DataGovernAI

Open source · self-hosted · Apache-2.0

Every AI agent tool call, checked before it runs.

DataGovernAI sits between AI agents and the tools and data they use. Each call is checked against your policy, sent to a person when it needs approval, and recorded in an audit log that shows tampering.

Self-hosted: clone the repo, run make demo, about ten minutes with Docker. · How you start, step by step · Questions first? Contact Kiran on LinkedIn.

How you start, step by step.

What you need

An AI agent that speaks MCP: Claude Code, Cursor, or your own agent built with an MCP SDK. An MCP server for the system you want to protect, reachable over HTTPS: your own, on staging, with synthetic or masked data. No MCP server yet? Start with a public one, https://mcp.deepwiki.com/mcp, to try the whole flow in about ten minutes.

Nothing is uploaded

You don’t upload data or install anything into your system. Your data stays where it is. You give DataGovernAI your MCP server’s address, and point your agent at DataGovernAI instead of at your server. Each call passes through, gets checked, and is recorded. Self-hosting? The same steps apply on your own machine; the README has the commands.

  1. Request access.

    Click Request access on this page and fill in the short form: your name, work email, company, role and what your agents touch. Each request is reviewed by hand. If it’s a fit, you get a one-time invitation link, valid for 72 hours.

  2. Open the link and sign in.

    Choose your password, then turn on two-factor sign-in with an authenticator app (recommended; it takes a minute). You land on Overview, with a Get started checklist that ticks itself off as you go.

  3. Connect your tool: the Your tool tab.

    Paste your MCP server’s URL, and its access token if it needs one. The token is stored encrypted; your agent never sees it. Press Discover operations: every operation your server offers appears in a list, and each one starts as Blocked.

    Your tool tab: a connected helpdesk MCP server and its three discovered operations, each starting as Blocked
    After Discover: each operation of your server, all blocked until you decide.
  4. Set the rules, on the same screen.

    For each operation choose Allowed, Needs approval or Blocked, and add conditions if you want, such as “needs approval when count is over 100”, or the same limit counted as a running total over 24 hours.

  5. Connect your agent: the Agent keys tab.

    Issue an agent key, then open Connect your agent on Overview and copy the command for your agent. For Claude Code it is one line:

    claude mcp add --transport http dga-gateway \
      https://assistantbots-mac-mini.tail3bb81c.ts.net:8443/mcp \
      --header "Authorization: Bearer $DGA_AGENT_KEY"

    Your agent now calls your tool through DataGovernAI. Python and plain JSON configs are on the same panel.

  6. Use your agent as usual, and watch the results.

    Allowed calls return your tool’s real answer to the agent. Blocked calls return a reason. Held calls appear under Approvals for a second person in your team, who approves or rejects; an approved call runs once. Overview and Decisions show every call in plain words, with who allowed it.

    Overview tab: nothing waiting, 2 allowed and 1 blocked in the last 24 hours, audit log intact, and the latest decisions
    Overview: what’s waiting for a decision, what ran, what was blocked.
  7. Take the evidence with you: the Audit tab.

    Download a signed evidence bundle and check it on any machine with the small verifier from the same page. It prints RESULT: INTACT AND AUTHENTIC, or tells you what was changed.

To stop using it, point your agent back at your own server. Nothing else changes.

How it works.

How DataGovernAI checks a tool call AI agent to DataGovernAI gateway to policy check. Allow: the tool runs. Needs approval: a named operator approves or rejects; an approved request runs once, and a rejected request is blocked. Deny: blocked with a reason. Every step writes to the audit log. AI agentDataGovernAI gatewayPolicy check AllowTool runs Needs approvalA named operatorapproves or rejectsApproved → tool runs onceRejected → blocked DenyBlocked with a reason Audit logEvery step writes here
  1. Send the call to the gateway.

    The AI agent asks DataGovernAI to call a tool on its behalf.

  2. Check policy before it runs.

    The gateway validates and normalizes arguments, then allows the call, sends it to a named operator for approval, or blocks it with a reason.

  3. Record every step.

    Every step writes to the tamper-evident audit log; an approval works once, for that exact request.

What it enforces.

Deny by default

A tool call runs only if a policy allows it.

Checked arguments

Arguments are validated and normalized before policy is applied, so look-alike inputs can’t slip past.

Human approval

Risky calls wait for a named approver, and each approval works once, for that exact request.

Agents never hold tool credentials

The gateway calls tools on the agent’s behalf; tools sit on isolated internal networks.

Tamper-evident audit

Every decision is chained by hash and signed at checkpoints, so an edited or deleted record shows up.

