Guide

User Guide

Last updated: 15 August 2026 — kept in step with the platform as features ship.

Getting started

Go to autogovern.io/app — the Governance Workbench. No sign-up is required for any tool there; it's free to use and runs immediately.

A short onboarding flow asks three quick questions — what kind of AI system, what area it operates in, what you want to do first — and jumps you straight to the most relevant tool.

You can always skip it and browse all 32 tools yourself.

The six Workbench journeys

The 32 tools are grouped into six outcome-oriented journeys in the left-hand nav, in the order most people actually use them:

1 Assess & approve an AI system

Start here for any single system you're evaluating:

System Intake classifies its EU AI Act risk tier and maps the exact obligations that apply.
The Fairness Scanner computes five real metrics across the three classical fairness criteria the research literature (and IBM's open-source AIF360 toolkit) organizes group-fairness around: disparate impact and demographic parity (Independence — are selection rates equal?), equal opportunity and equalized odds (Separation — are error rates equal?), and predictive parity (Sufficiency — is precision equal?). A model can score perfectly on the first four and still fail predictive parity if groups have different underlying outcome rates — see Methodology §4 for the exact formulas and thresholds.
Drift & Data Quality and the Doc Gap Scanner analyze real data or documentation you provide — entirely in your browser.
LLM Safety Check probes for prompt-injection weaknesses.
The Governance Copilot answers questions about your results.
Reg Timeline, Art. 50 Transparency, Model Risk Mgmt, Fair Lending, and FS AI RMF & Resilience cover specific regulatory angles depending on your sector.

2 Build my governance program

Once you're assessing more than one system, these tools build reusable, org-wide artifacts:

AI Org Governance generates a full governance charter and working plan.
Model Card Gen produces Annex IV documentation.
Compliance Crosswalk lets you answer a control question once and see it satisfy every framework it maps to.
Vendor Answer Pack does the same for vendor due-diligence questionnaires — fill it out once, reuse it everywhere.

3 Enforce & remediate in production

For systems already live. Most of these tools act on the running system; the first one only reasons about it, and says so below:

The Governance Agent runs a reasoning loop over your assessment — it reviews what you have, works out the highest-value next step, and writes up what it would do and why. It is analysis only: the gaps and recommendations it produces live for the length of the run and are never written to your score, your dossier or your registry, so copy anything you want to keep.
The Agent Control Plane answers allow/review/deny in real time for any action an agent asks it about, with a tamper-evident, hash-chained ledger and a kill switch. It is a checkpoint your systems call — not a proxy that intercepts them — so your agent has to make the call (via the API or SDKs) and has to honour the answer; an agent that never asks is not governed by it.
The Policy Compiler turns a plain-English rule like "customer-support agents may read account data but may not change payment details without human approval" into a structured policy the Control Plane actually enforces on every action it evaluates, once you activate it — see §3 below for the full workflow.
The Governance Autopilot continuously scans every registered system for gaps — missing owners, missing controls, overdue or failing tests, elevated residual risk — using data the tools above already compute, and recommends fixes. It applies them automatically only for one low-impact reversible action, only if you turn the automation level up, and only if an active policy doesn't block it.
The Remediation Center offers eight concrete fixes — add a policy check, add an approval gate, revoke an MCP server, disable a tool, block a deployment, add a control/test case, re-run an evaluation, or suspend a dormant Control Plane key — that a human proposes or approves — nothing there auto-applies either, and where a real follow-up check exists it's verified, not just marked done.
The Digital Twin & Simulation tool lets you test a change before it's real, across six real-data scenario types: backtest a candidate policy against your actual logged history, recompute residual risk with one control hypothetically failing, walk the real dependency graph from a vendor or model to see what would break, run a representative attack battery through your real active policies, or force a real upcoming regulation to "active today." Every scenario runs real data through the platform's own engines and never writes to a system, asset, control, or policy.

4 Detect emerging risk

The tools that look outward — at real-world failures, unmanaged AI, vendors, and attacks:

