Business User Harness — Wiseway Knowledge Assistant
Our call is Claude Desktop as the day-one chat surface — polished, zero-effort, MCP-native. The realistic swap-outs if priorities shift are Goose (vendor-neutral hedge, Linux Foundation governance) and LibreChat (self-hosted in your Azure AU East tenant, full data residency). This page lays the three options side-by-side on capability, vendor-published pricing, and the two ways each one reaches your Microsoft 365 data — so you can verify the call before signing anything.
01 — Options
Three chat apps worth taking seriously. Each wins a different argument. Claude Desktop · Goose · LibreChat — strengths, governance, and where each fits.
Anthropic's first-party desktop app. Hosted SaaS, Mac and Windows, consumer-grade chat UX, SAML SSO and audit logs on the Enterprise tier. Anthropic co-authored MCP, so every Wiseway MCP server plugs in without bespoke work.
Best when: you want staff productive in week one with zero infrastructure to operate.
Open-source desktop agent donated by Block to the Linux Foundation's Agentic AI Foundation in late 2025. Apache 2.0, model-agnostic across 15+ providers, deepest MCP extension ecosystem of the three. Runs entirely on each staff laptop.
Best when: vendor neutrality matters more than polish — the credibility hedge.
Independent open-source web app, MIT licence, ~24k GitHub stars. Self-hosted Docker deploy with Azure AD SSO, per-user RBAC, and your choice of model providers. Runs in your Azure AU East tenant — nothing leaves unless you point it out.
Best when: AU data residency becomes a hard contractual requirement.
02 — Capability
Fifteen decision dimensions, side by side. Governance, residency, identity, model-agnosticism, MCP, licensing — sourced and dated.
| Dimension | Claude DesktopAnthropic — hosted SaaS | GooseLinux Foundation (AAIF) — open source | LibreChatIndependent project — open source |
|---|---|---|---|
| Maintainer & governance | Anthropic — US company, commercial vendor with paid contracts and SLAs available on Enterprise. | AAIF — donated by Block to the Linux Foundation's Agentic AI Foundation in late 2025. Vendor-neutral. | Danny Avila + community — independent project, ~24k GitHub stars, MIT licence. |
| Where it runs | Native desktop app (Mac/Win) talking to Anthropic-hosted infrastructure. | Native desktop app (Mac/Win/Linux) running on each staff laptop. | Self-hosted web app. You run the server — Docker on Azure AU East, your tenant, your data centre. |
| Model-agnostic | No — Claude models only.Locked to Anthropic as the model vendor. | Yes — 15+ providers.Anthropic, OpenAI, Azure OpenAI, Bedrock, Gemini, Ollama (local), OpenRouter. | Yes — 10+ providers.Anthropic, OpenAI, Google Vertex, Bedrock, Mistral, Ollama, any OpenAI-compatible endpoint. |
| Australian data residency | Not today — data stored in the US. Sydney office announced Mar 2026, no GA timeline. M365 connector explicitly excluded from Microsoft's AU in-country processing. | Depends on the model — point Goose at Azure OpenAI Sydney or local Ollama and your data stays in AU. Goose itself runs on the laptop. | Yes — wherever you host it. Run it in Azure AU East and inference stays in your tenant. |
| Enterprise admin (SSO · RBAC · audit) |
Mature — SAML SSO, SCIM, audit logs, Compliance API on Enterprise tier. Org-wide Skills mgmt shipped Dec 2025. | None today — no SSO, no central admin console, no MDM-packaged installer. Config is per-laptop env-vars + an Extension Allowlist. | Built-in — LDAP / OIDC / Azure AD SSO, per-user model permissions, multi-user RBAC, audit-able conversation logs. |
| Open source | No — proprietary client and proprietary model. | Yes — Apache 2.0. No vendor can revoke your right to run it. | Yes — MIT. Permissive licence; fork and modify without obligation. |
| Vendor lock-in how hard to swap out |
Medium. Your IP (Skills, MCP servers, SharePoint policies) is portable. The surface — Claude Desktop itself — is not. If Anthropic raises prices or has an outage, the chat app stops; the MCP backend keeps working. | Low. Open licence, foundation-governed, brings its own provider list. The exit ramp is the product. | Low. MIT, no vendor, you control the deploy. Switching providers is a settings change. |
| Software cost | $20-$100/seat/mo.Pro $20 · Team Standard $20 · Team Premium $100 · Enterprise self-serve $20 + tokens at API rates. See pricing below. | $0.You still pay for inference through whichever model API key you configure. No per-seat licence. | $0 for the software.You pay for inference (your API key) + hosting (~$30-100/mo on Azure App Service AU East for ~300 staff). |
| MCP support the standard your platform is built on |
