DPO Radio

MCP Read Tools

Overview Image

Why MCP Read Tools Matter

A privacy engineer running a regulatory-research session in an external Claude client cannot defend an answer that is grounded in general training data. The regulator asks where the citation came from, and "the model said so" is not an answer. The session has to be grounded in the customer's actually-installed regulatory packs, with citations that map back to the source text.

Forseti's MCP read tools fix the grounding problem. Each read tool runs against the customer's installed packs and records, returns inline citations, and dispatches through the same internal handler the in-product chat uses — so the external session's answers are identical to the in-product session's answers. The privacy engineer can move between Anthropic Console and the in-product Forseti chat without re-grounding the conversation.

The read-tool surface is intentionally narrow. Eighteen read tools plus two memory read tools plus one persona-handoff advisory tool. No read tool can mutate any state. No read tool can return data outside the customer organisation. No read tool can return AI memory writes; memory writes happen only at workflow accept-time in the platform.

How It Works

Regulatory Search

Search the customer's installed regulatory packs for requirements, guidances, and procedures matching a query. Returns matches with inline citations naming the framework and the specific requirement. The search runs against the same regulatory index the in-product chat searches.

Document Search and Document Read

Search Forseti project documents for matching text, then retrieve a specific document for citation context. The document surface is scoped to documents the calling key has read permission on; documents outside the calling user's permission scope are filtered before results return.

Cross-Module Record Search

Search records across modules — DPIAs, RoPA activities, vendors, contracts, incidents, rights requests, governance documents, controls. The search also reaches the sector registers. These include penalty readiness and enforcement, authority directions, legal holds, regulated accounts, sector licences and marketing approvals. They also include credit-information supply and uses, data-service provider records, product classifications and conformity certificates. Investor verifications and disclosures, student records, publications, marketing consent and suppression, supervisory reporting and cross-border identity exchanges are covered too. The search is scoped to the customer organisation and to the record permissions of the calling key.

Installed Framework and Article Reading

Three tools read the frameworks the customer has installed. The first lists them, with each framework's status, effective date, jurisdiction, requirement count and article count. The second opens one framework in full. It returns the dates, parent and child frameworks, requirements a page at a time with a short obligation summary, the deadlines, the guidance titles, and a coverage summary across the articles. The third reads the articles themselves, each with its coverage state: implemented by a requirement, no duty on the controller, or not yet mapped. It can also read one article by number, with its heading in English and Vietnamese, the party the article is addressed to where the framework records one, and the requirements mapped to it. The answers are the ones the framework's Articles tab gives. A framework the customer has not installed returns "not installed", never content. Every answer states that the article is the smallest unit, so an external session cites articles and never invents a clause or point.

Reading the Instrument's Own Words

For Vietnamese instruments, three more tools read the text itself rather than its heading. The first searches the recorded Vietnamese text of the installed frameworks and returns short passages, each placed in its article where the passage falls inside one. The second reads one article in full, with the source's title and the pages it spans where the source has page marks. It also says how the text was obtained: read by OCR from the signed scan, extracted from the PDF text layer, or extracted from the DOCX original. The third reads a whole source a segment at a time, and every long answer says where to continue, so nothing is cut off without notice.

The text is the instrument's original Vietnamese as recorded from its signed source. It is not translated, later amendments are not merged into it, and it does not certify that an article is in force today; every answer says so. An article whose heading the source does not carry returns no text at all, never the heading in its place. A framework that groups other frameworks points to its members, and a draft that was later promulgated points to the instrument that replaced it. A framework the customer has not installed returns "not installed". Each source is checked against its recorded fingerprint before any of it is served.

When an answer cites an article, the citation counts as grounded only if one of these reads, or regulatory search, returned that article. A citation nothing grounded is marked for the reader to check.

Full-Record Reads

Record search returns summaries. A companion read tool opens one record, from a module record search reaches and under the same per-module read permission, returning the fields its module declares readable and naming every field it withholds. Anything that identifies a person, a protected custody reference and the tenant key are never among the readable fields, and because every withheld field is named, the calling session knows a field was withheld rather than reading it as empty. The data-subject register is searchable but refuses a full-record read.

Framework Conflict Detection

Surface known conflicts between the customer's installed regulatory packs. Returns the conflict description, the requirement IDs from both sides, the suggested resolutions, and the acknowledgement state.

Workflow Catalog Search

Browse the AI workflow catalog by name, status, or scope. Used to discover which built-in workflows are available and which user-defined workflows the organisation has authored.

Compliance Form Search and Read

Four tools cover the compliance-forms surface: search form templates by market, category, framework, or issuer; read one template's full schema; search form instances by status, template, or the parent record they hang off; and read one instance. Template reads and instance reads each require the matching form-library read permission, resolved against the calling key's own permissions.