Incident Intelligence matches your system's profile against a database of real, documented AI failures.
Regulatory Watch surfaces candidate regulatory changes for you to review and approve into your Regulatory Horizon.
Shadow AI Detection scans code or config you paste or upload for exposed provider keys, live model endpoints, agent-framework imports, and MCP server configs — AI use that never went through governance review.
The Vendor Risk Analyzer scans a vendor's contract or DPA text for the clauses that matter (training on your data, subprocessors, data residency, breach notification, audit rights), and can cross-check the vendor's name against the live incident database.
Red Team & Adversarial Testing runs a battery of real prompt-injection, jailbreak, and instruction-extraction attacks against your own model or agent endpoint, free and on demand — most AI-governance platforms treat this as a paid add-on or don't offer it. Prompt-injection/jailbreak findings are objectively verified with a unique marker token; instruction-extraction findings are explicitly labeled heuristic for manual review.
The Model & Artifact Security Scanner checks whether an uploaded pickle, PyTorch checkpoint, or .safetensors file can execute code when loaded — a real opcode-level pickle disassembler, not a keyword search, flagging any reference to a dangerous callable like os.system or eval. Runs entirely in your browser; the file is never uploaded.
The AI Threat Matrix maps your System Intake profile against two free, authoritative external datasets — the MIT AI Risk Repository (1,178 real risk entries) and MITRE ATLAS (173 real adversarial-AI techniques, 63 real case studies) — plus an AVID vulnerability feed. Every match states exactly which profile fact triggered it and links to a real, cited source; nothing here is a fabricated relevance score.

Policy Compiler

Open Policy Compiler in the Workbench. Write a governance rule the way you'd say it out loud — "Customer-support agents may read account data but may not change payment details without human approval" — and Compile policy converts it into a structured rule:

Who it applies to
What resource it covers
Which action types
A decision — allow, review, or deny

Review and edit every field before saving; nothing is guessed silently. Save it as a draft, then Simulate it against generated test cases (or your own) to see exactly what it would decide before it affects anything real.

Only active policies are consulted by the Agent Control Plane — activating checks for conflicts with other active policies covering the same ground with a different decision first. Every edit creates a new version; Rollback restores the previous one.

The full set of active policies is available as a signed, machine-readable bundle at GET /api/policy-bundle for any external system to fetch and verify — the same cryptographic key that signs Trust Passports.

5 Prepare for audit or executive review

The Compliance Dossier compiles everything into a single exportable report with a prioritized remediation roadmap.
The AI Registry lists every system, deadline, and control across your program.
Regulatory Horizon shows what's changing that affects you, with countdowns and concrete next actions.

6 Govern the agent economy

For when your AI agents start dealing with other organizations' AI agents:

The Agent-to-Agent Handshake decides whether one agent should transact with another. The calling agent presents its signed Trust Passport; the handshake verifies that signature, checks what the passport actually attests to, and returns an allow/review/deny — with the decision sealed into the same tamper-evident ledger the Agent Control Plane writes to. This is how an agent proves it is governed to a counterparty who has no other reason to trust it.

Data-flow badges

Every tool is tagged with exactly where its data goes — hover any badge in the Workbench for the full detail. See the Methodology page for which specific tools fall into each category:

Local-only AutoGovern-hosted Third-party AI processing Air-gapped / CLI-only

Local-only — the analysis runs entirely in your browser; nothing is uploaded.
AutoGovern-hosted — the result is saved to our database, so you can revisit or export it.
Third-party AI processing (Governance Copilot, Governance Agent) — your question or assessment context is sent to a configured LLM provider to generate a response.
Air-gapped / CLI-only — applies to the downloadable Aegis and Pulse scanners (see §12).

Organization accounts

Every free tool works with zero sign-up. Creating an account at /account is entirely optional, for teams who want:

A shared organization that multiple people can belong to
Member roles (owner, admin, member, viewer)
An audit log of sign-ins and account activity

To create one:

Go to /account and choose "Create Account".
Enter an organization name, your email, and a password (at least 10 characters).
You're automatically made the organization's owner.

As an owner or admin, you can invite teammates from the same page:

Enter their email and a role (member, admin, or viewer) and send.
They get a link to accept — if they already have an autogovern.io account they just sign in; if not, they set a password to create one.

If you belong to more than one organization, a switcher appears automatically so you can move between them.

Anything you save while signed in — a registered AI system, a saved assessment, a Compliance Crosswalk report, or a Vendor Answer Pack — is tagged to your organization and listed on the account page.

A registered AI system is additionally kept out of the public Workbench registry that unregistered visitors see. (Assessments, crosswalk reports, and vendor packs were never public either way — they're only ever reachable by their private share link.)

Owners and admins can change any other member's role or remove them from the member list. Only an owner can promote someone to owner, change another owner's role, or remove an owner — an organization always keeps at least one owner.

Anyone can leave an organization themselves at any time.

AI Inventory & relationships

From the account page, signed-in members can register the assets their systems actually use — models, agents, prompts, tools, MCP servers, datasets, vendors, business processes, and deployments — then link a system to each one it depends on.

This lets you trace any registered system to its full dependency graph — its models, tools, prompts, and datasets — from one place, rather than keeping that knowledge in someone's head.

Registering an asset with a name that closely matches an existing one of the same kind is flagged (not blocked) so you can catch accidental duplicates — the same check applies to systems in the AI Registry.

A relationship isn't limited to system→asset — either side of the "Link" form can be a system or an asset.

