In Part 1 I put Codex CLI behind open source agentgateway so it could use an OpenAI model on Amazon Bedrock without holding a key. That left one obvious hole: the gateway took requests from anyone who could reach the port.
This post closes it. When I type codex now, it sends me to the Microsoft sign-in page first. The gateway only accepts my requests if I come back with an Entra ID token and I am in the codex group.
What you end up with:
codex group, or the gateway answers 403./v1 (Bedrock) and /mcp (the Solo.io docs search server).
All of it is open source agentgateway v1.7.0 and Codex CLI 0.170.0. The full step list with every command is in the repo: sebbycorp/agentgateway-oss-k8s, entra-id/README.md.
flowchart LR
U[You type codex] --> C[Codex CLI]
C -- "1. no sign-in yet: open browser" --> E[Microsoft Entra ID]
E -- "2. token with your group IDs" --> C
C -- "3. Authorization: Bearer token" --> G[agentgateway proxy]
G -- "valid token + codex group" --> B[Amazon Bedrock]
G -- "valid token + codex group" --> M[search.solo.io MCP]
G -. "no token: 401, wrong group: 403" .-> C
Three parts have to agree:
| Part | What it does |
|---|---|
| Entra app registration | Issues tokens for “the gateway”, with the user’s group IDs inside |
| agentgateway policy | Checks the token’s signature, issuer, audience and group |
| Codex profile | Tells Codex where to send the user to sign in |
solo-search MCP route and a Codex profile.
az login
export TENANT_ID=$(az account show --query tenantId -o tsv)
export CTX=kind-agentgateway-codex
export GROUP_ID=$(az ad group create --display-name codex --mail-nickname codex \
--description "Access to LLMs and MCP servers through agentgateway" --query id -o tsv)
az ad group member add --group $GROUP_ID \
--member-id $(az ad signed-in-user show --query id -o tsv)
The app registration stands for the gateway. The gateway accepts only tokens that were issued for this app.
export APP_ID=$(az ad app create --display-name agentgateway-oss-codex \
--sign-in-audience AzureADMyOrg --query appId -o tsv)
OBJECT_ID=$(az ad app show --id $APP_ID --query id -o tsv)
az ad app update --id $APP_ID --identifier-uris "api://$APP_ID" \
--set groupMembershipClaims=SecurityGroup
SCOPE_ID=$(uuidgen | tr 'A-Z' 'a-z')
az rest --method PATCH --uri "https://graph.microsoft.com/v1.0/applications/$OBJECT_ID" \
--headers 'Content-Type=application/json' --body "{
\"api\": {
\"requestedAccessTokenVersion\": 2,
\"oauth2PermissionScopes\": [{
\"id\": \"$SCOPE_ID\", \"value\": \"access\", \"type\": \"User\", \"isEnabled\": true,
\"adminConsentDisplayName\": \"Access agentgateway\",
\"adminConsentDescription\": \"Call LLMs and MCP servers through agentgateway\",
\"userConsentDisplayName\": \"Access agentgateway\",
\"userConsentDescription\": \"Call LLMs and MCP servers through agentgateway\"
}]
}
}"
az rest --method PATCH --uri "https://graph.microsoft.com/v1.0/applications/$OBJECT_ID" \
--headers 'Content-Type=application/json' --body '{
"isFallbackPublicClient": true,
"publicClient": { "redirectUris": ["http://127.0.0.1/callback", "http://localhost/callback"] }
}'
az ad sp create --id $APP_ID
Four settings matter here, and each one fails in its own way if you skip it:
| Setting | If you skip it |
|---|---|
Scope api://<APP_ID>/access |
Codex has nothing to ask a token for |
requestedAccessTokenVersion: 2 |
Sign-in works, but the gateway answers 401 because the issuer is the old v1 one |
groupMembershipClaims: SecurityGroup |
The token has no groups, so the gateway answers 403 |
| Public client redirect URIs | The browser shows AADSTS500113: No reply address is registered |
I hit the last one myself. I ran Codex before adding the redirect URIs and got the AADSTS500113 page.
Two policies, in k8s/06-entra-auth.yaml. The first one attaches to the Gateway and covers everything:
apiVersion: agentgateway.dev/v1alpha1
kind: AgentgatewayPolicy
metadata:
name: entra-auth
namespace: agentgateway-system
spec:
targetRefs:
- group: gateway.networking.k8s.io
kind: Gateway
name: agentgateway-proxy
traffic:
jwtAuthentication: # who are you? 401 if the token is missing or invalid
mode: Strict
providers:
- issuer: https://login.microsoftonline.com/${TENANT_ID}/v2.0
audiences:
- ${APP_ID}
jwks:
remote:
url: https://login.microsoftonline.com/${TENANT_ID}/discovery/v2.0/keys
authorization: # are you allowed? 403 if not in the group
action: Allow
policy:
matchExpressions:
- '"${GROUP_ID}" in jwt.groups'
The second one attaches to the MCP route. It repeats the same checks and adds one block:
mcp:
provider: Entra # bridges the gaps between Entra and the MCP auth spec
clientId: ${APP_ID} # Entra has no dynamic client registration
resourceMetadata:
scopesSupported:
- api://${APP_ID}/access
bearerMethodsSupported:
- header
That block makes the gateway serve the OAuth discovery endpoints an MCP client looks for. Entra does not follow the MCP authorization spec on its own, and agentgateway’s Entra provider fills the gaps.
