Every AI client your team already runs, wired to the same reconciled model.
The MCP server gives developers governed, programmatic access to Rexfin's model and cited actuals: the same permission and citation rules the product enforces in the UI, open to whichever AI client speaks the protocol.
What every response carries
A figure only leaves the server once these checks already agree, not a "trust me" step a developer has to bolt on afterward.
The gateway
One protocol for every client your team already connects.
Instead of a bespoke integration per assistant, the MCP server exposes one governed surface (the model, its definitions, and its cited actuals) to whichever client asks, under whichever permissions the caller was granted.
Grounded by default
Bring your evidence into someone else's chat window.
A query arriving over MCP resolves against the same reconciled model as a query typed into Rexfin directly, the client just changes; the citation and the reconciliation status travel with the number either way.
Nothing gets re-typed and nothing gets re-verified on the way out. The figure the client renders is the figure the model already tied out.
| Region | Actual | Budget | Δ |
|---|---|---|---|
| UAE | 18,220 | 16,900 | +7.8% |
| KSA p.24 | 24,610 | 21,300 | +15.5% |
| Oman | 6,140 | 6,400 | -4.1% |
- · KSA is ahead of budget on trading volume.
- · Oman softened slightly against plan.
One gateway
Stop wiring a fresh integration for every assistant.
Every new AI client teams adopt used to mean another point-to-point connection to babysit. The MCP server replaces that with a single governed endpoint that any compliant client can call the same way.
Admins configure access once (by role, by client, by metric) and every client behind that endpoint inherits the same rules, the same definitions, and the same citations.
Traceable, step by step
Multi-step answers stay accountable the whole way through.
An agent rarely stops at one call. When a client chains a follow-up onto the last answer, each tool call resolves independently against the model, so the third question in a thread is exactly as grounded as the first.
Nothing carries forward as an assumed value; every step re-earns its own citation.
→ get_actual { metric: "revenue", period: "FY25-Q2" }
← 61,930 · cited p.24 · 2 extractors agree
→ compare_variance { baseline: "FY25-Q2-budget" }
← +28.5% · reconciled · assumptions attached
→ explain_variance { figure: "revenue" }
← "Trading volume up 34.9% YoY, see p.24–27."
Governed access
Every scope enforced at the source, not the prompt.
Roles and per-organization isolation apply to a programmatic caller exactly as they apply inside the product. A client scoped to Finance sees Finance figures; nothing wider, whatever it asks.
Admins approve which clients connect and monitor what they query, so visibility doesn't trade away for convenience.
Scoped to Finance · Controllers role
From the team
"We didn't want a second, quieter place for bad numbers to hide. If a client asks the MCP server for something that hasn't reconciled yet, the answer says so; the protocol doesn't get a shortcut the product wouldn't allow."
The Rexfin team
Client session
"What's the KSA revenue variance?"
+15.5% · p.24
One governed hub for every assistant in the building.
New AI clients keep showing up faster than anyone can vet them by hand. Each one plugs into the same reconciled model, the same permissions, and the same citations, never a fresh export for someone to babysit.
Further reading
How the trust chain holds up end to end
Filed, extracted, reconciled, cited, exported: walk the five steps every figure passes through before it reaches a client.
Read the walkthroughMeet the Analyst agent
The same evidence the MCP server exposes to external clients, asked in plain language, inside the product.
See the Analyst agentReview the security architecture
Encryption, isolation, and the controls that apply identically to a person in the UI and a client on the wire.
See security & trustFrequently asked
Questions from engineering teams.
What is the MCP server, in one sentence?
A governed endpoint that lets an AI client outside Rexfin (a chat assistant, an agent runtime, an internal script) query your reconciled model and cited actuals under the same permission rules the product itself enforces.
What can a connected client actually retrieve?
Figures, definitions, and scenarios from your model, always with their citation and reconciliation status attached. It reads the governed model; it does not get a side door to raw, unreconciled source files.
Can a call ever return a number that hasn't been reconciled?
No. The same checks that gate an export gate a query; if a figure has not cleared reconciliation, the response says so instead of guessing at a value.
Who decides which clients are allowed to connect?
Admins scope access by role, by client, and by metric, the same roles-and-access model used everywhere else in Rexfin, applied to programmatic callers too.
Does connecting more AI clients change our security posture?
No new posture, same one: TLS 1.2+ in transit, AES-256 at rest, per-organization isolation, and no shared-model training on your data, enforced identically whether the caller is the product UI or an MCP client.
Proof, not a promise
Works with what you file
- PDF filings
- Bank statements
- ERP exports
- ZATCA e-invoices
- Excel schedules
MCP server
Give every AI client the same evidence your team sees.
Book a demo and connect a real client to a sample of your stack: same permissions, same citations, same reconciled numbers.
Book a demo
Ready to give your developers a governed way in?
We'll connect a sample of your stack and show a real MCP call resolve to a cited, reconciled answer.