So you can also link two assets directly (an agent that delegates to another agent, a tool that's provided by a vendor, and so on), building out a fuller dependency graph than just what each system touches.

Systems now support the full lifecycle: proposed → experiment → design → development → testing → validation → awaiting approval → approved → production → suspended → retired.

Any system registered without an owner email is flagged with a warning on the account page so it doesn't get lost.

The Workbench's background reassessment process — which already re-checked regulatory obligations and evidence coverage on a schedule — now also watches for drift in a system's own details: vendor, geography, autonomy, jurisdictions, lifecycle stage.

It also watches the assets linked to it — a model's version changing, a tool or dataset's configuration changing. Anything it finds shows up in that system's reassessment history in the AI Registry.

Owners and admins can export their organization's entire inventory (every system, asset, and relationship) as a single file from the account page, and import one back in — including into a different organization, which always creates fresh records rather than overwriting anything.

An agent or MCP server asset can also be turned into a real, callable identity that authenticates to the Agent Control Plane — see §7 below.

Agent & MCP Registry

Any AI Inventory asset of kind agent or MCP server can be activated for the Control Plane — turning it from a passive description into a verifiable, revocable identity with its own API key. This reuses your existing inventory instead of asking you to maintain a second, separate registry.

From the account page's AI Inventory list, next to any agent/mcp_server row:

Click Activate for Control Plane. You get an API key, shown exactly once — copy it immediately, since it can't be retrieved again afterward.
Have that agent (or whatever calls the Control Plane on its behalf) send the key as Authorization: Bearer <key> on every POST /api/control-plane/evaluate call.
The action is logged under that asset's real registered name — whatever agent value the request body claims is ignored once a valid key is present, so one agent can't spoof another's identity.

From the same account page list, at any time:

Rotate key — issues a brand-new key and immediately invalidates the old one.
Suspend — the next call using that key is denied before policy evaluation even runs, the per-agent equivalent of the fleet-wide kill switch in the Agent Control Plane Workbench tool. Reversible.
Reactivate — resumes normal evaluation for a suspended key.
Revoke — terminal. The key can never be reactivated; activate a fresh key if you need to reuse that identity later.
View ledger — that identity's own slice of the tamper-evident ledger, scoped to your organization only. It never appears in the public anonymous demo ledger, and no other organization can see or manage your agent, even if they somehow guessed its ID.

Governance API

Once you've activated an agent (§7), you can call the same Control Plane your own Workbench uses from any external system — your own agent framework, an MCP server, a CI pipeline — for free, with no separate developer signup or sales call. The full reference lives at /governance-api; this is the short version.

Send Authorization: Bearer <key> plus a JSON action (at minimum a tool and an action_type) to POST /api/control-plane/evaluate.
You get back a decision — allow, review, or deny — plus the risk score, matched controls, and reasons behind it, the same shape the Workbench itself uses.
A review decision can be resolved later via POST /api/control-plane/action/:id/decision, using the ledger id from the response.

Two free, dependency-free SDKs are downloadable directly from the docs page — a single file each, no package manager install:

Node.js — autogovern.js (uses the native fetch, no dependencies)
Python — autogovern.py (uses only the standard library's urllib)

Both expose an evaluate() call and an assertAllowed() / assert_allowed() convenience method that raises unless the decision is a clean allow.

Authenticated calls get their own rate limit — 300 requests/minute per agent key — kept entirely separate from the 60-requests/minute-per-IP limit on the free anonymous Workbench demo, so real production traffic isn't throttled at demo-tier levels.

Controls Register & evidence packages

Open any system in the Workbench's AI Registry (Tool 30) and scroll to Controls Register to define what should be tested, how often, and who gets escalated to if it fails:

Title
Category — Model / LLM, Agentic, Data, or Operational
Test cadence, in days
An optional escalation email

Run the relevant Workbench tool yourself and log what it found with one click — Log pass or Log fail:

The Fairness Scanner for a bias control
Drift & Data Quality for a drift control
LLM Safety Check for a prompt-injection/jailbreak control
...and so on for the other control categories

A control that's logged as failing, or that's gone past its own test cadence without a fresh result (overdue), elevates that system's residual risk.

This is shown as a banner at the top of the system record, recomputed automatically every time the background reassessment daemon runs (every 6 hours) — not just when you happen to look.

A failing control also raises the reassessment alert level to critical.

If the system has an owner email or the control has an escalation contact, this sends a one-time email the moment it newly becomes critical — it won't re-send on every run while the issue is still open.

Click Generate evidence package to compile a system's obligations, evidence coverage, controls register, and recent reassessment history into one signed bundle.

It's cryptographically signed with the same key that signs Trust Passports, so anyone can fetch it back by its link and see it re-verify the signature — not just trust that the JSON hasn't been edited.

