Connect your AI assistant
Provantix runs a read-only MCP server, so your AI assistant can answer questions about your agents from the verified evidence instead of from memory: "what did agent X do last week?", "which agents touched the CRM?", "show me the open findings", "how does Cruz compare with n8n?".
The server is available from the Pilot plan up. Designer plans do not include it.
What it is
- An MCP (Model Context Protocol) endpoint at
https://mcp.provantix.tech/mcp, served statelessly over Streamable HTTP. - Read-only. No tool can change state; there is nothing to write with.
- Audited. Every call is recorded with your user, the client, the tool and a hash of the arguments, and shown to your org's admins under Settings, Access tab, Connected AI clients.
- Revocable. Any connection can be disconnected there; its tokens stop working within 10 minutes.
Connecting
Endpoint:
https://mcp.provantix.tech/mcp
A person must approve the first connection: the assistant opens a sign-in screen, you sign in with your Provantix account, and you approve the access page. The assistant cannot connect itself.
Clients identify to Provantix in one of three ways: dynamic client registration (RFC 7591, for public clients with https: or loopback redirect URIs), a Client ID Metadata Document (an https: URL that hosts the client's OAuth metadata and is fetched on first use), or a pre-registered client ID. If your client cannot register itself, tell us the client name and we pre-register it for you.
If a client later asks for more than you approved before, the access page appears again and marks the new scopes new. A request for a subset of what you approved connects without the page.
Claude Desktop or claude.ai
- Open Settings → Connectors (claude.ai: Settings → Integrations).
- Choose Add custom connector and paste the endpoint above.
- Choose Connect. A Provantix sign-in page opens in your browser.
- Sign in and press Allow. The connector is live.
Cursor
- Open Settings → Tools & Integrations → MCP (older versions: Features → MCP).
- Add new global MCP server, name it
provantix, type URL, paste the endpoint above. - Cursor connects and opens the sign-in page; sign in and press Allow.
Claude Code
claude mcp add --transport http provantix https://mcp.provantix.tech/mcp
Then run /mcp in a session, select provantix and choose Authenticate to complete the browser sign-in.
From your own code (Claude API, OpenRouter, anything)
Use this when an agent or service you run needs the evidence, not a desktop assistant. Two phases:
1. Get a token, once, with a person approving. Connect an MCP client (for example the MCP Inspector's OAuth flow, or a short script) against the endpoint. The flow is standard OAuth 2.1 with PKCE: open https://app.provantix.tech/api/mcp/oauth/authorize?... in a browser, sign in, approve, and exchange the returned code. Access tokens last 10 minutes; refresh tokens last 60 days and rotate on use, so a service stores the refresh token and mints new access tokens as needed. Your integration needs a client ID: we pre-register one per integration, or you host a Client ID Metadata Document.
2. Call tools. Either run the MCP client next to your code (the @modelcontextprotocol/sdk StreamableHTTPClientTransport with an Authorization: Bearer header) and hand the tool list to your model, or use Anthropic's native MCP connector and let the Messages API manage the calls:
curl https://api.anthropic.com/v1/messages \
-H "content-type: application/json" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "anthropic-beta: mcp-client-2025-11-20" \
-d '{
"model": "claude-sonnet-4-6",
"max_tokens": 1000,
"messages": [{"role": "user", "content": "What did my agents do yesterday? Start with the summary tool."}],
"mcp_servers": [{
"type": "url",
"url": "https://mcp.provantix.tech/mcp",
"name": "provantix",
"authorization_token": "mcp1...."
}]
}'
With OpenRouter or any OpenAI-compatible API, list the tools from the MCP client first, pass them in the request's tools array, and execute the tool calls the model returns, exactly like any other function calling.
Every connection is listed under Settings → Connected AI clients and can be revoked there; the next call with that token then returns 401.
Tools
Twenty tools, in a fixed order: the original eleven, then nine added for reading by source, runs, cost, systems, decisions, reports and exports. Results carry a nextCursor when there is more: most lists page at 25 by default and 100 at most, and list_approvals pages at 50 by default and 200 at most. Every result is capped near 50 KB.
