An open-source privacy compliance agent for India's DPDP Act and the EU/UK GDPR, built as 20 Claude Code skills.
It reads your code, schemas and data flows, writes findings that cite the exact section, rule or article, and exports them to GitHub code scanning.

Quick start · Sample output · Skills · Guardrails · Coverage · Contribute

---

The DPDP clock

India's Digital Personal Data Protection Rules switch on in stages. The Data Protection Board has existed on paper since November 2025. Consent Managers can register from 13 November 2026. On 13 May 2027 everything else lands at once: notices, consent, security safeguards, 72-hour breach reports, Data Principal rights, and penalties of up to ₹250 crore for a single failed safeguard. MeitY has already consulted on pulling that date forward.

A compliance checklist starts from the law and works down to the systems. Vigil starts from the other end: the code. It finds the Aadhaar column stored in plaintext, the marketing checkbox bundled into the terms, and the incident runbook with no 72-hour step. Each one comes back as a finding with a clause, a severity, a file and line, and a fix.

It won't tell you you're compliant. It will tell you which line of which file says you're not.

What a run looks like

That picture is sample-co, a made-up lending app. Its files were written by hand to show the format, and no code ships for it, so /audit sample-co has nothing to scan:

• summary.md: the human-readable audit, with posture, findings, penalty arithmetic and a fix-first list
• report.json: 9 schema-valid findings with owners, due dates and cross-references between DPDP and GDPR
• breach-log.jsonl: an incident from detection to CERT-In (5h 25m) and the EU authority (44h 10m), one append-only line per step

To watch a real run, audit examples/acme-demo, a small synthetic app with planted problems, and score the result against the list of what it should find. Quick start step 5 shows how.

How accurate is it?

Two blind runs so far, each on a synthetic app whose code states none of its gaps. The auditor never saw the answer list, and a separate session that did not write the report scored it.

Measure	kadamba-fit, fitness app	tamarind-lend, RBI-registered lender
Planted issues found	16 of 16	13 of 13
Citation anchors exact or more specific	45 of 48	17 of 17
Should-not-flag items flagged anyway	3 of 13; two were mistakes in the answer list, since corrected	5 of 12; three arguably mistakes in the answer list
Sector-only duties kept out of DPDP findings	Not measured	3 of 4
Fixture values copied into any output	0	0

tamarind-lend is the weaker run. A subagent of the session that built the app did the audit, and a commit subject in its context said sector-rule gaps were planted, so read its sector numbers as an upper bound.

AI can make mistakes. Have a licensed attorney check every finding before you rely on it. Both answer lists are public now, so a fair re-run copies the app out of the repo and never opens its .expected.json or run sheet.

Why Vigil

	Spreadsheet checklist	Hosted privacy platform	Vigil
Looks at your code, schemas and logs	No	Usually through integrations	Yes. It runs in your repo.
Cites the DPDP section or rule per finding	If someone typed it in	Varies	Every finding. Validation rejects a missing clause, or a section the law doesn't have.
Knows what is in force on which date	No	Varies	Phased commencement table, checked before a gap is called a breach
Where your data goes	Wherever the sheet lives	The vendor's cloud	Your machine and your model provider. No Vigil server.
Output	Cells	Dashboards	JSON, Markdown, SARIF, OSCAL, OCSF
Cost	Free	Licence fees	Free (Apache 2.0), plus model usage

Quick start

You need Node.js 22+ and Claude Code 2.1.295 or newer; older versions let a write through when the guard fails to start. There is nothing to npm install. On Windows, install Git for Windows too: the hooks run in Git Bash, and without it every write is blocked.

