Skip to content
Back to search

warden

Self-hosted MCP server that feeds AI agents new CVEs (NVD, CVE.org, KEV, EPSS) for the products you track

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
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/jacob-tate--warden/badge.svg)](/listings/jacob-tate--warden)

HTML

<a href="/listings/jacob-tate--warden"><img src="/api/listings/jacob-tate--warden/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

cargo build --release
./target/release/warden hash-key claude-code     # prints the key once + the config line
export WARDEN_API_KEYS=claude-code:<hash printed above>
export WARDEN_PUBLIC_URL=https://warden.example.com   # host is allowed on /mcp
./target/release/warden serve                    # listens on WARDEN_BIND (default 0.0.0.0:8080)

Several keys: `WARDEN_API_KEYS=claude-code:<hash>,routine:<hash>`. Each key name is that client's default `consumer`, so every client has its own "already seen" cursor.

Connect Claude Code:

claude mcp add --transport http warden https://HOST/mcp --header "Authorization: Bearer KEY"

Put an HTTPS reverse proxy in front (forward `Host`, do not buffer `/mcp`). `/healthz` is unauthenticated and reports only `ok` or `stale`.

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.

README.md · README · 73 lines
  1. # warden
  2. A small self-hosted monitor that tells an AI agent what changed in the software you run.
  3. The first monitor is a **CVE feed**: it watches new and changed CVEs, CISA KEV and EPSS, filters them to the products you track, and hands back a short, ranked JSON list of what's new since the last time you asked. It's exposed as an **MCP server**, so Claude (or any MCP client) can read the feed and add or remove tracked products as it goes.
  4. SSL certificate and HTTP uptime monitors are planned for later.
  5. ## Status
  6. The CVE sync, matching, discovery and the MCP server work, with API keys (Claude Code) and a built-in single-user OAuth server (claude.ai custom connectors and routines).
  7. ## Quick start
  8. ```bash
  9. cargo build --release
  10. ./target/release/warden hash-key claude-code # prints the key once + the config line
  11. export WARDEN_API_KEYS=claude-code:<hash printed above>
  12. export WARDEN_PUBLIC_URL=https://warden.example.com # host is allowed on /mcp
  13. ./target/release/warden serve # listens on WARDEN_BIND (default 0.0.0.0:8080)
  14. ```
  15. Several keys: `WARDEN_API_KEYS=claude-code:<hash>,routine:<hash>`. Each key name is that client's default `consumer`, so every client has its own "already seen" cursor.
  16. Connect Claude Code:
  17. ```bash
  18. claude mcp add --transport http warden https://HOST/mcp --header "Authorization: Bearer KEY"
  19. ```
  20. Put an HTTPS reverse proxy in front (forward `Host`, do not buffer `/mcp`). `/healthz` is unauthenticated and reports only `ok` or `stale`.
  21. ## claude.ai custom connector (OAuth)
  22. claude.ai connectors (and scheduled routines, which reuse them) cannot send a static key, so warden contains a small OAuth 2.1 authorization server for one user: you.
  23. 1. Pick a passphrase and hash it (reads stdin; hidden prompt on a terminal):
  24. ```bash
  25. ./target/release/warden hash-passphrase
  26. ```
  27. Put the printed `WARDEN_OWNER_PASSPHRASE_HASH='$argon2id$...'` in the environment. In docker compose or an `.env` read by compose, write every `$` as `$$`. Without this variable the OAuth endpoints answer 404 and only API keys work.
  28. 2. Set `WARDEN_PUBLIC_URL` to the exact public HTTPS origin (no path) and `WARDEN_TRUSTED_PROXIES` to your reverse proxy (see below).
  29. 3. In claude.ai: Settings > Connectors > Add custom connector, URL `https://HOST/mcp`. Leave client ID and secret blank: Claude registers itself (dynamic client registration). Claude opens warden's consent page; enter the passphrase. The "connection name" on that page (default `claude-ai`) becomes the consumer `oauth:<name>`, i.e. its own "already seen" cursor. Routines use the same connector, so they share that cursor.
  30. 4. Claude Code can keep using an API key (`--header`), or use OAuth (`claude mcp add --transport http warden https://HOST/mcp`, then `/mcp` login); it redirects to a loopback address.
  31. Redirect URIs are allowlisted (`WARDEN_OAUTH_REDIRECT_ALLOW`; default `https://claude.ai/api/mcp/auth_callback` and loopback `localhost`/`127.0.0.1`/`[::1]` on any port). Access tokens last 1 hour; refresh tokens 30 days, sliding (`WARDEN_OAUTH_REFRESH_DAYS`), rotated on every use. Revoke everything by deleting rows from `oauth_tokens` or removing the connector.
  32. Optional hardening: Anthropic's outbound ranges are `160.79.104.0/21` and `153.61.208.0/20`. You may restrict `/mcp`, `/token`, `/register` and `/.well-known/*` at the proxy to them. Do not restrict `/authorize` the same way: that page is opened by your own browser, not by Anthropic, so an allowlist of Anthropic ranges would block you. Exempt `/authorize` (or limit it to your own IP / VPN instead).
  33. Behind a reverse proxy, set `WARDEN_TRUSTED_PROXIES` (comma-separated IPs/CIDRs of the proxy **as warden sees it**; in Docker that is the bridge gateway or the proxy container, e.g. `172.16.0.0/12`, never `127.0.0.1`). It defaults to empty, and then all clients look like the proxy to the rate limiters; warden logs a warning at startup if OAuth is enabled without it.
  34. Once the connector is working, set `WARDEN_OAUTH_REGISTRATION=closed` (default `open`) to switch dynamic client registration off: `/register` then returns 403 and the metadata omits `registration_endpoint`. Existing connectors keep working; set it back to `open` to add another.
  35. The `consumer` argument of the MCP tools (a cursor name, `^[a-zA-Z0-9._:-]{1,64}$`) is not a security boundary: any authenticated caller can name any consumer, including the `oauth:<name>` cursor of a connector. warden is single-user, so this is deliberate.
  36. Docker image, compose file and reverse proxy setup: see [docs/deploy.md](docs/deploy.md).
  37. ## Planned design
  38. - **Language:** Rust (axum, [rmcp](https://github.com/modelcontextprotocol/rust-sdk)), SQLite, one Docker container.
  39. - **Data sources:**
  40. - CVE Program cvelistV5 delta files, for new records with CNA data.
  41. - NVD JSON 2.0 "modified" feed, for CPEs and NVD scores.
  42. - NVD API 2.0, only for single-CVE lookups.
  43. - CISA KEV.
  44. - FIRST EPSS.
  45. - **Matching:** CPE vendor:product, CNA vendor/product, and keywords.
  46. - **MCP tools:**
  47. - `get_new_cves`, `ack`, `cve_detail`
  48. - `list_tracked`, `add_tracked`, `update_tracked`, `remove_tracked`
  49. - **Auth:** static API keys and/or the built-in OAuth server, served over streamable HTTP behind a reverse proxy.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.