Integration of mcp Server with BYOT

[ervet]

Hello,

has someone integrated the https://github.com/Softeria/ms-365-mcp-server into LibreChat already and used the already present OAuth Tokens for this.

Or have anybody have different M365 MCP Server used in Librechat, expecially for Outlook.

Thx
Erik

[danny-avila]

I think @maxesse has unless it's a different MCP server

#10867

[maxesse]

Yes, that's exactly the one. It works really well, but you do need to do the OBO flow on the Librechat end, as provided by that PR, as the softeria MCP is not gonna do it for you.

[maxesse]

Actually one annoying thing is that with that mcp you'll often run out of context window, as gpt seems quite liberal in requesting 50 top items, without realising it's gonna be like 800k tokens. I'm experimenting with system prompts to make it be a little more conservative in how many items it requests.

[ervet]

Wow @maxesse thx for the implementation ! I would love to buy you a coffee for that :-) !!

Do you maybe have a working example for the librechat.yaml. I was thinking on how to do that. Starting like MS365_MCP_OAUTH_TOKEN={{TOKEN?}} @softeria/ms-365-mcp-server seems the way to go - at least from the documentation from softeria mcp ...

thx again.

[maxesse]

Yes, sure, it's all pretty straightforward tbh, since you don't need to configure the oauth flow in softeria m365 at all.
In librechat.yaml this'll do (plus the PR I pushed as that variable won't exist until it's in your librechat build):

  Microsoft365:
    type: streamable-http
    url: http://mcp-m365.k8s-internal.local/mcp
    timeout: 45000
    initTimeout: 150000
    startup: true
    headers:
      Authorization: "Bearer {{LIBRECHAT_GRAPH_ACCESS_TOKEN}}"
    serverInstructions: "Use M365 tools to interact with Microsoft 365 services including Outlook, OneDrive, SharePoint, Teams, and other Microsoft Graph APIs."

Then for the actual MCP server deployment, I deployed it like this - I am building an image weekly from their repo since they don't offer pre-built docker images. Using a base alpine image + npx their mcp wasn't great, at least i don't really like that it gets pulled at container startup, it's slower and I'd rather like containers to be immutable.

I added a couple of comments to the startup arguments. Note how I didn't specify any Entra client ID / secret / anything - all we want is for each and every request to have a valid graph token, while Librechat takes care of refreshing it etc.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: mcp-m365
  namespace: mcp-servers
spec:
  replicas: 2
  selector:
    matchLabels:
      app: mcp-m365
  template:
    metadata:
      labels:
        app: mcp-m365
    spec:
      containers:
        - name: mcp-m365
          image: your.container.registry/ms-365-mcp-server:latest@sha256:pinneddigest
          args:
            - "-v" # Without this you literally won't see a single line of log from the containers, adding verbosity at least tells you if it's working
            - "--http" # This starts it in streamable http mode, which we need
            - "9000" # On port 9000
            - "--org-mode" # This enables all the various teams commands etc. which aren't available in personal mode
            - "--toon" # In theory it 'saves' a few tokens, but it's not that helpful when gpt requests 800k tokens worth of emails to start with
            - "--discovery" # so far this has worked decently for me, instead of expanding all the 90+ tools, it only exposes a master tool and then the model finds out which tools to call subsequently
          ports:
            - containerPort: 9000
          readinessProbe:
            tcpSocket:
              port: 9000
            initialDelaySeconds: 30
            periodSeconds: 5
          livenessProbe:
            tcpSocket:
              port: 9000
            initialDelaySeconds: 60
            periodSeconds: 10
          resources:
            requests:
              memory: "256Mi"
              cpu: "100m"
            limits:
              memory: "512Mi"
              cpu: "500m"

[ervet]

thx @maxesse what a legend ! Thank you for that ... Now I owe you not just a coffee, but rather a beer and a Döner. :-)

Will document my stuff here also but need some days to get this up and running

[leondape]

I'm also desperately looking for a multi-user capable oauth MCP Server for Outlook, wondering why microsoft is not deploying safe remote-MCP Servers.

This is a PSA for the softeria Repo:

Many MCP servers out who claim they support oauth2 are implemented in a stateful manor. It's left to the MCP Server to handle the tokens (!bad!)
MCP Servers should be stateless and leave the token handling to MCP Clients like Librechat (good).

In Summer of 2025 I became so desperate that I vibe coded Oauth2 support bolted on top that softeria repo which is mentioned above. It works nicely in MCP Inspector and any MCP Client that supports oauth as per spec (naturally).

I clearly stated into my PR that it was vibe-coded and needs skillful security review about the whole oauth flow. "Unfortunately" my PR was immediately merged into main, which made me back off far from that repo. Never would I use it in prod, if my prod is a company with multiple users.

Since there still is no better solution around I started doing a "vibe-code-audit" and it found a serious bug: If you use it in oauth mode all tokens are shared by the server so that race-conditions can occur. In single user mode this is not a huge problem, but for organizational deployments it really is. See here: Softeria/ms-365-mcp-server#187

I'm now "vibe-fixing" this but still I don't feel super safe about this repo...

Hoping very much Microsoft will step in and deliver Microsoft audited remote MCPs to close this systemic issue.

[MTR3K]

Hi @leondape, they actually are in the process of deploying some. but of course they will want $$$$$ for it.

See here:
https://learn.microsoft.com/en-us/microsoft-agent-365/mcp-server-reference/mail