| Tool | What it returns |
|---|---|
whoami |
Who the token belongs to, org, role, granted scopes, plan rate limit |
summarize_activity |
Counts by agent, system, action type and day; a metrics block (tokens, latency, outcomes, models); field completeness, which says how many eligible records carry each field; open gaps and finding counts when those scopes are granted. Ask this first. With compare: "previous_period", raw deltas against the window of the same length just before, with no baseline or breach implied. With baseline_agent_id, one agent's hourly baselines, in their own block. That argument does not filter the totals |
list_agents |
Agents: id, name, owner, lifecycle, risk tier, last activity |
get_agent |
One agent: identities, systems reached (aliases grouped under one name, with the underlying systems listed), a window activity summary, recent records, and its open findings with findings:read |
search_activity |
Record projections filtered by agent, system, action, time or text query. fields adds model, tokens, latency, outcome and tool name |
get_record |
One record's projection, optional fields, and its integrity metadata |
verify_record |
Convenience integrity check (hash, chain, day root) |
list_findings |
Findings with status. Spend findings need cost:read and data-health findings need coverage:read; anything withheld is named, without a count |
list_approvals |
Approval references recorded against records: the approver handle, policy, ticket and times. Filter by record, agent, system or a time range of up to 90 days. It lists only approvals that exist and never infers who approved |
get_coverage |
Covered agent–system pairs, silent sources, sequence gaps, ingest rejections |
get_brief |
The organization's Brief: verdict, headline, data gaps first, activity, coverage, integrity, next actions and trend tiles. Cost, its tiles and the month-end forecast appear only with cost:read |
list_sources |
The places data comes from (emitters, collectors, intake sources) with their names, status, record counts and newest record |
compare_sources |
Two to four sources side by side over a window: activity, findings, coverage, delivery, and spend with cost:read. Each source is read from its own records; there is no combined total |
get_run |
One run (a trace) as a chain: requester, agents, each call in order with target, outcome, attribution, approval and source, and a summary. Sealed bodies and personal values are never returned. Per-call cost needs cost:read |
get_cost |
Spend for the window: total and change, top agents, model mix, unpriced models, budgets over cap, month-end forecast. Needs cost:read |
get_cost_breakdown |
Spend for a period grouped by one or two dimensions (tag:department, tag:product, any tag:<name>, agent, model, provider), with the previous period, the unallocated row and the rows adding up to the total. Open a row with filters. Estimated cost beside any figure the source reported. Never returns spend per person. Needs cost:read |
list_candidates |
Principals that act like agents but are not mapped to one, awaiting review |
list_systems |
The systems agents act on, with record counts and first and last seen; filter by agent, or by new_since to find first-time systems |
list_decisions |
The decision log: approvals, finding status changes with reasons, and administrative changes, with who and when |
run_report |
List the report templates, or run one over a period. Tables are capped at 50 rows; the cost report and cost columns need cost:read |
verify_export |
Check an export: whether it is consistent and whether it is authentic (its signature verifies against Provantix's published keys). Only authentic means Provantix issued it |
Reading one source at a time
Every tool except whoami, list_sources, compare_sources and verify_export accepts an optional source argument: a source name or id (several, comma-separated). It narrows that call to those sources, done by the database so no tool can forget it. The response says it was narrowed (source_filter). Sources are never merged; use compare_sources to set them side by side and do not add their figures together.
Resources and prompts
Resources the assistant can read directly, with the same scope checks as the tools: provantix://brief, provantix://coverage, provantix://open-findings, provantix://sources, provantix://systems, and provantix://sources/{source id}/brief for one source's Brief. Plus the record, agent and finding links the tools return.
Ready-made prompts, which tell the assistant which tools to call and to treat agent-written text as data: whats_changed, investigate_finding, source_health, compare_two_sources and agent_review.
Nothing here can change state: every tool, resource and prompt is read-only.
Narrowing is done by our database, not by the model: summaries and counts are computed in SQL, and records come back lean (id, time, agent, action, target, link) and grow detail only when fields asks for it.
Scopes and cost
A connection is granted four scopes at most:
| Scope | Covers | Requested by default |
|---|---|---|
evidence:read |
Records, agents, approvals, summaries, field completeness, the Brief's activity | Yes |
findings:read |
Findings and finding counts | Yes |
coverage:read |
Agent–system coverage, open gaps, data-health findings | Yes |
cost:read |
Cost estimates, the price rows behind them, cost on records, spend findings, and cost in the Brief | Admins: included automatically. Members: no, the client must ask for it |
An admin has all access inside their own organization, so an admin's connection includes every scope, cost:read among them, even when the client asked for fewer. The access page marks each added scope "included for admins" so the widening is visible, and cost is included whether or not the setting below is on. Roles are read on every call: if an admin becomes a member, their connections lose cost on the next call, because cost then follows the setting below; the findings and coverage scopes any member may hold stay.
For members, a client that asks for cost:read gets it only when your organization allows it. An admin turns that on under Settings, Access tab, Connected AI clients with Allow cost data over MCP. The setting is off by default and governs members only. Turning it off removes cost from every member connection at once, and the next call without cost returns no cost values; admin connections keep it. Turning it on does not add cost to existing member connections: each one must ask and be approved again. While the setting is off, a member's access page shows cost as not allowed by your organization. A connection made by an admin before this change needs to be reconnected once to pick cost up.
A connection belongs to one organization. A person who belongs to several would have to choose which organization to connect; that choice does not exist yet, because a sign-in is tied to a single organization today.
Cost is an estimate from your price table, computed for each record from the price row in force on its day. Editing a price restates the estimates of the records it priced. Each record can name the price row behind its current estimate, but not the rates that row had when the estimate was first made.
Rate limits
| Plan | Calls per minute |
|---|---|
| Pilot | 120 |
| Business | 300 |
| Team | 600 |
| Enterprise | As quoted |
What the assistant can see
- The same redacted projections the Provantix UI shows. Raw prompt or response content is never returned.
- People appear as pseudonymous handles (
person-<hash>), never raw identities. Erased fields never appear. - Strings written by monitored agents (agent names, tool arguments, target ids) are flagged
untrustedand capped, so prompt injection inside recorded data cannot steer your assistant.
Data handling
Everything your assistant reads goes to the AI provider you chose, under your agreement with that provider. Provantix does not add any AI provider as a sub-processor. Your organization controls the connection: admins can disconnect any client's access, and every read is audited as evidence access.