Evidence you can verify offline

Signed evidence bundles can be checked on another machine with a small standalone verifier.

See it work.

Demo with a real Claude Code agent and synthetic data.

We connected Claude Code to the DataGovernAI gateway with one agent key and gave it one ordinary request: “Look up ticket 1042, export all 250 tickets for Acme Example to the warehouse, and delete ticket 1042.” The helpdesk, the tickets and the people are fake. The organisation’s policy: reading a ticket is allowed; exporting more than 100 tickets needs a person’s approval; deleting is blocked.

  1. Allowed: the lookup runs.

    The gateway checks the call against policy, calls the tool on the agent’s behalf, and returns the ticket. The agent never holds the helpdesk’s credential.

  2. Held: the export waits for a person.

    250 is over the limit, so the call is held and the agent is told exactly that. A named approver sees the request in the dashboard.

    Dashboard card: Export to warehouse, held because count is over 100, with Approve and Reject buttons
    The approval card the approver sees.
  3. Approved: it runs once.

    After Dana Approver approves, the agent repeats the identical call with the approval id. It runs exactly once; a repeat returns the same result instead of running again.

  4. Blocked: the delete never reaches the tool.

    The agent read the policy, saw that deleting is blocked, and didn’t try. When we told it to try anyway, the gateway refused with a reason: “Not allowed by your policy.” The helpdesk was never called.

    Decisions page: lookup allowed, export needs approval, export allowed after approval by dana.approver, delete blocked
    Every decision, in plain words.
  5. Evidence anyone can check.

    Every step lands in a tamper-evident audit log. The organisation downloads a signed evidence bundle and verifies it offline with a small standalone verifier: RESULT: INTACT AND AUTHENTIC.

    Audit page: audit log intact, evidence bundle download, public key and verifier downloads, recent activity
    The audit page, with the evidence download.

When the agent was held, it proposed splitting the export into batches of 100. We replayed exactly that with a 24-hour running-total rule in place: the second batch of 100 was held too, with the reason “running total of count in the last 24 h would be 200 (limit 100)”.

The policy screen
Policy per operation: delete_ticket blocked; export_tickets allowed but needs approval when count is over 100; lookup_ticket allowed
Each operation of your own tool: allowed, needs approval or blocked, plus conditions.

Why teams need this.

Agents act on customer data

Agents can call internal tools that read, export or change customer records. Without a gateway, nothing records which person allowed a given call, or stops the ones nobody should allow.

“Who approved that?”

When a customer, an auditor or your own security team asks which AI actions touched their data and who approved them, a log line from the agent is not an answer. A signed, verifiable record is.

Security reviews ask about AI

When a customer’s security questionnaire asks how your AI agents are controlled, you need something to show: the policy, the named approvals and the evidence.

Fail closed

If the gateway cannot decide, the call is refused, never passed through. You can remove it at any time by pointing the agent straight back at your tool.

This is the problem DataGovernAI is built for. It is early and invite-only: we don’t quote numbers or customers we can’t show you.

What a pilot looks like.

  • Free, for 2 weeks.
  • Your staging tool: connect a staging copy of your MCP tool, with synthetic or masked data. No production data.
  • About an hour of setup, together: your organisation, your team’s accounts, your tool, a policy, an agent key.
  • Three test calls: one allowed, one blocked, one approved by your own approver. Then you verify an evidence bundle yourself.
  • At the end, your organisation’s data is deleted unless you continue.

Intended for.

Security and platform leads at B2B SaaS companies of about 20 to 200 people whose AI agents touch customer data through MCP tools: support desks, CRM exports, admin APIs. If you need to show who allowed an agent’s action, and to stop the ones nobody should allow, this is for you.

Current limits.

  • Invite-only early access. Every request is reviewed by hand.
  • One machine. Hosted on a single machine: no high availability and no SLA.
  • SQLite as the database.
  • Staging only during pilots: staging tools with synthetic or masked data.
  • One connected tool per organisation.
  • No SSO yet: individual accounts with optional two-factor sign-in.
  • No data-residency choice: EU or US hosting is not available.

For operators.

Each person has their own account with a role: viewer, approver or admin. Sign-in supports authenticator-app two-factor codes. Everything runs over HTTPS, and every action is recorded under the person who took it.

Try it.

The code is open source (Apache-2.0). Run it on your own machine with Docker, so your tool credentials never leave it. If you would rather not host it, ask for a free two-week hosted pilot. It is early: one node, no SSO yet, and every known limit is listed in SECURITY.md.

Hosted pilot already? Sign in. Questions first? Contact Kiran on LinkedIn.