Kerberos single sign-on for MCP against FreeIPA: authenticate with the ticket a developer already holds from logging in, authorize by directory group, and keep no passwords, API keys or per-developer secrets anywhere.
Status: under heavy development. Names, installer flags and file layouts
still change between commits, so read the log before upgrading a deployment.
Known rough edge: the Firefox that setup.ps1 installs inside WSL for Windows
workstations is not stable yet. The bridge, wslssh, wslgit, krb-fetch and
krb-git do not depend on it.
The server's optional policy editor. Each tool maps to the FreeIPA groups that may call it and, if it may act downstream as the caller, to the one service it may reach.
This is a system you deploy, in two halves, with one hook for your own code. It is not a library to import, and it is not an MCP client. You run the server on one host in your FreeIPA realm, hand the client kit to every workstation, and put your tools in one site file that this repository never carries.
| Part | What you get | Where it runs | In the tree |
|---|---|---|---|
| The server | a Kerberised MCP server on the official Python SDK, behind nginx. Every request is authenticated offline by SPNEGO against a keytab; every tool call is authorized deny-by-default by FreeIPA group; every decision is an audit line. Optional: a tool may act downstream as the caller (constrained delegation), and a browser editor for the policy (pictured) | one Linux host in the realm | server/ |
| Its installer | one script that creates the service account, venv, keytab, systemd unit, nginx vhost and certificate, then publishes the client kit; a read-only verifier that checks the result | the same host | server/install/ |
| The client kit | a stdio bridge that mints a fresh Negotiate token per request, the mcp-krb launcher an MCP client is pointed at,krb-fetch for byte-exact files andkrb-git for git over Kerberos, and the reverse bridge: a shared dev host that holds no ticket uses your workstation's through a socket forwarded byssh -R |
every workstation and shared dev host | client/bridge/ |
| Workstation provisioning | setup.sh (Linux: enrol, then install),setup.ps1 (Windows without domain join: WSL2, Kerberos ssh, VS Code Remote-SSH, a browser that can SSO, the bridge),setup-macos.sh , uninstallers driven by install manifests, and a provisioning web page the installer serves |
workstations | client/ ,client/web/ |
Your own tools go in one Python file, loaded through MCP_SITE_TOOLS, each
declaring the groups that may call it. The page's site-specific sections go in
one HTML fragment, through CLIENT_SITE_SECTIONS. Neither lives in this
repository, so a deployment never maintains a fork of a file it did not write.
Behind all of it sits SECURITY.md, the review with the threat
model, ranked findings and deployment checklist, and a hermetic test suite that
runs on Windows and Linux with no KDC.
What that gives a team, concretely:
- A developer types one password, at login. From then on an MCP client such as Claude Code reaches the internal server with nothing to configure and no secret stored anywhere. Disabling the account in FreeIPA ends the access.
- The server sees the real principal on every call, decides by directory group, and logs who did what under their own name.
- A tool can call CI, a forge or any other Kerberised service as the person who asked, so the downstream system attributes the action to them and bounds it by their permissions there.
- On a shared dev host with no ticket, the same things work: reaching the MCP server, fetching a file, and git clone, pull and push, all through a socket the workstation forwards. No credential is copied to the host.
- A Windows workstation that is not domain-joined gets passwordless ssh, VS Code Remote-SSH and single sign-on in a browser, through WSL2.
What it assumes: FreeIPA (or another MIT Kerberos realm with an LDAP directory you can adapt the group lookup to), Linux hosts for the server and the shared machines, and an MCP client that speaks stdio. The tools it ships are stubs; the scaffolding around them is the product.
MCP has no enterprise single-sign-on story. Claude Code's MCP client can attach an OAuth flow or a fixed header to its requests, but it cannot speak HTTP Negotiate/SPNEGO, the standard way a browser or a command-line tool proves a Kerberos identity to a web service. So it cannot authenticate directly to a Kerberized internal MCP server. This repository is a worked, reviewed answer for FreeIPA shops: it closes that gap using the Kerberos ticket a developer already holds from logging in to a FreeIPA-enrolled machine. No passwords, no API keys, no per-dev secrets.
It is a reference implementation. The MCP tools it ships (whoami,
list_projects, restart_service, trigger_build) are stubs; the value is the
authentication, authorization, delegation and deployment scaffolding around them,
meant to be read, reviewed and adapted to your own tools. Your own tools go in a
separate file that this repository never carries, loaded through MCP_SITE_TOOLS
(see Configuration), so a deployment does not end up
maintaining a fork of a file it did not write.
