VITNA – Containment for rogue AI agents
Server Details
Guard wraps stdio MCP servers: risky calls wait for a person; no answer, no run. Remote: cooperative
- Status
- Healthy
- Uptime
- 100.0% over the last 24 hours
- Last Tested
- Transport
- Streamable HTTP · MCP 2025-03-26
- URL
- Repository
- COSTRINITY/vitna-compliance-mcp
- GitHub Stars
- 0
- Server Listing
- @costrinity/vitna-compliance-mcp
- ChatGPT
- –
- Claude
- –
TDQS
Scored across 23 tools
The set goes out of its way to disambiguate near-neighbours: breach_classify vs us_state_breach_deadline, india_regulators_directory vs india_sectoral_check, consent_check vs vitna_preflight, and vitna_help vs vitna_claim all explicitly cross-reference each other with 'use X instead for Y'. Remaining fuzziness is modest — global_compliance_map overlaps conceptually with the three sectoral/cross-border analysis tools, and pii_test vs vitna_preflight both inspect content, though their descriptions separate them adequately.
Most names use a noun_verb ordering with a consistent vocabulary of verbs (validate, classify, check, mask, test, get), and the vitna_* prefix cleanly groups the three meta/oversight tools. The deviations are the six or so noun-only reference names (global_compliance_map, india_cross_border_status, japan_cross_border_status, india_regulators_directory, sub_processors_register), which break the verb-suffix pattern but remain readable.
23 tools pushes against the upper bound for comfortable selection, and the surface bundles two fairly different products — AI-agent safety oversight and a multi-jurisdiction identity/privacy compliance suite (Aadhaar, CPF, GSTIN, IBAN, PAN, SIN validators). Each tool is individually justified, but the aggregate feels heavy and only loosely unified by the 'VITNA' branding.
Read/write lifecycle is one-sided: the server evaluates and records decisions but exposes no tool to list or retrieve the audit records it claims to produce, and consent_check references a matching_consent_id with no tool to create, list, or revoke consents. Core evaluation flows are well covered, but an agent cannot close the loop on recording, auditing, or managing the state these tools depend on.
Available Tools
23 toolsaadhaar_maskAInspect
Mask + Verhoeff-validate an Aadhaar number. Returns masked form, validity, and an owner-scoped reference token. No persistence of the raw value. Stateless validator: records no decision and leaves no dashboard timeline trace.
Inputs: aadhaar (required).
ai_act_classifyAInspect
Before you build or ship an AI feature, check where it lands under the EU AI Act (Regulation 2024/1689). Describe the use case (with biometric / remote-identification / automated-decision / social-scoring / GPAI flags) and VITNA returns the risk tier (prohibited / high-risk / limited-risk / minimal-risk), GPAI obligations, and the per-tier obligations you would have to meet. A classification for you to act on: VITNA evaluates and records, it does not gate the build.
Inputs: use_case (required), sectors, biometric, social_scoring, +4 more.
breach_classifyAInspect
After a security incident, check whether it is legally reportable before you decide how to respond. Give the incident facts (affected count, data categories, sensitivity, recovery state) and VITNA returns reportability + reasoning + the notification deadline + who to notify, across DPDP §8, GDPR Art 33, CPRA §1798.82, LGPD Art 48, PDPA §26B, and US-FED sectoral. This makes the full incident decision from the facts; for a quick per-US-state deadline/recipient/threshold table without incident facts, use us_state_breach_deadline. VITNA evaluates and records; acting on the result is up to you.
Inputs: sensitivity (required), affected_count (required), recovery_state (required), data_categories (required), +2 more.
consent_checkAInspect
Before you process someone's personal data, ask VITNA whether an active consent actually permits it for this purpose. Give the data principal + purpose (and optional category); returns { allowed, reason, matching_consent_id, principal_id }, a determination you must honour yourself since VITNA evaluates and records but does not enforce. Use this for personal-data processing legality; for a dangerous technical action (shell / file / DB / network) use vitna_preflight instead.
Inputs: purpose (required), category, principal_id, principal_ref.
cpf_validateAInspect
Validate a Brazilian CPF (mod-11 check digits, rejects all-same). Stateless validator: records no decision and leaves no dashboard timeline trace.