First-class — Anthropic co-authored MCP. Remote MCP, OAuth, Streamable HTTP all supported. Custom connectors live in Settings. | First-class — Block co-authored MCP. 70+ extensions out of the box. Extension Allowlist for IT control. | Listed MCP client on the official MCP registry. Configurable via librechat.yaml or Smithery. |
| M365 / SharePoint connector | First-party, GA Apr 2026.Anthropic-hosted connector. Caveat: searches the whole tenant scoped by user permissions — no site-specific scoping. | Via MCP servers.Use Microsoft's reference SharePoint MCP, or the one Wynne AI is building for this engagement. | Via MCP servers.Same MCP server works across all three — the integration isn't trapped in any one chat app. |
| Fit for non-technical staff warehouse · forwarders · ops |
Day-one usable — consumer-grade chat UX, file uploads, voice on mobile. Zero training needed. | Power-user flavour — Recipes are YAML files, UX still feels engineer-built. Block uses it internally on non-engineers but external evidence is thin. | Web UI is fine — looks similar to ChatGPT. But onboarding 300 staff is a self-serve account creation flow you operate. |
| Business features that the Knowledge Assistant needs |
Citations · file upload (30 MB / 20 files per conversation) · web fetch · Skills · conversation history · mobile apps · email summarisation · Teams not native. | Recipes (parameterised workflows) · file ops · code execution · screen interaction · MCP everything · no mobile app · no email integration. | Document chat · agents · RAG built-in · code interpreter · multi-modal · per-user model permissions · web only (no native mobile). |
| Vendor support & SLA | Commercial — paid Enterprise tier includes account team and SLA. There is someone to call. | Community only — Linux Foundation governance, GitHub issues, no paid support contract from AAIF. | Community only — independent project, Discord + GitHub. No paid support contract. |
Status colours: green = clear strength, amber = workable with caveats, red = real gap for Wiseway today.
03 — Verdicts
The table told you what. This tells you so what? Strengths, weaknesses, and where each tool fits in Wiseway's stack.
Option 1 — hosted
Strengths
Weaknesses
Use as: the day-one Knowledge Assistant surface. Polished enough that you don't need a change-management programme to roll it out.
Option 2 — open desktop
Strengths
Weaknesses
Use as: the credibility hedge. The thing you point to when the board asks "what if Anthropic doubles their prices?" Pilot on a tech-savvy team (finance, data) before considering wider rollout.
Option 3 — self-hosted web
Strengths
Weaknesses
Use as: the sovereign option. The day AU data residency becomes a hard contractual requirement, LibreChat is what you switch the Knowledge Assistant to without rebuilding the platform.
04 — Pricing
Vendor-published prices. Every figure sourced and dated. Per-seat and per-token, hosted and self-hosted — monthly cost models for ~300 staff.
Anthropic — Claude plans
Source for every row: claude.com/pricing and linked Anthropic help centre articles.
| Plan | Per seat (annual) | Monthly equivalent | Tokens | Min seats |
|---|---|---|---|---|
| Claude Proclaude.com/pricing | $20/seat/mo | — | Includedcaps apply | 1 |
| Claude Team — Standardclaude.com/pricing | $20/seat/mo | $25/seat/mo | Included"more than Pro" | 5 |
| Claude Team — Premiumclaude.com/pricing | $100/seat/mo | $125/seat/mo | Included5× Standard | 5 |
| Claude Enterprise — self-servesupport.claude.com — Enterprise plan | $20/seat/mo | annual only | Billed separatelyat API rates → below | 20 |
| Claude Enterprise — sales-assistedclaude.com/contact-sales | contact sales | — | Billed separately | 50 |
Anthropic does not publish exact per-user message caps for Pro/Team in their pricing pages — only that they are "higher than Free" and that Premium gives "5× more usage" than Standard. See usage limit best practices.
Read carefully
Same headline number on claude.com/pricing, different shape underneath.
$20/seat/mo
Source: claude.com/pricing
$20/seat/mo + tokens
Anthropic — API token rates
Direct from platform.claude.com — pricing docs. These rates apply on Enterprise self-serve and on direct API use.
| Model | Input | 5-min cache write | 1-hr cache write | Cache read (hit) | Output |
|---|---|---|---|---|---|
| Claude Opus 4.7 | $5/MTok | $6.25 | $10 | $0.50 | $25/MTok |
| Claude Sonnet 4.6 | $3/MTok | $3.75 | $6 | $0.30 | $15/MTok |
| Claude Haiku 4.5 | $1/MTok | $1.25 | $2 | $0.10 | $5/MTok |
Batch API offers a flat 50% discount on input and output for asynchronous jobs. See pricing docs for full matrix including Opus 4.1, Sonnet 4.5, Haiku 3.5 and batch rates.