A tool can also act on behalf of the caller. One of those stubs, trigger_build,
shows the delegation path: a tool calls a downstream Kerberized service as the human
who invoked it, so the downstream logs the real person rather than a shared service
account. It is off by default, reaches any Kerberized service (CI is only the
example), and has the one genuinely subtle security story here. Its own section,
On-behalf-of delegation, covers it.
One idea holds this together: a user credential exists in exactly one place, the workstation. Every other machine either has its own machine identity, or has nothing at all. Nothing is copied, forwarded or minted to make a second you.
┌────────────────────────────────────────┐
│ FreeIPA: KDC · DNS · CA │
│ │
│ issues TGTs to people │
│ keytabs to machines │
│ the CA that signs every host │
└────────────────────────────────────────┘
▲ ▲ ▲
① kinit │ ② keytab │ ③ ca.crt │ plain HTTP,
(TGT) │ at enrol │ to a new │ hash checked
│ │ machine │ out of band
╔═════════════════════╪════════════════╪════════════════╪═════════════════╗
║ WORKSTATION ┘ │ │ ║
║ the ONLY place a user ticket lives │ │ ║
║ (on Windows: inside WSL) ║
║ ║
║ MCP client ───stdio──▶┌────────┐ ║
║ │ │ bridge │───── ④ SPNEGO ──────────────────┐ ║
║ │ runs └───┬────┘ │ ║
║ ▼ │ ⑦a --listen │ ║
║ krb-fetch ──┐ ▼ │ ║
║ krb-git ────┤ ~/.mcp-krb.sock 0600 │ ║
║ ├──▶ ~/.mcp-krb-fetch.sock 0600 ⑦b │ ║
║ └──▶ ~/.mcp-krb-git.sock 0600 ⑦c │ ║
║ ▲ │ ║
║ ssh ────── ⑤ ─────┐ │ │ ║
╚══════════╪═══════════╪═══════╪══════════════════════════════════════╪═══╝
│ │ ║ │
│ │ ║ ssh -R, one per socket │
│ │ ║ the SOCKETS are forwarded, │
│ │ ║ never the credential │
│ │ ║ │
╔══════════╪═══════════╪═══════╪══════════════════════════════════════╪═══╗
║ SHARED │ DEV HOST ║ │ ║
║ ◀────────┘ ⑤ ssh in ▼ │ ║
║ /run/user/N/mcp-krb.sock ◀── ⑦a │ ║
║ /run/user/N/mcp-krb-fetch.sock ◀── ⑦b │ ║
║ /run/user/N/mcp-krb-git.sock ◀── ⑦c │ ║
║ ▲ ▲ │ ║
║ stdio │ │ asks the workstation │ ║
║ MCP client ──▶ remote bridge krb-fetch · krb-git │ ║
║ │ ▲ │ ║
║ └────────── runs ──────────────────┘ │ ║
║ │ ║
║ has a HOST keytab. NO user ticket. nothing here to steal. │ ║
╚═════════════════════════════════════════════════════════════════════╪═══╝
│ │
│ ⑧ fetch a file, byte-exact · clone, pull, push │ ④
▼ ▼
┌────────────────────────────────┐ ┌──────────────────────────────┐
│ any Kerberised service │ │ the MCP server │
│ git · CI · wiki · artifacts │◀─ ⑥ ───│ 401 on every path │
│ │ on-be- │ holds a SERVICE keytab │
│ whatever your shop runs; this │ half-of│ your tools live here │
│ repo assumes none of them │ │ refuses a forwarded TGT │
└────────────────────────────────┘ └──────────────────────────────┘
┌────────────────────────────────┐
│ the provisioning page │ the one deliberately anonymous surface:
│ served by this installer │ a machine that is not enrolled yet has
│ client scripts, anonymous │ no ticket, so the bundle it needs in
│ on purpose │ order to enrol cannot sit behind SPNEGO
└────────────────────────────────┘
① kinit The only step involving a human secret, on the only machine holding
one.
② Host keytabs Each enrolled machine gets its own identity. That proves a machine is itself; it says nothing about who you are and cannot be used to become you. A Mac never does this, which is why macOS leaves the realm untouched where Linux does not.