Compliance Form Field Drafting

A fifth forms tool asks Forseti to draft values for a template's fields and returns them, per field, with citations. It never writes those values to the form. A server-side policy decides which fields may be drafted at all. References to users, organisations, vendors and systems, attachments and signatures are excluded by construction, and a dropdown field may only ever be drafted with a value from its own option list. The operator accepts or rejects each drafted field in the platform, and only that acceptance writes anything.

Memory Queries

Look up organisation-scoped memories by anchor or run a free-text search across the organisation's memories. Both memory tools are read-only. Memory writes happen at workflow accept-time in the platform, not through any MCP path.

Persona Handoff Advisory

The persona-handoff tool recommends a Forseti specialist persona for a question excerpt. It is advisory; calling it does not mutate the chat session. The recommendation is a suggestion the calling integration can act on.

The Difference

Graphic Image

Without Forseti MCP Read Tools

Graphic Image

With Forseti MCP Read Tools

IconExternal AI sessions answer regulatory questions from general training data, not from the customer's installed packs.
IconEvery read tool runs against the customer's installed packs and records, with inline citations to the source.
IconCitations either are not provided or cannot be verified back to the source pack.
IconThe same internal handler dispatches in-product and MCP read tools; answers are identical.
IconA question about what an article says is answered from a summary, because the session has no access to the article's text.
IconCross-module record search reaches DPIAs, RoPA, vendors, contracts, incidents, rights requests and the sector registers, and opens one record, returning the fields its module declares readable and naming every field it withholds — scoped to the calling key's permissions.
IconExternal sessions cannot reach the customer's own records to ground answers in the operational state.
IconFramework reading answers which articles of an installed framework no requirement covers yet, with the same answer the Articles tab gives.
IconMemory of prior organisational decisions is not available outside the in-product chat.
IconFor Vietnamese instruments, an article can be quoted in its own recorded words, with the source and page it came from.
IconThe in-product Forseti chat and the external integration drift apart over time as each acquires its own answer state.
IconMemory queries (read-only) surface organisation-scoped context across chats.

Built for Auditable AI

Build For Image

Every read tool emits the same audit event the in-product chat emits, with an explicit MCP source marker. The Audit Trail page can filter on the MCP source to isolate external traffic.

Build For Image

No read tool can mutate any state. The MCP memory tools are read-only by construction; memory writes happen only at workflow accept-time in the platform, not through any MCP path.

Build For Image

Cross-organisation read attempts are impossible by construction. The customer organisation identifier is resolved from the API key row, not from any client-supplied claim.

Overview Image

Use the Installed Sources

The connector can read the relevant recorded regulatory sources and permitted records within its tool set. Use Chats for Forseti's in-product playbook handoff and source review.

Background Image

See MCP Read Tools in Action

Ready to see an Anthropic Console session call regulatory search against your installed packs and return cited answers in real time? Request a personalised demo.

Demo Image
Ronni K. Gothard Christiansen

Ronni K. Gothard Christiansen - Technical Privacy Engineer & CEO

Technical Compliance Expert, 32+ Years Open Source Advocate, X-BoD Open Source Matters Inc.

Or contact via

ronni@aesirx.io+84 909 500 760

Next Steps

Icon Image

Start a Compliance Pilot

Issue an MCP key scoped to read tools only and run a 30-day pilot against an Anthropic Console regulatory-research session.

Icon Image

Discuss Your Compliance Needs

Talk to our team about which read tools each integration should be scoped to.

Frequently Asked Questions About MCP Read Tools

Yes. Read tools dispatch through the same internal handler the in-product chat uses. The customer's installed regulatory packs, records, and memory are the same on both surfaces. Answers are identical.

No. It is the instrument's original Vietnamese text, recorded from its signed source and shipped with the platform. Recognition errors are possible where the source was a scan, no translation is provided, and the text does not state whether an article is currently in force. For a filing, rely on the official publication.

No. Every read tool is scoped to the customer organisation that issued the API key. Cross-organisation reads are impossible by construction — the organisation identifier is resolved from the key row at authentication.

No Forseti token allowance applies to them; there is no prepaid allowance to draw down or run out. The external client's own model usage is billed by whoever runs that client. Forseti's own AI features run on the provider your organisation configures, and the Usage page reports what they consumed.

The tool returns an explicit "no matches" response. The calling integration receives the empty result; the model does not synthesise an answer from training data. This is part of the citation-grounding contract — no grounding, no conclusion.

Yes. Every read tool call emits an audit event with an explicit MCP source marker, the requesting key's prefix, the tool name, and the call outcome. The Audit Trail page can filter on the MCP source to isolate external traffic.