Inputs: cpf (required).
dpia_threshold_checkAInspect
Before you start a new processing activity, check whether the law requires a DPIA first (GDPR Art 35 / DPDP §10 / LGPD Art 38). Give the purpose + data categories (and scale / systematic-monitoring / automated-decision / cross-border / vulnerable-subjects flags); returns dpia_required + the 9-criterion WP29 analysis + jurisdiction guidance, so you know whether to pause and assess before proceeding.
Inputs: data_categories (required), processing_purpose (required), scale, cross_border, +4 more.
global_compliance_mapAInspect
Master catalogue of every privacy/security/sectoral regime VITNA has fabric for (28 entries covering 24 named statutes). Stateless lookup: records no decision and leaves no dashboard timeline trace.
Inputs: no named parameters.
gstin_validateAInspect
Validate a GSTIN format + mod-36 check digit; returns state code lookup. Stateless validator: records no decision and leaves no dashboard timeline trace.
Inputs: gstin (required).
iban_validateAInspect
Validate an IBAN format + ISO 7064 mod-97 check digit; supports 71 countries. Stateless validator: records no decision and leaves no dashboard timeline trace.
Inputs: iban (required).
india_cross_border_statusAInspect
Before you transfer personal data out of India, check the destination country's DPDP §16 status (permitted / restricted / sectoral_restricted) plus any RBI / SEBI / IRDAI caveats. Pass the ISO-3166 alpha-2 country code (e.g. US). Stateless lookup: records no decision and leaves no dashboard timeline trace.
Inputs: country (required).
india_regulators_directoryAInspect
Static reference directory of Indian data and sector regulators (DPB, RBI, SEBI, IRDAI, TRAI, DoT, PFRDA, MeitY, MCA), optionally filtered by sector: a lookup of who exists and what they cover. To instead work out which of them apply to a specific processing activity, use india_sectoral_check. Stateless lookup: records no decision and leaves no dashboard timeline trace.
Inputs: sector.
india_sectoral_checkAInspect
Before you process personal data under Indian law, find out which sectoral regulators actually bind your specific activity (RBI / SEBI / IRDAI / TRAI / DoT / PFRDA) from its processing profile, so you know whose rules apply before you act. This analyses your processing to say what applies; for a plain directory of every Indian regulator regardless of your activity, use india_regulators_directory.
Inputs: data_categories (required), processing_purpose (required), sector_hint, counterparty_types.
japan_cross_border_statusAInspect
Before you transfer personal data out of Japan, check the destination country's APPI Art 28 status (adequacy / standard basis / high scrutiny). Pass the ISO-3166 alpha-2 country code. Stateless lookup: records no decision and leaves no dashboard timeline trace.
Inputs: country (required).
pan_classifyAInspect
Classify a PAN entity type from the 4th character (P=Person, C=Company, H=HUF, F=Firm, ...). Stateless validator: records no decision and leaves no dashboard timeline trace.
Inputs: pan (required).
pii_testAInspect
Dry-run VITNA's PII / threat detection on a sample event before you send real data, to preview what would be tagged, how it would be redacted, and whether severity would escalate. Nothing is persisted and nothing is filtered: a safe rehearsal you act on, not an enforced gate — it records no decision and leaves no dashboard timeline trace.
Inputs: sample_event (required), jurisdiction.
privacy_notice_getAInspect
Generate the operator's jurisdiction-templated privacy notice. Returns markdown or JSON. Called with an API key, the operator's name and contact details appear as placeholders; the signed-in owner gets the real values. Stateless generator: records no decision and leaves no dashboard timeline trace.
Inputs: format.
sin_validateAInspect
Validate a Canadian SIN (Luhn checksum); returns series region + masked form. Stateless validator: records no decision and leaves no dashboard timeline trace.
Inputs: sin (required).
sub_processors_registerAInspect
Return the public sub-processor register (Supabase, Vercel, Resend, etc.). Stateless lookup: records no decision and leaves no dashboard timeline trace.