③ CA bootstrap A new machine trusts nothing, so it fetches the realm CA over plain HTTP and checks it against a SHA-256 obtained elsewhere. That comparison is the whole check, which is why the hash must not come from the same infrastructure that serves the certificate.
④ MCP The client speaks stdio to the bridge; the bridge speaks SPNEGO to the server. The ticket never moves, because the bridge is already where it is.
⑤ SSH GSSAPI with GSSAPIDelegateCredentials no. Your ticket is not
forwarded, which is why a shell on a shared host has no credentials. Design, not
gap.
⑥ On-behalf-of A tool can act as you against a downstream Kerberised service using constrained delegation with a narrow evidence credential, refusing a forwarded TGT. The example tool targets CI; the mechanism cares about none of that. See On-behalf-of delegation.
⑦ Forwarded sockets --listen and --fetch-listen serve MCP and fetching
over 0600 Unix sockets. ssh -R forwards them to a shared host, where
mcp-krb-remote-bridge.py joins a client's stdio to them. The far end holds no
credential, imports no crypto, and loses the channel when the session ends.
⑧ Fetching a file --fetch writes a URL to disk over SPNEGO, for content
that must arrive byte-exact and therefore must not pass through a model. On a
shared host it goes via ⑦b, so the GET and the allowlist both happen where the
ticket is.
Fair question, since the assistant already has a shell and your ticket, and could call an internal API itself.
Authorisation cannot be enforced on the client. A tool that runs on your machine is a tool you can edit. "Only this group may trigger a build" living in a script on a workstation is a suggestion; behind an authenticated server it is a rule, because the check happens somewhere the caller does not control.
Audit needs a chokepoint. Every call is recorded against the Kerberos principal that made it, allowed or denied. Fifty workstations calling an API directly produce no such record.
Delegation needs a service principal. Acting as you against another service, without holding your TGT, requires an identity the realm knows and rules it enforces. A shell script cannot be granted that, and should not be.
And the corollary: things you could already do belong in the shell, not
behind a tool. krb-fetch and krb-git are deliberately not MCP tools. The
assistant already has your ticket, so wrapping an HTTP GET or a git pull in a
server call would add a hop, a schema and an audit line while changing nothing
about what is possible.
The test each tool should pass is: could the caller do this themselves, unlogged,
if the tool did not exist? When the answer is yes, it does not belong here.
Worth stating plainly, because shared hosts are the case they exist for and several people at different privilege levels may be logged into one at once.
The sockets are 0600, so an unprivileged peer cannot use them. Root on that
host can, while your session is open, which on a box where colleagues hold sudo
means those colleagues. After you disconnect the socket file remains but nothing
answers it.
So the exposure is bounded by your session rather than by a ticket lifetime. The
alternative, running kinit on the shared host, leaves a ticket cache that root
there can read and use to become you everywhere in the realm, for its full
lifetime, still valid after you log out.
It is structurally ssh-agent forwarding, with a narrower grant: agent forwarding
gives SSH-to-anywhere, these give one service and one allowlist. The operational
rule that follows is to avoid mixing privilege levels and sudo on one host.
krb-git adds one hop to this picture, a loopback port, because git cannot open
a Unix socket. A port is reachable by every user on the host, so the forwarder
behind it checks the uid of each connection against the kernel's TCP table and
refuses anyone else. The port grants nothing the socket did not.
A developer logs in to their workstation. That is the only time anyone types a password. From then on Claude Code reaches the internal MCP server with no further prompt and nothing to configure:
klist # a ticket is already there, put in place at login
claude mcp list # internal-tools: connected
The server learns who they are on every request, decides what they may call from their directory group membership, and records the decision under their real name. Because nothing was set up with a shared secret, there is nothing to rotate, leak, or hunt down later: removing the account in FreeIPA removes the access.
server/install/run.sh is the installer, and it covers every step so nothing has
to be done by hand: the service account, the --system-site-packages venv, the
code deploy, the keytab retrieval and permission contract, mcp-server.service,
the nginx vhost, the certificate, and the certbot deploy hook. It is idempotent, so
re-running it converges, and it refuses to proceed on a value it cannot resolve
rather than defaulting to a placeholder.
