcp-drift-guard
- /plugin install mcp-drift-guard@getvda
- ```
- - **At session start** it checks your remote servers against the lockfile. It stays silent unless something drifted;
- then it tells you, and tells Claude to treat the changed descriptions as unreviewed.
- - **A skill** that locks, checks and explains drift on request, and never accepts a change without you.
- - **The MCP Drift Observatory's tools** (`drift_check`, `drift_diff`, `drift_probe`, `drift_status`, `drift_hash`) for public
- servers. The skill asks before sending a server URL to them.
- ## Observatory fallback
- A remote server can't be read from a run when its secret isn't in that job, when a firewall blocks the runner, or when
- it's down. Before, that server just showed "not checked". Now, for public `https` servers, mcp-drift-guard asks the
- [MCP Drift Observatory](https://drift.getvda.ai) what it saw at its last nightly crawl (`drift_check`, plus `drift_diff`
- when there were changes) and compares that hash with your pin. The hashes are the same bytes, so the comparison is exact.
- | Locally | Observatory | Result |
- |---|---|---|
- | read OK | (not asked) | full diff, as before |
- | can't read | knows it, same hash as your pin | ✅ unchanged (via observatory, with the date it was seen) |
- | can't read | knows it, different hash | ⚠️ drifted, with the observatory's change summary and a link to its tool-level diff |
- | can't read | doesn't know it | not checked |
- - `lock` pins such a server from the observatory's hash alone. There is no stored definition to diff yet, so the first
- `update` after you can read it stores the full definitions.
- - Observations older than 7 days are marked **stale**.
- - **What leaves the run:** each such server's origin and path, for example `https://mcp.example.com/mcp`. Never the
- query string, headers, credentials or tool definitions. Hosts that look private (`localhost`, `10.x`, `192.168.x`,
- `.internal`, `.corp`…), plain `http`, and URLs with an unexpanded `${VAR}` in the host or path are never sent.
- - It covers what the observatory can see: public servers that answer without auth. Turn it off with `observatory: false`,
- `--no-observatory` or `DRIFT_GUARD_OFFLINE=1`.
- ## The public history cross-check, over RSS-A
- The [MCP Drift Observatory](https://github.com/mikerawsonnz/mcp-drift-observatory) crawls public MCP servers every night,
- hashes what each declares and publishes every change as a **signed feed**. It uses [RSS-A](https://github.com/getvda-ai/rssa)
- (RSS for Agents), an open protocol: an ordinary Atom/RSS feed that is linked from the publisher's A2A Agent Card,
- with every entry signed by the publisher's key.
- On each run, mcp-drift-guard reads the last 14 days of that feed, **verifies every signature**, and adds a public history
- to the report, for example "this server changed its tools on 6 of the last 14 days". Only the feed is fetched: your
- server list, config and tool definitions never leave the run. The request carries a short anonymous client id (a hash
- of the repository name in CI, a random id locally) so the publisher can count distinct readers, never who they are.
- Turn it off with `public-feed: false`, `--no-public-feed` or `DRIFT_GUARD_OFFLINE=1`.
- You can subscribe to the same feed in any feed reader:
- <https://hub.rssa.getvda.ai/network/@getvda-mcp-drift-observa-e131.atom>
- ## What it can't see
- It compares **declarations**, not behaviour. A server that keeps its tool list byte-identical while changing what the
- tools do is invisible to this check, and to every other client-side check. Servers that are unreachable or need
- credentials you didn't provide are reported as "not checked", never as "unchanged".
- ## Licence and privacy
- Apache-2.0. Maintained by [getvda.ai](https://getvda.ai) (Rawson Consulting B.V.). See [PRIVACY.md](PRIVACY.md).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.