Other vendors
Where the vendor publishes a number on their public pricing page, it's shown here. Where they don't, the row links to their sales enquiry page.
| Plan | Per seat (annual) | Per seat (monthly) | Notes |
|---|---|---|---|
| Microsoft 365 Copilot Chatmicrosoft.com — Copilot pricing | Included | Included | Requires eligible M365 subscription |
| Microsoft 365 Copilot Businessmicrosoft.com — Copilot pricing | $18/seat/mopromo until 30 Jun 2026 | $25.20/seat/mo | Standard rate $21/seat after promo. Up to 300 users. |
| ChatGPT Businesschatgpt.com/pricing · help.openai.com | see vendor page | see vendor page | OpenAI's pricing page blocks programmatic fetches — confirm in browser before quoting. |
| ChatGPT Enterpriseopenai.com/chatgpt/enterprise | contact sales | contact sales | No per-seat price published by OpenAI. |
| Gemini Enterprisecloud.google.com/gemini-enterprise | contact sales | contact sales | Google's product page directs all pricing enquiries to sales. |
Open-source clients
These projects publish their code under open licences. The client costs nothing; the inference still costs whatever the model vendor charges through your API key — see API rates above for Anthropic.
| Project | Software cost | Licence | Form factor | MCP & model providers |
|---|---|---|---|---|
| Goosegithub.com/block/goose | Free | Apache-2.0 | Native desktopmacOS / Linux / Windows | MCP native (70+ extensions). 15+ providers: Anthropic, OpenAI, Google, Bedrock, Ollama, OpenRouter. BYO key or existing subscription. |
| LibreChatgithub.com/danny-avila/LibreChat | Free | MIT | Self-hosted webDocker / Railway / Sealos | Listed as an official MCP client on modelcontextprotocol.io. Anthropic, OpenAI, Google Vertex, Bedrock, Mistral, DeepSeek, Ollama. SSO via Azure AD / Cognito / OAuth. |
| Cherry Studiogithub.com/CherryHQ/cherry-studio | Free | AGPL-3.0commercial licence on request | Native desktopmacOS / Linux / Windows | MCP server support. OpenAI, Gemini, Anthropic, plus Ollama / LM Studio for local models. |
| Open WebUIgithub.com/open-webui/open-webui | Free | Open WebUIbranding-preservation clause | Self-hosted webDocker | MCP Registry built in. Ollama (primary), any OpenAI-compatible API, Mistral, Groq, OpenRouter. |
| AnythingLLMgithub.com/Mintplex-Labs/anything-llm | Free | MIT | Desktop or self-hosted | MCP-compatible. Wide provider list incl. Anthropic, OpenAI, Bedrock, Ollama. RAG built-in. |
AGPL-3.0 (Cherry Studio) requires that derivative works also be open-sourced under AGPL — relevant if you fork it. MIT and Apache-2.0 don't have that requirement.
Build it in-house?
What it would actually take, and why we don't recommend it for Wiseway today.
1 senior engineer
~4 months to MVP, then ~1 day/week to maintain
Plus, on call
Pros
Cons
05 — Microsoft 365
Two ways to plug Claude into your M365 tenant — and the trade-offs each forces. First-party connector vs self-hosted MCP — residency, read vs read+write, what each costs to operate.
How each one is set up
Both options talk to your SharePoint, OneDrive, Outlook and Teams data through Microsoft Graph — but only one of them can write back. Anthropic's connector is read-only retrieval; the self-hosted MCP exposes the full Graph CRUD surface, so Claude can edit documents, send mail and update calendars when you grant it those scopes. What also changes is who runs the server that calls Graph — and therefore where the queries travel and which clients can use it.
Side-by-side comparison
Decision-relevant axes only. The headline difference is the second row — read-only retrieval versus full read & write against your tenant. Sources for every claim are linked in section 07.