Two prerequisites are human on purpose, because they consume admin credentials the installer deliberately refuses to hold (it preflights for both and stops with the exact command if either is missing):
- The host is enrolled in FreeIPA (
ipa host-addwith a one-time OTP, thenipa-client-installon the box). - The IPA service principal
HTTP/<mcp-fqdn>exists and this host is allowed to retrieve its keytab.run.sh --create-ipa-servicewill do this for you, but only if you already hold an admin ticket.
Then, with server/install/site.env.example copied to /etc/mcp-server/site.env
and filled in:
sudo sh server/install/run.sh --site-env /etc/mcp-server/site.env
sh server/install/verify.sh <fqdn>
The server's Python environment is the one part worth understanding before you run
it. The venv must be created with --system-site-packages so it keeps the OS
python3-gssapi, with mcp and uvicorn installed from
server/requirements.lock.txt. An isolated venv produces a server that starts
cleanly and then fails every SPNEGO handshake. The installer gets this right; the
note is here for anyone building the environment by hand.
Two warnings that bite in practice:
- The installer runs as root and deploys whatever it finds in the source tree, so
it walks every parent directory up to
/and refuses a checkout that is not root-owned or is group/world-writable. Agit pullinto a home directory fails this by design; copy the tree to a root-owned path first. run.shrestartsmcp-serverunconditionally at the end, becausesystemctl enable --nowdoes nothing to an already-running unit and would otherwise leave yesterday's process serving new code with every check still green.
verify.sh is read-only and worth re-running after any change: 14 checks that turn
a silently broken install (a 401ing ACME challenge, a policy that denies everyone, a
stale process, a delegation flag half-configured) into a loud one. Full installer
detail, what verify.sh asserts, and what the host serves live in
server/ and SECURITY.md.
One command per platform, documented end to end in
client/README.md: client/setup.sh for IPA-enrolled Linux,
client/setup.ps1 for non-domain-joined Windows (it stands up Kerberos SSH and the
bridge inside WSL2, where the ticket lives). Both fetch the kit over HTTPS pinned to
the realm CA and never pipe a download to a shell.
macOS has no script. It is a short, documented manual Kerberos-client setup
(see client/README.md and the provisioning page the MCP host
serves), separate because macOS ships Heimdal rather than MIT krb5 and needs a
specific kdc = tcp/... line that a naive config gets wrong. Both halves of the
kit apply there: SSH via GSSAPI is proven, and the MCP bridge has a documented
path too (pip install gssapi; the wheel links the system Heimdal, so the bridge
reads the same ticket kinit writes), verified by linkage though not yet
exercised end to end from a Mac.
Shared team workspace: once users reach a host with their own Kerberos identity, a
directory the whole team can edit needs no extra machinery. FreeIPA already puts
everyone in the irs group, so one owned root:irs at mode 2770 (setgid)
with a default ACL (setfacl -d -m g:irs:rwx) is group-writable, and new files
inherit both the group and the writability. Keep it under /srv rather than
/home/<name> (IPA auto-creates home directories there), and keep secrets out of a
group-writable path.
- One password, typed at login, and nothing stored anywhere afterwards. No API key to leak, rotate, or discover in a config file two years later. Disable the account in FreeIPA and the last service ticket dies within hours by its own validity window.
- Every request is authenticated on its own, offline, against the keytab. A
half-finished handshake is never mistaken for a completed one, which is the bug
class behind NTLM-style pre-auth bypasses (
[C1]: the acceptor gates onctx.complete, true GSS completeness per RFC 2743). - No session to steal. The server runs
stateless_http=True, so noMcp-Session-Idis ever issued or honoured as a bearer credential ([S1]). - Authorization is deny by default, per tool, by IPA group, resolved through SSSD's
local cache with no network call on the request path (
[S2]). - Who may call what is code rather than configuration:
authz.TOOL_GROUPSanddelegation.TOOL_TARGETSlive in security-owned files, so changing them is a reviewed change. A bounded operator overlay (MCP_DELEGATION_TARGETS) exists for site-specific targets, syntax-checked, capped, and unable to override a reviewed one. - The replay cache stays on, the service runs under a full systemd sandbox including
MemoryDenyWriteExecute, and the only long-lived secret is a keytab readable by the service account and root ([R1],[K1], an accepted residual documented in SECURITY.md). - A tool can act as the caller against any Kerberized downstream service, off by
default (
[D1]). An earlier finding, that a client could forward a full TGT to escape the target allowlist, is closed by a runtime check; seeOn-behalf-of delegation .
