Repository navigation
Read-only MCP server for DefectDojo Community Edition #16712
systechgd
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
DefectDojo Pro ships a built-in MCP server. Community Edition doesn't, so I built one for
our own triage and have open-sourced it:
https://github.com/systechgd/defectdojo-community-mcp
Unofficial, not affiliated with DefectDojo Inc. If you're on Pro, use the built-in one
— it's vendor-maintained and considerably more capable. This exists only because CE has
nothing.
The problem it came out of
We run a microservice estate: dozens of repositories, one DefectDojo product per repo,
and several scanners feeding each one — SAST, dependency scanning, secret scanning, DAST,
container and IaC scans. That setup is normal, and it produces a specific kind of pain
that has nothing to do with DefectDojo being bad at its job:
transitive dependency and the same vulnerability lands in most repos at once. DefectDojo
reports it honestly — once per product, because they genuinely are separate findings.
But it's one fix.
same CVE in twenty products isn't a duplicate, so answering "which services are affected
by this one?" means paging through products one at a time or writing API scripts.
and several engineers independently do the same upgrade, or you file one ticket with a
hand-assembled list of affected repos — and whatever you missed quietly keeps the
vulnerability in production.
the SAST tool's rule ID and the DAST tool's endpoint can all point at one root cause,
with nothing obvious tying them together.
out-of-scope and already-mitigated findings all come back unless you filter them out
every single time.
editor or a chat window. Triage becomes copy-paste between the two.
What you can do with it on CE
It talks to the ordinary REST API v2 with a normal API token — nothing to enable
server-side, no minimum DefectDojo version. Seven read-only tools, and in practice you
drive them by asking:
every product, where actionable means active and not false-positive, duplicate,
risk-accepted, out-of-scope or mitigated.
group_critical_by_vulngroups criticals across every product by CVE, falling back toCWE + title, and returns each group with the full list of affected products, widest
blast radius first. One group becomes one ticket with a complete affected-services list,
instead of a pile of near-identical tickets or an incomplete one.
what's theirs, without them needing a DefectDojo login.
mitigation, references, CWE/CVE, component and version, file and line. Enough to write
the ticket without opening the UI.
the issue text, grouped by root cause. It's tracker-agnostic: the server never talks to
your tracker, it just hands over clean data.
duplicates, which is usually a signal about scanner configuration rather than code.
reachability, so a misconfiguration doesn't masquerade as "no findings".
On read-only
It's a structural guarantee, not a convention. The HTTP client class exposes no method
capable of POST/PUT/PATCH/DELETE, so a mutating tool can't be added by accident, and a
test asserts that every request every tool issues is a GET. If write support ever happens
it will be a separate server, so operators can grant read without granting write.
That matters for a practical reason: it makes the server safe to hand to developers. The
worst case is someone reads findings they could already see in the UI — nobody can mark
something a false positive or mitigate it by talking to a chatbot.
Two things worth saying plainly:
access, and a leaked read-only token can't be used to alter your findings.
messages all end up in finding text, and that text goes to a model. Treat tool output as
data, never as instructions. SECURITY.md covers the threat model.
Install
Also on GHCR, and listed on the official MCP Registry as
io.github.systechgd/defectdojo-community-mcp.Apache-2.0. CI covers Python 3.10-3.13 and both MCP SDK generations.
What I'd like back
Mainly whether the grouping heuristic matches how other people triage. CVE first, then
CWE + title, works for our scanner mix, but I'd expect it to fall over on setups I
haven't seen — particularly where findings arrive without CVEs. Issues and PRs welcome,
and I'm happy to be told the approach is wrong.
All reactions