Next to it, Generate AI-BOM compiles the same system's full dependency graph — every model, dataset, prompt, tool, MCP server, and vendor linked to it in your AI Inventory, each with its version, owner, and status.

It also includes evaluation results and security findings, signed the same way.

If a prior AI-BOM exists for the system, the result also shows what changed since then (assets added or removed, versions bumped) — a shareable, independently verifiable answer to "what does this system actually depend on?"

Trust Passport

Once you've saved a Compliance Dossier, you can mint a Trust Passport — a cryptographically signed, publicly verifiable summary of your system's risk tier, readiness score, and governance checks.

Anyone with the link (or the embeddable badge) can verify it was genuinely issued by autogovern.io and hasn't been altered, without needing an account of their own.

Issue one from the Compliance Dossier panel (optionally add a scope, e.g. "Production deployment, EU region"); verify any passport at /passport/:id.

You can revoke a passport you issued from the same panel — a revoked passport shows a clear "Revoked" banner on its public page rather than looking like a generic verification failure.

Every passport carries a level, computed automatically and never something you can just select:

Self-Assessed — the default.
Evidence-Verified — if it's linked to a registered system with real evidence coverage and no failing controls.
Independently Assured — if an external auditor has signed off on it through the Auditor Portal below.

It also expires (180 days by default) — the public page shows a distinct "Expired" state once that date passes.

From a system's page in the AI Registry, generate an Auditor Portal link to invite an external auditor to review it — no account required on their end, just the link.

They see your evidence coverage and controls-register status (never raw ledger data), can submit findings, and you can post a management response from the same AI Registry page.

When they're satisfied, they sign off with a name and email (self-asserted, not a verified credential — disclosed as such) to issue a signed, public assurance report at /assurance/:id, which automatically upgrades any linked Trust Passport to Independently Assured.

AutoGovern Rating & Diligence Reports

Every publicly listed Trust Passport also carries an AutoGovern Rating — a single, comparable A–F score recomputed live from readiness, assurance level, your Control Plane telemetry, and Incident Intelligence exposure, weighted by how much governance quality actually matters for that system's domain (its materiality).

The full, ranked list — /ratings
The published methodology behind every number — /standard
A live-generated aggregate report — /report

The rating panel on your passport page includes a copy-paste embeddable badge.

From a registered system's page in the AI Registry, generate a signed Diligence Report — either an investor report (M&A / portfolio diligence) or an insurer report (underwriting submission). It compiles that system's Rating, AI-BOM, evidence coverage, and incident exposure into one document at /diligence/:id.

A board or investor audience can start at /board instead of the practitioner Workbench.

Aegis & Pulse (free downloads)

Aegis is a free, standalone security scanner (macOS/Windows/Linux) that finds real misconfigurations on your machine — fully offline, air-gapped by default.
Pulse is a free command-line tool that re-ranks your existing vulnerability report (or scans a path directly) against live CISA-KEV and EPSS exploit data, so you can tell which CVEs are actually being exploited.

Both are downloadable from the homepage (Aegis, Pulse) with no account required.

The free public reference (no account)

Everything below is open, needs no sign-in, and is maintained by the same agents that run the Workbench. Each page is a permanent URL you can bookmark, cite or send to a colleague.

60-second check — six questions, then your risk tier and the deadlines that actually apply. Nothing is stored; the answer lives in the URL, so it is shareable.
Regulation reference — one page per regulation we track: status, catalogued requirements, the controls that satisfy each one, and the dates attached to it. Every page shows when our agents last read that regulation's own official source, and warns you when it has moved since we wrote the summary.
Compare regulations — two or three regulations side by side on the same obligation themes. It shows where they overlap, what only one of them demands, and which deadline lands first. Useful when you already comply with one regime and need to know how much of the next one you get for free. An empty cell means we have not catalogued a requirement there, not that the law is silent.
Regulatory deadlines — every dated obligation we track, filterable, with a calendar feed you can subscribe to so new ones appear on their own.
Incident radar — curated AI failure case studies plus a live feed from the news, each curated one with its own page and the control that would have prevented it.
Show your work — our own agents' record, including the failures: what they changed with no human involved, what our own integrity agent currently says is wrong with us, every prediction this blog has committed to and whether it came true, and the official links we publish that turned out to be dead. It is the standard we argue for on this site, applied to us.
Benchmark index, Trust Registry, Ratings leaderboard, State of AI Governance and the weekly digest.

These are plain-English summaries written by us, not legal advice. Every page links its official source — read that before you rely on anything.

Learn more

Methodology & Scoring — the exact formulas behind every score and classification.
Privacy Policy, Data Processing Addendum, Terms of Service.
Questions: info@autogovern.io.

Methodology & Scoring Governance Workbench Home