The client is a small local stdio bridge (client/bridge/):
Claude Code runs it as a subprocess and it forwards every JSON-RPC message to the
server, normally over HTTPS with a freshly minted Negotiate token. Two exceptions
are deliberate: the bridge accepts http:// only for localhost, 127.0.0.1 and
::1 and refuses any other scheme, and MCP_KRB_NOAUTH=1 drops the Authorization
header for local testing. Both let a developer exercise the transport without a KDC,
and neither belongs in a deployed config.
Server and bridge are both Python. The bridge is stdlib only: python3-gssapi
ships with ipa-client, so there is nothing to pip install on an IPA-enrolled
Linux workstation. (Inside WSL the distro is not enrolled and has no ipa-client,
so python3-gssapi and krb5-user are an explicit apt install, which setup.ps1
performs.)
The server is the official MCP Python SDK (FastMCP, streamable HTTP) with the Kerberos auth as a self-contained ASGI middleware in front, behind nginx (TLS). Non-domain-joined Windows workstations need one extra step, covered under Provision a workstation above.
One tool call, end to end. Every arrow carries [T] the transport and [A] what authenticates it.
Three arrows say [A] none. Those are deliberate, and each one is explained below the diagram.
WORKSTATION (IPA-enrolled) | NETWORK | MCP HOST mcp.example.internal | REALM
------------------------------------+---------+-----------------------------------+-------------
| | |
[ Claude Code ] | | |
| | | |
| (1) [T] stdio pipes, one JSON-RPC message per line |
| [A] none: the OS process boundary. The bridge runs as you, with |
| your ticket. This hop adds no authentication of its own. |
v | | |
[ mcp-krb-bridge.py ] | | |
| \ | | |
| \ (2) [T] Kerberos TGS-REQ, SPNEGO mech 1.3.6.1.5.5.2 ------------> [ FreeIPA KDC ]
| \ [A] your TGT, put in your ccache by SSSD when you logged in. | ipa.example
| \ The bridge never prompts for a password. | .internal
| \ Flags asked for: mutual_auth, out_of_sequence. | |
| \ never delegate_to_peer. See [CL1]. | |
| \ | | | |
| <---- service ticket for HTTP/mcp.example.internal -----------+-------+
| | | |
| (3) [T] HTTPS 443, TLS 1.2+, realm CA. New token every request. |
| [A] Authorization: Negotiate <base64 AP-REQ> (RFC 4559) |
+---------------------------->|-------->[ nginx ] |
| | TLS terminates here. Rate + conn
| | limits, 1 MB body cap, X-Forwarded-For
| | overwritten, security headers.
| | |
| | (4) [T] UNIX socket /run/mcp-server/mcp.sock,
| | root:<nginx-group> 0660, parent dir 0755
| | [A] none: file permissions are the control.
| | 0666 here would let any local user bypass
| | nginx entirely. Set by ExecStartPost as
| | root, because the sandboxed service holds
| | no CAP_CHOWN and cannot set it itself.
| | v
| | [ uvicorn ]
| | |
| | ===== Gate 1: who are you =====================
| | SpnegoAuthMiddleware + spnego_auth
| | reject NTLM, cap token at 64 KB
| | accept using KRB5_KTNAME's keytab
| | (that file is root:<grp> 0640)
| | require ctx.complete <- the real gate [C1]
| | pin mech to krb5/SPNEGO, regex the principal
| | require the realm to match MCP_REALM
| | fail -> 401 Negotiate / 403. Reason to the audit
| | log only, never to the caller. [C4]
| | |
| | pass -> scope['krb_principal'] = alice@...
| | v
| | ===== Gate 2: may you do this =================
| | require(ctx, '<this tool's own literal name>')
| | authz.TOOL_GROUPS[tool] -> IPA group set
| | os.getgrouplist via SSSD's local cache
| | no network. Any error -> deny. [S2]
| | |
| | every decision, allow or deny, -> JSON audit
| | line on stderr -> journald
| | v
| | [ the tool runs ]
| | :
| : : (5) optional, off by default. [D1]
| : : forward_header(ctx, '<own name>')
| : : is_narrow_evidence()? a forwarded TGT
| : : is refused here
| : : TOOL_TARGETS[tool] -> exactly one SPN
| : : [T] S4U2Proxy TGS-REQ ------> [ KDC ]
| : : [A] the evidence credential naming you
| : v
| : [ downstream ] sees alice, not the MCP
| : service account. Site supplies the HTTP call.