| Dimension | Official Anthropic connector M365 Connector for Claude | Self-hosted community MCP e.g. Softeria ms-365-mcp-server |
|---|---|---|
| Built & maintained by | Anthropic, first-party.App registrations and server operated by Anthropic. | Open-source community.Softeria (~720 stars, MIT) is the production-shaped option; alternatives: Lokka, PnP CLI MCP. |
| Document & data operations | Read-only — search, retrieve, cite. Surfaces SharePoint, OneDrive, Outlook and Teams content into the conversation. Cannot create, edit, rename or delete files, send mail, create calendar events, or post Teams messages — it has no write path back into the tenant. | Read & write — full Graph surface. Exposes the underlying Microsoft Graph endpoints as MCP tools: upload & overwrite SharePoint / OneDrive files, edit Excel ranges, draft and send Outlook mail, create / move / delete events, post Teams chats. Scope of what Claude can change is governed by the Graph permissions you grant the app registration — narrow them to keep the blast radius small. |
| Security posture & prompt-injection blast radius | Smaller blast radius — read-only by construction. Tokens are stored on Anthropic infrastructure; you outsource operational security and patch cadence to Anthropic (SOC 2 Type 2 attested). A prompt-injection payload hidden inside a SharePoint document can still mislead Claude's answer or coax it into leaking content already visible in the conversation — that's a property of LLMs, not the connector — but it has no write tools to abuse, so it cannot cause Claude to email, overwrite or delete anything in your tenant. Trade-off: source is closed, you cannot independently audit the server. | Auditable, but you own the blast radius. Open source is a security positive: independent review, pinnable versions, faster supply-chain triage, tokens stay inside your Azure tenant. The risk shifts to what an injection can do once it lands: because write tools are exposed, a poisoned document could instruct Claude to email a file externally, overwrite a SharePoint page, or create calendar invites — if the Graph scopes allow it. Mitigations: grant least-privilege scopes (Sites.Selected, Files.ReadWrite.Selected, omit Mail.Send unless explicitly required), keep destructive tools behind per-call user confirmation in the client, and own the upstream CVE-watching yourself. |
| Where it runs | Anthropic infrastructure, US.Path: Claude → Anthropic backend → Anthropic-hosted MCP → Graph. | Your Azure tenant, your choice of region.Container Apps in AU East keeps the query hop in Australia. |
| Source code | Closed — you cannot inspect or modify the server. | Open source — MIT licence; fork, audit, or modify. |
| Client compatibility | Claude Desktop / claude.ai only.Locked to Anthropic clients — Goose or LibreChat cannot reach it. | Any MCP client.Same server speaks to Claude Desktop today, Goose or LibreChat tomorrow — protocol-level interop, no rewrite. |
| Setup effort | Minutes — Entra Global Admin grants tenant-wide consent once; users complete per-user OAuth. | 3–5 days initial deploy — Entra app registration, container build, scope tuning, runbook. |
| Ongoing maintenance | None — Anthropic operates it.SLA included on Enterprise tier. | ~0.5–1 engineer-day / month for patches and upstream updates. |
| Cost | $0 — included with every Claude plan.Free / Pro / Team / Enterprise all include it. | ~AU$50 / month Azure Container Apps + Log Analytics, idle workload. |
| Australian data residency | No — queries traverse US.Explicitly excluded from Microsoft's in-country processing commitments for Anthropic-routed traffic. | Yes — deploy in AU East and the proxy hop stays in-country.Inherits Microsoft's AU residency on the Graph side. |
| Tenant scoping | Tenant-wide only.No site-specific restriction inside the connector — limit via SharePoint permissions only. | Site-scoped possible.Restrict the MCP server's Graph scopes so it can only reach specific sites or libraries. |
| Retrieval quality | High — tuned ranking, OCR on images, document chunking, multi-modal extraction. | Lower out of the box — raw Microsoft Graph search. Add OCR / chunking yourself if needed. |
| Citations & rendering | Polished inline citations + file previews rendered by Claude. | JSON responses; whichever client (Claude Desktop, Goose, LibreChat) decides how to render them. |
Status colours: green = strength, amber = workable with caveats, red = real gap.
06 — Brief
Your brief, line by line — answered against all three options. Direct quotes from sections 3 and 6 of the Wiseway AI Brief, mapped to verdicts.
| What you said in the brief | Claude Desktop | Goose | LibreChat |
|---|---|---|---|
| "Work well with our existing Microsoft 365 and SharePoint setup" | Strong — first-party M365 connector. | OK — via MCP server. | OK — via MCP server. |
| "Keep our data in Australia" | Gap — US-hosted today. | If you choose AU models. | Strong — you host where you like. |
| "Don't lock us into a single vendor or model provider" | Locked to Anthropic for the surface. | Strong — model-agnostic, foundation-governed. | Strong — model-agnostic, MIT. |
| "Can be extended to future use cases without rebuilding" | Strong — Skills + MCP designed for it. | Strong — Recipes + MCP designed for it. | Strong — Agents + MCP, fully open. |
| "We want to own the outcome — maintain and extend it ourselves" | Partial — your Skills/MCP repos are yours; the client is Anthropic's. | Strong — Apache 2.0, you own everything. | Strong — MIT, you own everything. |
| "One platform, not three silos" | Works — same skills + MCP servers reused by your other assistants. | Works — same skills + MCP servers reused. | Works — same skills + MCP servers reused. |
| "Pragmatic delivery — pilot small, prove value, expand" | Best — buy 5-10 seats, pilot in a week, no infrastructure. | Workable — install on a few laptops, but no central admin. | Workable — stand up one Azure deploy, but it's real infrastructure work. |
07 — Sources
Every claim, every link. Sourced and dated. Vendor pricing pages, support articles, GitHub repos — verify any figure on the page.