1. Clone and check the workspace.

   git clone https://github.com/prathikcannotcode/Vigil.git
   cd Vigil
   npm run validate -- --strict

2. Open Claude Code in the repo.

   claude

3. Look at the demo. Type /posture sample-co or ask "What's sample-co's penalty exposure?"
4. Add your organisation. Type /onboard or say "Add my company acme-tech, a SaaS company processing data in India and the EU."
5. Audit a codebase. Add the codebase to the session (/add-dir ../your-app), then run /audit. Findings land in assessments/. To try it on the synthetic app first, copy it out of the repo so the run can't read the answer list, then add it the same way:

   cp -r examples/acme-demo ../acme-demo

6. Send findings where your team already looks.

   npm run export:sarif -- --company acme-tech --output findings.sarif

   Upload findings.sarif with github/codeql-action/upload-sarif and findings with a file and line should appear in the Security tab next to your code. Findings with no file, such as a missing contract, and closed findings are left out of SARIF. The exporter names them, and they stay in report.json, OSCAL and OCSF.

Delete company-profile/sample-co/ and assessments/2026-10-19-sample-co-baseline/ whenever you like.

Or install the skills as a plugin

claude plugin marketplace add prathikcannotcode/Vigil

claude plugin install vigil@prathikcannotcode

The plugin carries the 20 skills, namespaced as /vigil:audit, /vigil:dpdp-consent and so on. Questions and lookups, such as /vigil:gdpr-qa and /vigil:dpdp-catalog, work in any project. The skills that save files expect a Vigil workspace: /vigil:audit and /vigil:onboard write assessments/ and company-profile/ into whatever project is open, with no schemas there to check them, and /vigil:export calls npm scripts only the workspace has. The plugin carries no hooks, so nothing guards those writes. To audit, clone the repo.

Skills

Every skill is a real slash command. You can also describe the task in plain English and Claude picks the right one.

Workflows

Command	What it does
/onboard	Gathers organisation details, writes the company profile, detects which frameworks apply, takes a first posture snapshot
/audit	Full DPDP + GDPR audit in four phases: discovery, DPDP, GDPR, synthesis. /audit dpdp or /audit gdpr for one framework
/remediate	Turns findings into a plan ranked by penalty exposure, with owners, due dates and code fixes
/refresh	Quarterly re-assessment: regulatory changes, PII re-scan, data map drift, before/after posture

DPDP Act 2023 and Rules 2025

Command	What it does	Anchored in
/dpdp-consent	Notices, valid consent, withdrawal, Consent Managers, dark-pattern review, parental consent	s.5, s.6, s.9; Rules 3, 4, 10
/dpdp-breach	Assess, contain, notify the Board and every Data Principal, CERT-In and RBI in parallel	s.8(6); Rule 7
/dpdp-rights	Access, correction, erasure, grievances (90-day ceiling), nomination, SLA tracking	ss.11-14; Rule 14
/dpdp-dpia	Screening, full DPIA, annual review for Significant Data Fiduciaries	s.10(2)(c); Rule 13
/dpdp-dpo	DPO charter, independence check, monitoring calendar, quarterly board report	s.10(2)(a), s.8(9); Rule 9
/dpdp-catalog	Section and rule map, commencement dates, penalty Schedule, sector regulators	Act, Rules, RoD Order 2026

EU GDPR and UK GDPR

Command	What it does	Anchored in
/gdpr-audit	Code and system audit: personal data, lawful basis, minimisation, security, retention, transfers	Art. 5, 6, 9, 25, 32, 44-49
/gdpr-data-flows	Data flow mapping and Record of Processing Activities check	Art. 30
/gdpr-draft	Privacy notice, DPA, consent banner, retention policy, rights procedure	Art. 13-14, 28
/gdpr-qa	Answers with article citations, recent CJEU and EDPB developments, UK divergence	Whole Regulation, DUAA 2025
/gdpr-catalog	Article map, penalty tiers, transfer mechanisms, UK DUAA changes	Art. 83, 44-49

Cross-framework tools

Command	What it does
/pii-scan	Finds personal data in code, schemas, logs and configs, including Aadhaar, PAN and UPI IDs. Reports locations, never values
/posture	Six-dimension posture across DPDP and GDPR, saved as a snapshot
/penalty	Maximum exposure under the DPDP Schedule and GDPR Art. 83, with Art. 83(3) grouping so linked infringements are not double-counted
/reg-monitor	Upcoming dates, pending changes and their impact on you
/export	SARIF 2.1.0, OSCAL Assessment Results, OCSF Compliance Findings

Prompts to try, by role

Compliance officer

• What's our compliance posture?
• Which frameworks apply to acme-tech, and what are the penalty caps?
• Which DPDP obligations are in force today, and which start in May 2027?