The bridge turns each MCP message into an HTTPS request carrying a fresh SPNEGO
token, which nginx terminates and hands to the Python server over a UNIX socket.
Gate 1 (spnego_asgi.py with spnego_auth.py) validates the ticket offline against
the keytab and answers who the caller is [C1]. Gate 2 (authz.py, the first line of
each tool) answers whether that caller may call the tool, deny-by-default by IPA
group [S2]. The three [A] none arrows are hops a stronger control already covers,
the OS process boundary and the UNIX socket permissions. Delegation (arrow 5) is off
by default. SECURITY.md is the reference for all of it.
Off by default. A tool can call a downstream Kerberized service as the caller, so
the downstream sees the real human rather than this server's shared service account.
The shipped example trigger_build forwards to a CI system, but the mechanism knows
nothing about CI: it reaches any Kerberized service the caller could reach, such as
an internal REST API, a directory, a database proxy, or a second MCP server.
The mechanism is evidence-based S4U2Proxy constrained delegation. It does not use
protocol transition, the variant that would let a service mint a ticket for a user
who never authenticated. When the caller authenticates, MIT composes a credential
naming them from the ticket they already presented; the server shows that to the KDC
and asks for a ticket to one named downstream service. Two limits follow: it cannot
act for a user who never called, since that user's ticket is the evidence, and it
cannot reach a service delegation.TOOL_TARGETS has not named (deny by default, one
target per tool, security-owned in code plus a bounded operator overlay).
The subtle part, and the reason this is more than a config switch, is that a
hostile client can set GSS_C_DELEG_FLAG and hand the server its full forwarded TGT
in place of a narrow evidence credential, which the realm's target allowlist does
not constrain. The server rejects that credential. is_narrow_evidence() reads
MIT's GSS_KRB5_GET_CRED_IMPERSONATOR, a marker the Kerberos library writes only on
the non-delegating accept path and from the server's own name, so a client cannot
forge it, and it accepts only an S4U2Proxy evidence credential composed by this
acceptor. It fails closed on every unresolvable case, including a GSSAPI too old to
answer the question. So an earlier finding, that a client could forward a TGT to
escape the allowlist, is closed by that runtime check, and the realm's allowlist
holds for everything the server uses.
Enabling it has a cost the docs spell out: the acceptor credential becomes usable for outbound authentication, which raises what a stolen keytab is worth under [K1]. Turning it on is a deliberate deployment decision. The shipped client does not delegate ([CL1]), and the full analysis is [D1] in SECURITY.md.
Skip this if constrained delegation is familiar. It is the least well known corner of Kerberos and the rest of this section is hard to follow without it.
Ordinary Kerberos proves one thing: I am stergios, and I am talking to this service. It says nothing about that service then talking to a third one on your behalf. But that is exactly what is wanted here: the MCP server has to reach Gitea as you, so Gitea's own permissions and audit log apply to the real person, without the MCP server ever holding your password or your TGT.
The naive answer is credential forwarding: you hand the service your TGT and it becomes you, everywhere, for the life of the ticket. That is what this design refuses, because a single compromised service then owns every user who called it, against every service in the realm.
S4U2Proxy, Service for User to Proxy, is the KDC feature that does it narrowly:
- You authenticate to the MCP server normally. That handshake leaves it holding your service ticket to itself . That ticket isevidence you turned up.
- The MCP server returns to the KDC: here is evidence stergios authenticated to me; give me a ticket to Gitea, on his behalf.
- The KDC checks whether this service is allowed to delegate to that target , and issues only if so.
Two properties follow, and they are the whole reason for the design.
No evidence, no ticket. The server cannot invent a session for somebody who never called it. Its reach is bounded by who actually used it, not by who exists in the realm. (The sibling feature S4U2Self, "protocol transition", removes that bound and lets a keytab mint sessions for arbitrary users. This implementation deliberately does not use it; see [D1].)
The KDC decides, not the service. Step 3 is enforced by the KDC, so it keeps
holding even if the service host is fully compromised. It is the only control in
this design that survives that, which is why it is worth configuring properly
rather than relying on the server's own MCP_DELEGATION_TARGETS list.
