FinOpsAI Agentic Platform — MCP · A2A · Ask FinOpsAI

Your cloud and AI spend, answered by any assistant.

FinOpsAI is built agent-first. The same 37 tools, the same role permissions and the same audit trail serve Gemini Enterprise, Claude, Microsoft Copilot Studio, Amazon Bedrock agents and the assistant inside the app. Every figure comes from the bill or the vendor, per currency; every change is confirmed by a person and written to the audit log.

Connector documentation All endpoints
Three doors, one server

Every assistant your teams already use becomes a FinOpsAI front end

An external assistant, the assistant inside FinOpsAI, and other agents all call the same Model Context Protocol server. Nothing is duplicated: one catalogue, one permission model, one audit log.

Door 1 · External assistants Gemini Enterprise · Claude Copilot Studio · Bedrock agents Door 2 · Ask FinOpsAI the assistant inside the app no setup · permissions follow your role Door 3 · A2A agents agent card at /.well-known/agent.json JSON-RPC, human-in-the-loop skills OAuth 2.1 PKCE · scopes by role token = one workspace FinOpsAI MCP server mcp.finopsai.app/mcp 37 tools · Streamable HTTP readOnlyHint · destructiveHint confirm before any write FinOpsAI API same role gates as the app FOCUS billing data GCP · AWS · Azure · AI vendors Audit log every call, every action Scopes by role — Viewer 22 read tools · Member 30 (adds AI reads) · Owner / Admin 37 (adds confirmed actions)
One server behind three doors. Permissions follow the signed-in person: a Viewer sees 22 read tools, a Member adds the 8 AI reads, an Owner or Admin adds the 7 confirmed actions.
Model Context Protocol

37 tools, typed and annotated, for any MCP client

A remote MCP server over Streamable HTTP (protocol 2025-11-25, 2025-06-18 and 2025-03-26). Each tool is a typed wrapper over one FinOpsAI API route, so an assistant can never see more than the person who connected it.

22 reads · cloud

Cost, savings & planning

Workspace, cost overview, anomalies and root cause, savings position and history, recommendations and efficiency signals, commitments, carbon, forecast, variance, budgets, allocation, Kubernetes, licences, multi-cloud explorer, data coverage, resource utilisation, invoice reconciliation, recommendation backlog, issued reports.

8 reads · AI

AI spend & usage

AI spend in two layers (vendor APIs and cloud-managed AI), spend by model, usage by person with token economics, explorer by seven dimensions, AI budgets and token allocations, AI efficiency opportunities, shadow-AI findings, AI commitments (GPU/TPU, PTU).

7 actions · confirmed

Act, with a person in the loop

Apply or dismiss a recommendation, set an AI budget, set a token allocation, approve an AI service, acknowledge an anomaly, raise a Jira ticket. Each carries destructiveHint, is restated by the assistant, confirmed by the user and written to the audit log. FinOpsAI never changes resources in your cloud.

PersonAssistantFinOpsAI MCP serverAudit log "Why did our bill jump, and close that anomaly if it was planned" list_anomalies → explain_anomaly (reads, no prompt) figures per currency · MEASURED / ESTIMATE · as-of date · ids answer + "I can close anomaly … as accepted cost — confirm?" confirm acknowledge_anomaly (destructiveHint) who · tool · arguments · time "Closed — audit entry recorded 14:02 UTC"
One question, end to end. Reads run without a prompt; a write is proposed, confirmed, executed once, and audited.

Connect from any client

Gemini Enterprise — Data stores → Custom MCP server, OAuth 2.0 with the client your admin registered.
Claude — Settings → Connectors → Add custom connector → https://mcp.finopsai.app/mcp, sign in and pick your workspace.
Copilot Studio · Bedrock agents — add an MCP tool server with the same URL. Discovery follows RFC 9728 and RFC 8414.

What the assistant is told

The server's instructions are short and public: use the figures the tools return and never convert currencies; say the provenance in words — measured, estimate or not measured, never 0 for absent; quote the as-of date; call the explaining tool for "why"; before any state-changing tool, restate the change and ask the user to confirm.

Agent-to-Agent

An A2A agent that explains what it finds and fixes it with your approval

FinOpsAI publishes an A2A v1.0 agent card and a JSON-RPC endpoint, so other agents and orchestration platforms can hand it a task and receive structured artefacts back. Four skills are exposed: anomaly detection, root-cause analysis, remediation planning and human-in-the-loop governance.