Inputs: no named parameters.
us_sectoral_checkAInspect
Before you process personal data under US law, find out which US federal sectoral regimes bind you (HIPAA, GLBA, COPPA, FERPA, FCRA, SOX) for a given processing profile, so you can factor them in before you act. US-scoped; for Indian sectoral regulators use india_sectoral_check.
Inputs: data_categories (required), processing_purpose (required), ai_decisions, counterparty_types, +1 more.
us_state_breach_deadlineAInspect
Quick reference lookup of a single US state's breach-notification window, AG recipient and resident threshold (e.g. 'CA' gives 500 residents, CA AG, without unreasonable delay). This is a static table, not an incident ruling. When you have the actual incident facts and need a reportable / not-reportable decision with reasoning, use breach_classify instead. Stateless lookup: records no decision and leaves no dashboard timeline trace.
Inputs: state (required).
vitna_claimAInspect
Ask whether this VITNA account has been claimed by a real person yet. Call this when the user asks about their VITNA account or when you are told evidence is not being kept. Returns { claimed, claim_url: null, claim_note, what_claiming_gets_you }. You are never given the claim link, and you must not look for it: only a person may claim the account that oversees you. If claimed is false, tell the user that the account's evidence is NOT being durably recorded and that they can claim it with the claim link VITNA showed them when the trial started (the stdio server prints it in its MCP server log).
Inputs: no named parameters.
vitna_helpAInspect
What is VITNA and how do I use it to keep myself in check? Call this FIRST after connecting to learn the safety and oversight checks available: how to check risky actions BEFORE running them, what a deny / hold decision means, trial vs claimed mode, and how the user can monitor and audit what you do. Runs entirely locally: no account, no API call, and no dashboard timeline trace.
Inputs: no named parameters.
vitna_preflightAInspect
SAFETY / OVERSIGHT CHECK before a dangerous or destructive action (shell command, file deletion, DB statement, network call). Call this to have VITNA check the action BEFORE you run it: it flags dangerous shell / SQL / secret-exfil / prompt-injection / suspicious-network patterns and returns { decision: allowed|blocked|flagged }. RESPONSE SHAPE DEPENDS ON THE KEY: an unclaimed trial key gets the decision label only, because the pattern detail is what makes the classifier worth stealing. A claimed key additionally gets threat_category, severity, reason (the rule names that matched, never the matched text), a threats[] array and a redacted echo of what was scanned, plus a signed audit record the user can review. VITNA evaluates and records; it does NOT enforce, so treat blocked / flagged as a stop and get human approval. Act on decision: when the account's owner has switched it to shadow mode, decision is allowed even where flagged is true, and shadow.would_have says what VITNA would have answered when enforcing. This is how a user keeps you in check. Heuristic pattern match, not a sandbox: novel or obfuscated payloads can pass.
Inputs: action, payload, action_type.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Related MCP Connectors
Runtime behavioral trust scoring for MCP servers. Check reliability before calling unknown tools.
Remote MCP server for supportsheep: run AI interviews and manage support content for your blog.
Remote MCP server for supportsheep: run AI interviews and manage support content for your blog.
Related MCP Servers
AlicenseNot gradedqualityCmaintenanceGoverns any stdio MCP server by interposing between the agent and the server, enforcing policy on every tools/call and refusing denied actions before they reach the upstream server.13 npmApache 2.0- AlicenseAqualityBmaintenanceStdio policy proxy that runs in front of another MCP server: blocks paths outside allowed folders, enforces tool allowlists, asks on the terminal before risky calls, masks secrets in server output and writes a hash-chained audit log.16MIT
- AlicenseAqualityBmaintenanceGuardian is the trust layer for agent commerce. guardian-mcp sits in front of your MCP servers: before a tool call runs, Guardian allows it, asks a person, or denies it, by rules the owner sets, with no AI in the decision. Fails closed. Signed records you can verify offline. Dollar limits on tools that spend money. Free plan: https://odingard.com/guardian3773 npm1Apache 2.0
- AlicenseNot gradedqualityDmaintenanceA local-first stdio proxy for MCP servers that traps destructive tool calls and holds them for manual commit or discard, preventing accidental mutations.AGPL 3.0
Glama MCP Gateway
Add one secure layer between your agents and this server.