In the KDC log a successful one looks like this, and it is worth knowing the shape because it is the fastest way to confirm the whole chain works:
HTTP/mcp.example.internal for HTTP/git.example.internal
... CONSTRAINED-DELEGATION s4u-client=stergios@EXAMPLE.INTERNAL
Three things have to line up, and the KDC is the only one this repository cannot arrange for you.
First, authorise the delegation in FreeIPA, as a realm admin. It is two objects, not one, and the split trips people up: a target is a reusable named list of services that may be delegated to, and a rule says which service may use that list. Read a configured pair as one sentence: this service may act as a caller towards these destinations, and nothing else.
Neither object is exposed in the FreeIPA web UI. There is no Service Delegation
page; ipa or LDAP are the only ways to see or change them.
ipa servicedelegationtarget-add mcp-targets
ipa servicedelegationtarget-add-member mcp-targets \
--principals=HTTP/ci.example.internal@EXAMPLE.INTERNAL
ipa servicedelegationrule-add mcp-delegation
ipa servicedelegationrule-add-member mcp-delegation \
--principals=HTTP/mcp.example.internal@EXAMPLE.INTERNAL
ipa servicedelegationrule-add-target mcp-delegation \
--servicedelegationtargets=mcp-targets
Repeat the flag for each value (--principals=a --principals=b). A
comma-separated list is accepted and then silently adds nothing: ipa reports
success with the rejects in a failed field most callers never read, and the
first symptom is a runtime KDC_ERR_BADOPTION that _explain() cannot
distinguish from a missing rule.
Verify with ipa servicedelegationrule-show mcp-delegation, as a realm admin.
That emphasis is the single most expensive thing on this page to get wrong.
Reading these objects needs the System: Read Service Delegations permission,
carried by the Service Administrators privilege, which an ordinary user does
not hold. LDAP ACIs hide entries rather than refusing them, so an unprivileged
ipa servicedelegationrule-find returns:
------------------------------
0 service delegation rules matched
------------------------------
on a realm where delegation is fully configured and working. No error, no warning, nothing to suggest the answer was filtered. It is indistinguishable from a realm where nothing has ever been set up.
This has bitten in practice, on this codebase, and it cost real time: an audit run as an ordinary user concluded the KDC-side allowlist did not exist, and only a direct LDAP query as directory manager showed two rules quietly doing their job. It applies to a host principal too, so checking from the MCP host is equally useless.
Two things follow. Audit delegation as admin or over LDAP, never as
yourself. And if you are the person who will later have to verify this, grant
your own account the Service Administrators privilege now, while you still
remember these objects exist. They are invisible in the web UI, so there is
nothing to stumble across that would remind you.
To see them regardless of IPA permissions, from the IPA server itself:
ldapsearch -LLL -Y EXTERNAL -H ldapi://%2frun%2fslapd-EXAMPLE-INTERNAL.socket \
-b cn=s4u2proxy,cn=etc,dc=example,dc=internal \
"(objectClass=*)" cn memberPrincipal ipaAllowedTarget
Second, callers need forwardable tickets. Without protocol transition the KDC
hard-requires it. On Windows that is setup.ps1 -Forwardable; elsewhere it is
forwardable = true in krb5.conf. A non-forwardable caller is refused with the
same opaque error as a missing rule.
Third, turn it on in site.env: MCP_DELEGATION=1 plus a
MCP_DELEGATION_TARGETS row per forwarding tool. run.sh validates the grammar,
refuses a target naming a tool that does not call forward_header(), and warns
that the keytab is now usable outbound.
server/install/site.env.example is the single source of site values (domain,
realm, KDC, MCP URL, CA hash, delegation toggles). Copy it to
/etc/mcp-server/site.env, fill it in, and keep it out of git; the installer reads
site values from there and nowhere else.
MCP_SITE_TOOLS points at a Python file loaded at startup, after the shipped
stubs and before the ASGI app is built. It defines one function:
def register(mcp, require, forward_header, register_tool_policy):
@mcp.tool()
def list_tickets(ctx: Context) -> str:
"""List the caller's tickets."""
p = require(ctx, 'list_tickets') # authorize first, always
h = forward_header(ctx, 'list_tickets') # optional: act as the caller
...
register_tool_policy('list_tickets', {'support-staff'})
Keep that file outside the deployed code directory. run.sh converges that
directory on this repository's file set, so anything left beside the shipped
modules is removed on the next deploy, and your tools would go with it.