Data Protection Officer

• Set up DPO governance for acme-tech.
• Prepare the quarterly board report.
• Has DPDP or GDPR changed anything that affects us?

Engineering

• Scan this codebase for personal data.
• Are there any dark patterns in our consent flow?
• Draft a fix for F-0003 and tell me which clause it closes.

Legal and privacy counsel

• Draft a GDPR privacy notice for our SaaS product.
• Is storing IP addresses in access logs GDPR compliant?
• Draft a Data Processing Agreement for our analytics vendor.

During an incident

• We've had a data breach. Walk me through the response.

How it works

flowchart LR
    you([You]) -->|"/audit"| cc["Claude Code"]
    cc --> skills["20 skills<br/>dpdp-* · gdpr-* · tools"]
    skills -->|"Write / Edit"| guard{{"write-guard<br/>PreToolUse hook"}}
    guard -.->|"blocked: outside layout,<br/>log rewrite, secret"| cc
    guard -->|"allowed"| ws[("company-profile/<br/>assessments/")]
    ws --> val{{"validate-on-write<br/>PostToolUse hook"}}
    val -.->|"schema errors"| cc
    ws --> exp["npm run export:*"]
    exp --> sarif["SARIF: GitHub code scanning"]
    exp --> oscal["OSCAL: GRC tools"]
    exp --> ocsf["OCSF: SIEM"]

Loading

Skills are Markdown that tell Claude what the law requires and how to check for it. Everything Claude writes with its file tools passes through two hooks and must match a JSON Schema. The agent can be wrong about the law. Through those tools it cannot write a finding that cites a section the law doesn't have, rewrite a breach log, or drop a file outside the workspace. A shell command skips the hooks, which is why the shell stays behind a permission prompt.

Guardrails

Guard	Enforced by	What it stops
Workspace layout	write-guard.mjs reading layout.json	Writes to any path the layout does not list, path traversal out of the repo, edits to the hooks, schemas and settings themselves, and hook events it cannot read
Append-only logs	write-guard.mjs	Any write to breach-log.jsonl or rights-log.jsonl whose result does not start with the existing file. Corrections are new lines.
Schema validation	validate-on-write.mjs, npm run validate	A finding without a clause or severity, a breach entry without an event, a posture with an unknown status. 9 schemas.
Clause numbers	The same two	A clause naming a section, rule or article its framework doesn't have, or another instrument entirely. Sub-clauses and cross_reference are not checked.
Secrets	write-guard.mjs	AWS, GitHub, OpenAI and Anthropic-style keys, private keys, password assignments
No personal data in findings	write-guard.mjs, skill rules, schema shape	Aadhaar, PAN and +91 mobile numbers in company or assessment data. Findings carry file and line, never the value found there. Emails, dates of birth, card and account numbers and health data pass the hook and rest on the skill rules.
Tests	npm test, GitHub Actions	Regressions in all of the above. 27 tests, Node's built-in runner.

These guards catch an agent's mistakes in Claude Code's file tools. They are not a security boundary. A shell command through Bash is not intercepted, so .claude/settings.json pre-approves only the validate and test commands. The personal-data filter knows three formats, and a clause that passes is well formed, not legally checked. SECURITY.md lists every limit.

Your own company profiles and assessments are gitignored by default; only the synthetic demo is tracked. Use git add -f if you do want them in a private fork.

Headless, CI and KPIs

Run a workflow without an interactive session, for example from a nightly job:

node .claude/scripts/run-headless.mjs --workflow audit --company acme-tech --framework dpdp --target ../your-app

--target is the codebase to read. Add --dry-run to print the prompt, or --harness opencode --model <provider/model> to run on OpenCode (no --target there, and the guardrail hooks are Claude Code settings, so don't assume they run). Every argument is validated before it reaches the shell, and the transcript lands in kpis/data/raw/sessions/, including a partial one if the run fails.

