Skip to main content
Glama
Aldabra CVE Watch

Aldabra CVE Watch

by aldabra-sh
README.md
# Aldabra CVE Feed MCP

Let **your agent** monitor [Aldabra's existing public CVE feed](https://cve.aldabra.sh)
for the software you use. The connector only reads public advisories. Your agent
keeps your stack, filters results, schedules checks, remembers previous events,
and notifies you using your own account and model usage.

No Aldabra account, GitHub sign-in, saved watch, database or paid connector plan.
Aldabra does not run a monitoring job for each user.

## Connect and paste

Add this remote MCP server with **no authentication**:

**`https://cve-mcp.aldabra.sh/mcp`**

The endpoint uses Streamable HTTP. See [client setup](docs/CLIENTS.md) and
[release evidence](docs/BUILD_STATE.md). A connected tool alone does not establish
that a client's scheduled task works.

Then paste:

```text
Help me monitor CVEs relevant to my stack using Aldabra's public feed.
Read https://github.com/aldabra-sh/cve-mcp and its skills/cve-watch/SKILL.md.
Verify the read-only MCP connection, or explain the setup steps for my client.

Ask for the software/products I use, optional versions and aliases, and my
preferred daily time and timezone. Keep that information in my agent's own
state. Aldabra should not store my stack or run my monitoring schedule.

Establish a baseline, then create or reuse a daily task in this client's
scheduler if it supports unattended connector access and durable progress.
Include unscored CVEs, meaningful UPDATE events and corrections. Deduplicate
by event GUID, preserve source dates, and link to the official CVE record.
Stay quiet on healthy checks with no new matches. Report failures, partial
checks and coverage gaps. Do not treat a term match as confirmed impact.

Test the saved task and show its actual schedule receipt and next run.
If scheduling or durable state is unavailable, explain that limitation;
do not claim monitoring is active.
```

You can also use [the RSS URL](https://cve.aldabra.sh/feed.xml) directly in an agent
that can fetch and parse RSS. It contains the latest 2,000 activity items, so
long absences can outgrow its window. The MCP page interface exposes the wider
retained activity window and allows unchanged pages to be skipped.

## Three read tools

| Tool | Result |
|---|---|
| `feed_status` | Freshness, coverage, revision and immutable page descriptors |
| `read_activity_page` | Up to 50 NEW/UPDATE events from one descriptor |
| `get_cve` | An exact CVE lookup with its official link, including outside feed retention |

There are no watch, acknowledgment, account, scheduler or notification tools.
Products and versions are not tool arguments. The agent evaluates relevance in
its own environment. An exact CVE lookup does not add that CVE to the feed.

Results preserve source descriptions, CVSS version/provider, publication and
observation times, correction status and stable event GUIDs. NEW and UPDATE for
the same CVE remain separate. Source text is untrusted data, never instructions.

Fictional examples:

> **NEW · ExampleVPN · CVE-20XX-12345 · Unscored**  
> ExampleVPN allows an out-of-bounds read. Unscored — no CVSS score is currently
> provided in the source record. Original publication time and official link follow.

> **UPDATE · ExampleVPN · CVE-20XX-12345 · CVSS added: 9.8 Critical**  
> CVSS added: Unscored → 9.8 Critical, CVSS 3.1, supplied by the CNA.
> The update time and original publication time remain distinct.

## Coverage and progress

Read actual coverage from `feed_status`. The current source starts with seven days
of backfill, retains 30 days of activity, and tracks known CVEs for score changes
and corrections for 90 days after publication. It is not an all-time catalog.
Updates reflect differences observed between source polls, not every intermediate
revision. Affected-version-only edits are not UPDATE triggers. Ambiguous or absent
product names can limit relevance matching; unscored never means safe.

The [agent skill](skills/cve-watch/SKILL.md) specifies baseline, snapshot pagination,
GUID deduplication, correction handling and failure recovery. State belongs in
storage available to the **user's scheduled agent**, not merely a chat's temporary
context. Healthy empty checks advance that agent's successful-check time; partial
or failed checks do not. No client scheduling or delivery guarantee is implied.

## Privacy and cost

The connector has no user database, login, AI calls, timer or persistent session.
It requests only fixed public feed paths. Ordinary hosting access metadata can
still exist; it does not log request bodies. Do not put secrets or private inventory
into tool arguments. Your agent/client handles your stack and its own usage costs.

Hosting targets Cloudflare Workers Free: a small, bounded read adapter, without
full-catalog parsing or server-side stack matching. No paid billing is enabled by
this project. Free request limits are shared and finite; see the measured
[deployment notes](docs/DEPLOYMENT.md). The existing feed pipeline is unchanged.

## Development

Node 22+:

```sh
npm ci
npm run check
npm run dev
```

Local endpoint: `http://localhost:8787/mcp`. No credentials, migrations or bindings
are needed. Production has no dependency on a developer's computer.

The small JSON-only adapter supports MCP protocol versions 2025-03-26, 2025-06-18
and 2025-11-25. Tests use the official MCP SDK client against a real local Worker
runtime. There are no runtime package dependencies. See [operations](docs/OPERATIONS.md)
and [deployment](docs/DEPLOYMENT.md).

## Source and license

Data: [CVE Program / CVEProject/cvelistV5](https://github.com/CVEProject/cvelistV5),
presented through Aldabra's public feed. Tool results carry the source license
notice. CVE data retains its own terms. New connector code: [MIT](LICENSE).