/etc/mcp-server/site_tools.py is the natural home: the installer owns that
directory and never prunes it. Own it root:root 0644, the same as the code, and
keep it in whatever repository holds your site configuration.
is fail-loud. A path that is set but unreadable, unloadable, or missing
register() stops the server at startup rather than quietly serving a tool set
that lost half its entries. Two limits worth knowing: the invariant test in
tests/python/ parses mcp_server.py only, so it does not check a site tool's
require() wiring, and delegation targets for site tools still come from
MCP_DELEGATION_TARGETS like any other.
server/ # the MCP server (official SDK)
spnego_auth.py - hardened Kerberos acceptor (fixes [C1] by construction)
spnego_asgi.py - self-contained SPNEGO ASGI auth middleware (wraps the SDK app)
mcp_server.py - FastMCP server (stateless) + tools; wires in authz + audit
authz.py - security-owned per-tool IPA-group policy + SSSD group lookup
authz_editor.py - optional, disabled-by-default browser editor for that policy
delegation.py - security-owned on-behalf-of forwarding policy; off by default
requirements.txt - server-only deps (mcp, uvicorn); the bridge stays stdlib
requirements.lock.txt - the ==-pinned tree actually validated; install from this
install/ # everything about getting it running, kept apart from what runs
run.sh - the installer: account, venv, code, keytab, unit, vhost, cert, and
the client bundle (served at /client/ by default, or --client-export DIR)
verify.sh - read-only post-install verifier (last step of the install), 14 checks
site.env.example - the single source of site values; copy, fill, keep out of git
mcp-server.service, nginx-mcp.nginx
client/ # everything that runs on a workstation
setup.sh - Linux: enroll in FreeIPA, then install the MCP client
setup.ps1 - Windows: WSL2 Kerberos SSH, VS Code Remote-SSH, Firefox-in-WSL, MCP bridge
install-bridge.sh - install the client (downloaded over HTTPS, then run; never piped to a shell)
JsoncEdit.ps1 - helper used by setup.ps1 to edit JSONC config in place
README.md - provisioning a Linux or Windows workstation, end to end
bridge/ # what install-bridge.sh installs on a workstation
mcp-krb-bridge.py - the bridge itself (stdlib + python3-gssapi)
mcp-krb-remote-bridge.py - the inert half, for a host that holds no ticket
krb-fetch - fetch one URL byte-exact, over whichever of the two applies
mcp-fetch - krb-fetch's former name: a shim that runs it, kept so the name still works
krb-git - git through the same choice: Negotiate itself on a workstation, relayed on a shared host
examples/ - mcp.json, mcp.json.windows, managed-mcp.json
tests/ # hermetic unit tests (fake gssapi, no KDC needed)
run-tests.sh, python/
docs/ # the images this README shows; nothing here is installed anywhere
The installer directory is server/install/. If you find a doc or a script
referring to server/deploy/, it is stale.
The full threat model, RFC compliance mapping, ranked findings, CVE inventory and deployment checklist live in SECURITY.md. Read it before deploying.
Headline posture: authentication is offline SPNEGO/Kerberos on every request, the replay cache stays on, authorization is per-tool deny-by-default by IPA group, and the only long-lived secret is a keytab readable by the service account and by root. The finding to read before deploying is [SC1], which is not fixed: client distribution rests on HTTPS plus a CA pin with no signature, so a compromised publisher can serve anything, and whoever controls those bytes runs code as root on every workstation that installs them. It is an accepted risk with a named upgrade path, and the docs treat it as accepted rather than solved.
- SECURITY.md : the security review. RFC compliance mapping, ranked findings, CVE inventory, and the deployment checklist. Read before production.
- client/README.md : provisioning a Linux, Windows, or macOS
workstation end to end, including non-domain-joined Windows via WSL2 (Kerberos SSH
- VS Code) and the trust model for the client kit.
Run the hermetic unit tests with sh tests/run-tests.sh: they pass on Windows and
Linux with no native packages, no KDC and no MCP SDK, using a fake gssapi, and
include a source-level check that every @mcp.tool is wired to the authorization
policy and that a tool which forwards names itself. Security-owned paths (the
acceptor, the middleware, authz.py, delegation.py) require review under
.github/CODEOWNERS; a fork should point that file at its own
reviewers. Report vulnerabilities per
SECURITY.md, not in a public issue.
Apache-2.0, copyright Overpass Connect. See NOTICE.