The MCP route also needs two extra path matches so those discovery requests reach it. Without them the gateway answered route not found on /.well-known/oauth-protected-resource/mcp:
- path:
type: PathPrefix
value: /.well-known/oauth-protected-resource
- path:
type: PathPrefix
value: /.well-known/oauth-authorization-server
Apply both files. The IDs are filled in at apply time, so they stay out of the repo:
kubectl --context $CTX apply -f k8s/05-mcp-solo-search.yaml
sed -e "s/\${TENANT_ID}/$TENANT_ID/g" -e "s/\${APP_ID}/$APP_ID/g" -e "s/\${GROUP_ID}/$GROUP_ID/g" \
k8s/06-entra-auth.yaml | kubectl --context $CTX apply -f -
I tested the gateway with curl first, so I knew the policy worked before adding a second moving part.
| Request | LLM /v1 |
MCP /mcp |
|---|---|---|
| No token | 401 | 401 |
| Garbage token | 401 | 401 |
| Valid Entra token for a different API | 401 | not tested |
| Valid token, group not matching | 403 | 403 |
| Valid token, user in the group | 200 | 200 |
For the “group not matching” row I pointed the policy at a group ID that does not exist. I did not test with a second real user.
Codex CLI has a gateway_oauth table on a model provider. Give it the Entra endpoints and the app’s client ID:
model = "gpt-oss-120b"
model_provider = "agentgateway"
[model_providers.agentgateway]
name = "agentgateway"
base_url = "http://localhost:<port>/v1"
wire_api = "responses"
[model_providers.agentgateway.gateway_oauth]
authorization_url = "https://login.microsoftonline.com/<TENANT_ID>/oauth2/v2.0/authorize"
token_url = "https://login.microsoftonline.com/<TENANT_ID>/oauth2/v2.0/token"
client_id = "<APP_ID>"
scopes = ["api://<APP_ID>/access", "offline_access"]
[mcp_servers.solo-search]
url = "http://localhost:<port>/mcp"
Save it as ~/.codex/agentgateway.config.toml. There is no token and no secret in the file. If you added the two blocks that make Claude models work in the previous post, keep them at the end of the file.
Now start Codex:
codex --profile agentgateway
Codex opens the Microsoft page. If it cannot open a browser, it prints the link:
Authorize the model provider by opening this URL:
https://login.microsoftonline.com/<TENANT_ID>/oauth2/v2.0/authorize?response_type=code&client_id=...
After the sign-in Codex continues, and it keeps the token encrypted under ~/.codex/secrets/. It refreshes the token by itself, so the next run does not ask again.
The MCP server does not reuse the model sign-in. It has its own, and this is the part that took the longest to get right.
Codex adds an ID to the end of its MCP callback path, and Entra matches redirect paths exactly. So you have to find that ID and register it. Print the path:
codex --profile agentgateway mcp login solo-search --no-browser </dev/null 2>&1 \
| grep -oE 'redirect_uri=[^&]+' | python3 -c "import sys, urllib.parse as u; print(u.unquote(sys.stdin.read()))"
Register it on the app, without the port:
CALLBACK_ID=dFBqJtJnaDT6 # yours will differ
az rest --method PATCH --uri "https://graph.microsoft.com/v1.0/applications/$OBJECT_ID" \
--headers 'Content-Type=application/json' --body "{
\"publicClient\": { \"redirectUris\": [
\"http://127.0.0.1/callback\", \"http://localhost/callback\",
\"http://127.0.0.1/callback/$CALLBACK_ID\"
] }
}"
Then sign in:
codex --profile agentgateway mcp login solo-search
The ID is derived from the MCP server URL. It stayed the same when I renamed the server and changed when I changed the host in the URL. So if your gateway address changes, you redo this step.
echo 'alias codex="codex --profile agentgateway"' >> ~/.zshrc
source ~/.zshrc
I used an alias instead of editing ~/.codex/config.toml, because that file is shared with the ChatGPT desktop app.
To prove the sign-in is enforced, I moved the stored credential away and started Codex. It stopped and printed the sign-in link again. With the credential back, it ran:
codex exec --skip-git-repo-check \
"Use the solo-search MCP search tool to look up: agentgateway. Reply with the top result's title only."
Out of the box the request log identifies the user by the token’s sub claim, which is an opaque string. One setting changes it to the sign-in name:
kubectl --context $CTX -n agentgateway-system patch agentgatewayparameters agentgateway-proxy-params \
--type merge -p '{"spec":{"rawConfig":{"config":{"standardAttributes":{"user":"jwt.preferred_username"}}}}}'
I also turned on prompt logging, which stores the prompt and the answer of every call. On Kubernetes the checkbox for it in the UI is read-only, so it goes in the proxy parameters:
spec:
rawConfig:
frontendPolicies:
accessLog:
database:
llm: full
The Logs page now shows the prompt on each row:
And the user, the group and the whole conversation when you open one (I blurred the user and group here):
Think before you turn prompt logging on for a team. Every prompt lands in Postgres in plain text, next to the name of the person who typed it.
az account get-access-token as a token command. It worked, but it silently reused my Azure CLI session and never asked me to sign in. The built-in gateway_oauth flow is the one that prompts.codex mcp list shows Not logged in until you run codex mcp login. http://127.0.0.1/callback covers any port for the model sign-in. The MCP callback has an extra path segment and needs its own entry.groups claim for such users, and the policy answers 403. The fix is to emit only groups assigned to the application.gpt-oss-20b and claude-haiku-4-5 next to gpt-oss-120b. All three sit behind the same policy, and codex -m claude-haiku-4-5 uses the same sign-in.
The repo has this as a 10-step guide with every command, a troubleshooting table and a script for the Entra setup: entra-id/README.md.
Related posts: