Quotations MCP Server

Standalone MCP server exposing the quotations-app backend as tools. It's a thin HTTP client over the quotations REST API — no database of its own — so it stays decoupled and can point at local dev or the live Render backend.

MCP agent ──MCP──► server.py ──HTTP──► quotations-app backend ──► MongoDB

Other services in this repo

This is one of three independently-deployable pieces that happen to share a git repo — no shared state, no shared process:

Related MCP server: finance-mcp

oscar_ai_bot_legacy/ — the "Oscar AI" DM bot (moved 2026-09-26)

What it was: a second, focused RFQ path inside the Oscar backend (AlumnxAILabs_epa), separate from the automatic background poller (rfq_service/ above). A team member could DM a special "Oscar AI" account inside the Oscar chat app and ask "any new RFQ emails?" — the bot would check the mailbox on demand and reply in-chat with quotation details, instead of waiting for the background worker's own polling cycle.

Why it's here now: all RFQ-processing code was consolidated into this repo. This folder holds every file that on-demand path depended on:

⚠️ NOT currently runnable standalone. email_processing_service.py imports Oscar-core services (item_service, comment_service, notification_service) that only exist inside the Oscar backend's own Python path — those aren't in this repo. To bring the DM bot back to life here, one of two things has to happen first:

1. Rewire it to call Oscar's /internal/* API (the same pattern rfq_service/oscar_client.py in this repo already uses for the background worker) instead of importing item_service/comment_service directly — i.e. make email_processing_service.py's task-creation step an HTTP call, not a local import.

2. Or run it co-located with a copy of the Oscar backend's models/services package on the Python path (not recommended — reintroduces the tight coupling this move was meant to remove).

Automatic assignment behavior (unchanged from when it lived in Oscar): every quotation this pipeline creates is auto-assigned to the team lead of whichever team the asking user belongs to (resolve_team_lead in email_processing_service.py) — not to the asker, and not to a fixed person. A task carrying the quotation link is created and assigned the same way. If the asker has no team, it falls back to RFQ_TEAM_LEAD_USER_ID, then the owner of RFQ_ORG_TEAM_ID.

Config this path needs (same names as rfq_service/'s own env vars — see that folder's .env.example): GMAIL_* (OAuth creds for the shared inbox), OPENAI_API_KEY, QUOTATION_API_URL / QUOTATION_FRONTEND_URL, RFQ_TEAM_LEAD_USER_ID, RFQ_ORG_TEAM_ID, RFQ_TASK_DUE_MINUTES / RFQ_TASK_DUE_HOURS, OSCAR_AI_BOT_ID (the provisioned bot account's user id in Oscar's own database).

As of this move, the "Oscar AI" DM feature is OFF inside the live Oscar app — the code that triggered it (main.py's DM-interception hook) has been removed there. Reviving it means finishing option 1 or 2 above, then re-adding an equivalent trigger wherever DMs are handled going forward.

Tools

Purpose is conversational intelligence — let an agent answer natural-language questions about quotations (counts, values, who-has-what, lookups) and take actions.

Conversational / query (read):

READ-ONLY by design. This MCP only queries the customer's data and answers questions — it never writes (no create/reassign). Creating or modifying quotations is the customer app's job. The backend exposes only list + get, so filtering/aggregation is done here; prices/totals come only from the server-computed data — never invented.

Completion / status

A quotation is "completed" only if the customer's data stores that state. Today their backend hard-codes status:"draft" and never changes it, so quotation_stats / search_quotations(status=...) will only ever report draft. The moment their app writes a real status (e.g. completed/sent), these read tools surface it automatically — no change here.

Install

pip install -r requirements.txt   # mcp[cli], httpx

Run

# stdio (dev / desktop MCP clients)
python server.py

# HTTP endpoint (for the MCP agent to connect to)
MCP_TRANSPORT=http MCP_PORT=8200 python server.py
# → MCP endpoint at http://<host>:8200/mcp

# inspect/try tools
mcp dev server.py

Env

Notes

• Deployed separately from the quotations backend and from Oscar.

• The Oscar agent will later connect to this server (as an MCP client) via a new endpoint — this repo/folder does not depend on Oscar.