Nightly detectorScoutBrainApprovalAction + audit robust median / MAD gateBQML ARIMA_PLUS ensemblemateriality floor per currencyseverity bands gathers the evidence:daily series, services,projects, commitments,recent changes remediation cardwaits for a personapprove · defer · dismissreason recorded confirmed action runs onceaudit entry writtenrealised saving measuredon the next close Gemini 2.5 Flash on Vertex AIasia-south1 · PII redactionplain-language root causeremediation plan with exactcommands, as cardsnothing is changed here Reachable three ways: A2A task (agent.json) · MCP tools list_anomalies → explain_anomaly → acknowledge_anomaly · Ask FinOpsAI in the appSovereign-resident design: processing in asia-south1, no customer cloud credentials held, every write confirmed and audited.
The anomaly agent. Detection is statistical and reproducible; language is used only to explain and to plan; execution waits for a person.
Agent card

/.well-known/agent.json

Name, description, skills, input and output modes and the JSON-RPC endpoint — the standard A2A discovery document, served from finopsai.app.

Skills

Four, with artefacts

Anomaly detection returns the ranked list with ids; root-cause analysis returns the explanation; remediation planning returns cards with exact commands; governance returns the approval state and audit reference.

Identity

Workspace API key bearer

A2A tasks authenticate with a FinOpsAI workspace API key, scoped to one workspace. The same role gates, product gates and handover lock apply as everywhere else.

Ask FinOpsAI

The assistant inside the app — same tools, no setup

Ask FinOpsAI runs the MCP catalogue in-process. A question becomes up to three tool rounds; figures appear as tiles with their currency and provenance; a proposed action appears as a confirm card. Owners and Admins confirm writes; Members read AI data; Viewers read cloud data.

Grounded

Every number traceable

Each figure in an answer is one a tool returned — the tile shows the tool, the currency and whether it is measured, an estimate, configured or not measured. Nothing is extrapolated by the model.

Bounded

Limits by design

Three tool rounds, six tool calls and a 25-second wall clock per question; conversation history is held server-side and never leaves the workspace; transient model congestion is retried and, if it persists, reported plainly.

Accountable

Confirm, then audit

A write is never executed from text alone. The assistant proposes, the person confirms in the card, the action runs once, and the audit log records who confirmed what and when — the same entry an external assistant would produce.

QuestionTool rounds ≤ 3Grounded answerPending actionConfirm natural languagehistory server-sideone quota unitin-process MCP cataloguescopes from the user's role≤ 6 calls · 25 s wall clockfigure tiles per currencyMEASURED / ESTIMATE /CONFIGURED / NOT MEASUREDwrite tool proposedarguments shown10-minute window Owner / Admin clicksruns onceaudited Engine: FinOpsAI Brain on Google Cloud, asia-south1 · same catalogue, permissions and audit trail as every external assistant
Ask FinOpsAI is the third door into the same server — which is why what you see in the app and what Gemini Enterprise or Claude answers never disagree.
Governance

Built so security teams can say yes

Identity

OAuth 2.1 with PKCE

FinOpsAI runs its own authorisation server. Clients are pre-registered; tokens are bound to one workspace and one role and live sixty minutes; scopes finops.read, finops.ai.read and finops.write map to Viewer, Member and Owner/Admin. Discovery per RFC 8414 and RFC 9728.

Least privilege

Nothing broader than the person

Every tool call is executed as the signed-in user through the same API the app uses, so role, product, subscription and handover gates apply unchanged. The tool list itself is filtered by scope — an assistant never sees a tool its user cannot run.

Safety annotations

readOnlyHint · destructiveHint

30 tools are read-only and idempotent. 7 change state; each is confirmed by the user before it runs, runs once, and is written to the audit log with the arguments. FinOpsAI does not hold credentials that could change resources in your cloud.

Residency

asia-south1, end to end

Billing data, the assistant's model calls and the audit log are processed in Mumbai. FinOpsAI connects to your clouds with customer-granted read identities — a GCP IAM grant, an AWS role via web identity, an Azure Entra workload identity — and stores no cloud keys.

Edge

WAF and rate limits

A Cloud Armor policy with OWASP rules fronts the API; per-user rate limits and per-plan AI quotas bound usage; the server reports its own busy state plainly rather than failing silently.

Transparency

Usage you can see

The MCP Connectors page in the app shows every connected assistant, who granted it, which permissions, tool calls over 30 days, and the actions recorded — and lets an Owner revoke any grant at once.