On a fresh clone, run claude once interactively and accept the trust dialog first. Until then Claude Code ignores the permissions.allow list in .claude/settings.json.

npm run kpis turns transcripts and each company's latest report into six numbers. Add -- --company acme-tech for one company; otherwise sample-co is pooled with yours until you delete it.

KPI	Source	On sample-co
Cost per audit	total_cost_usd from Claude Code's JSON output	no runs recorded yet
Coverage	examined / applicable controls in each report, as counted by the assessor	84.8%
Actionability	findings with clause, recommendation, owner and due date	33.3%
Time to remediation	remediated_at minus found_at, median and p90	5 days
Posture trend	dimensions improved or regressed between snapshots	1 snapshot
Penalty exposure trend	exposure per snapshot	₹650 crore, €90M

Regulatory coverage

Last verified 2 November 2026, and the CERT-In, RBI Digital Lending and KYC rows of /dpdp-catalog on 11 November 2026. Dates and clauses trace to the Gazette of India, EUR-Lex and legislation.gov.uk; see /dpdp-catalog and /gdpr-catalog.

Framework	What Vigil covers
DPDP Act 2023	Sections 2-17, 27-33, 44, the penalty Schedule, the Removal of Difficulties Order 2026
DPDP Rules 2025	All 23 rules, with their phased commencement dates (13 Nov 2025, 13 Nov 2026, 13 May 2027)
EU GDPR	Principles, lawful bases, rights, controller and processor duties, breaches, DPIAs, transfers, Art. 83 penalties, recent CJEU and EDPB developments
UK GDPR + DUAA 2025	Recognised legitimate interests, Art. 22A-22D, the transfer "data protection test", complaints handling, PECR changes
Indian sector rules	CERT-In Directions 2022, RBI Cybersecurity Directions 2026, RBI Digital Lending Directions 2025, the RBI KYC Master Direction, SEBI CSCRF, IRDAI guidelines, CCPA Dark Patterns Guidelines 2023, as context only. They appear in cross_reference and the sector guides, and nothing validates them.

Plus 7 consent notice templates (including Hindi) in examples/consent-notices and sector guides for fintech, healthtech and edtech. Sector counsel has not reviewed the guides. Sector detection has been measured once, on the fintech run above; healthtech and edtech have no run yet.

What's in the repo

Path	Purpose
.claude/skills/<name>/SKILL.md	The 20 skills
.claude/schemas/	9 JSON Schemas and layout.json, the list of writable paths
.claude/scripts/	Hooks, validator, exporters, KPIs, headless runner, tests
company-profile/{id}/	Per-organisation data: details, data map, consent records, DPIAs, logs, posture
assessments/{id}/	Point-in-time audits: report, posture, summary
examples/	Consent notice templates, sector guides and a synthetic app to audit
docs/assets/	README images, logo, brand sheet and design tokens

FAQ

Does my data leave my machine? Vigil has no server and no telemetry. Claude Code sends the files it reads to your model provider, as it does on any repo, so point it at the same code you would let a coding agent read.

Why audit before May 2027? Because fixing a plaintext Aadhaar column is a migration today and an incident report later. Before the main commencement date, Vigil labels DPDP gaps as readiness findings instead of violations.

Is this legal advice? No. It is tooling that shows its working: every finding names the clause so a lawyer can check it in minutes rather than rebuild it from scratch.

Can I add another framework? Yes. Add <prefix>-* skills and a catalog skill; CONTRIBUTING.md has the shape.

Contributing

Regulatory updates, sector guides and notices in more Eighth Schedule languages are the most useful contributions. Read CONTRIBUTING.md, then AGENTS.md, which binds human and AI contributors alike. Report vulnerabilities through SECURITY.md. Changes are listed in CHANGELOG.md.

Disclaimer

AI can make mistakes. Vigil is compliance tooling, not legal advice. Have a licensed attorney check every finding before you rely on it.

License

Apache License 2.0.