AI agents can dramatically reduce the time required for prior-art discovery, patent-status checks, and technology landscaping—but only when they can interact with authoritative sources in a controlled, auditable way. A WebMCP tool for Indian patent research provides that bridge: it exposes narrowly defined actions that an agent can call to search, retrieve, filter, and cite records from Indian patent-office databases.
The challenge is not simply scraping a website. Indian patent data may be distributed across public search interfaces, PDF documents, registration records, status pages, and external international databases. Interfaces can change, searches may require session state or CAPTCHA handling, and legal conclusions must never be inferred from incomplete data. A robust implementation therefore combines WebMCP tool design, browser automation, data normalization, provenance, and human review.
What Is a WebMCP Tool?
WebMCP can be understood as a tool interface that enables an AI agent to perform structured actions on the web. Instead of giving an agent unrestricted browser control, you expose typed capabilities such as:
- Search Indian patent applications by keyword, applicant, inventor, or publication number
- Retrieve a publication or application record
- Extract bibliographic metadata from a PDF
- Check the latest visible procedural status
- Find related documents using IPC or CPC classifications
- Return source URLs, timestamps, and evidence snippets
The model decides when to call the tool, but your server controls what the tool can do. This separation is essential for reliability and security. The agent should not be allowed to invent URLs, submit arbitrary forms, bypass access controls, or claim that a patent is valid, granted, expired, or enforceable without appropriate legal verification.
A useful design principle is: make every tool action small, deterministic, and evidence-producing. A response should include not only an answer, but also the source page, publication number, retrieval time, query parameters, and confidence or completeness indicators.
Define the Indian Patent Research Use Case First
Before writing code, specify the exact workflow. “Search patents” is too broad for a dependable agent tool. Start with one or two high-value jobs, for example:
1. A startup enters a technical description and receives potentially relevant Indian publications.
2. A patent professional checks whether a publication number appears in the Indian record system.
3. A research team builds a landscape by applicant, technology, IPC class, and filing year.
4. An agent creates a cited shortlist for human prior-art review.
Avoid presenting the tool as an automated legal opinion engine. The tool can retrieve and organize public records; it should not decide novelty, inventive step, ownership, infringement, or enforceability. Those conclusions depend on claim interpretation, prosecution history, jurisdiction, dates, and professional judgment.
For India-aware workflows, distinguish at least these identifiers and events:
- Application number
- Publication number
- Grant or patent number
- Filing date
- Publication date
- Priority date
- Applicant and inventor names
- Patent office branch or jurisdictional information
- IPC classification, and CPC where available
- Examination, opposition, grant, abandonment, or other visible status events
Do not merge these fields into a single “patent status” value. A record can be published but not granted, or show a portal status that requires confirmation from official documents.
Recommended Architecture
A production-grade WebMCP patent tool should have five layers:
1. Agent-facing tool layer
This exposes a small set of JSON-schema-defined actions. Examples include search_indian_patents, get_patent_record, get_patent_document, and find_related_documents.
2. Orchestration layer
This validates inputs, applies rate limits, selects the appropriate source adapter, manages retries, and records the request. It should reject unsupported operations before any browser session starts.
3. Source adapters
Each public source gets its own adapter. An adapter understands page structure, pagination, filters, document links, cookies, and failure conditions. Do not scatter source-specific selectors throughout the agent layer.
4. Normalization and evidence layer
This converts inconsistent source responses into a stable internal model. It also stores raw HTML or PDF references where permitted, extracted fields, source URLs, timestamps, and parsing warnings.
5. Cache and audit layer
Cache repeat searches and document metadata, while preserving a complete audit trail. A cache reduces load on public services and makes agent responses faster and more reproducible.
A simple flow is:
Agent request
↓
WebMCP schema validation
↓
Policy and rate-limit checks
↓
Search or retrieval adapter
↓
Normalization + evidence extraction
↓
Confidence and completeness checks
↓
Structured response with citationsDesign the Tool Schema Carefully
The schema is the contract between the agent and your service. Keep inputs explicit and constrained. For example:
{
"name": "search_indian_patents",
"description": "Search publicly accessible Indian patent records and return cited results.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string", "minLength": 2, "maxLength": 500 },
"fields": {
"type": "array",
"items": { "enum": ["all", "title", "abstract", "applicant", "inventor", "number"] }
},
"ipc_classes": {
"type": "array",
"items": { "type": "string", "pattern": "^[A-H][0-9]{2}[A-Z]" }
},
"filing_year_from": { "type": "integer", "minimum": 1900 },
"filing_year_to": { "type": "integer", "maximum": 2100 },
"page": { "type": "integer", "minimum": 1, "maximum": 20 },
"page_size": { "type": "integer", "minimum": 1, "maximum": 50 }
},
"required": ["query"],
"additionalProperties": false
}
}The response should be equally structured:
{
"query": "battery thermal management",
"results": [
{
"publication_number": "IN2020XXXXXX",
"title": "...",
"applicants": ["..."],
"inventors": ["..."],
"filing_date": "2020-01-15",
"publication_date": "2021-07-23",
"ipc": ["H01M 10/00"],
"source_url": "https://example.gov.in/record/...",
"retrieved_at": "2026-09-03T10:00:00Z",
"evidence": [
{ "field": "title", "text": "...", "source_page": 1 }
],
"warnings": []
}
],
"next_page": 2,
"completeness": "partial"
}Use ISO 8601 dates internally, but preserve the original display value when parsing is uncertain. If the source does not expose a field, return null rather than guessing.
Connect to Indian Patent Office Data Sources Responsibly
The Indian Patent Office’s public systems and related government portals can change layouts, session requirements, and availability. Verify the current official access method before deployment. Prefer documented APIs or permitted download mechanisms where available. If no API exists, browser automation should operate only within the website’s terms, robots guidance, access controls, and reasonable rate limits.
A source adapter should handle:
- Initial session creation and cookies
- Search form tokens and hidden fields
- Pagination and result limits
- PDF or image-document links
- Temporary server errors and timeouts
- Empty-result states
- Maintenance pages and CAPTCHA challenges
- Language or encoding variations
- Duplicate records and inconsistent number formats
Never build a CAPTCHA solver or a mechanism intended to defeat access controls. If a portal presents a challenge, stop, return a clear tool error, and route the task to an approved human or official integration.
For broader discovery, you may combine Indian records with international sources such as WIPO or other legally permitted databases. Label every result by source. An international family record is not proof that the same application has a particular Indian status.
Implement Search, Retrieval, and Citation Workflows
A good agent workflow separates discovery from verification.
Discovery
The agent begins with a normalized query. Convert natural language into searchable concepts, but preserve the original query for auditability. Expand obvious synonyms cautiously—for example, “solid-state battery” and “solid state battery”—without silently broadening into unrelated technologies.
Run multiple constrained searches when useful:
- Exact phrases in title or abstract
- Applicant-name variants
- Inventor-name variants
- IPC class filters
- Publication-number lookup
- Date ranges
Deduplicate by a stable identifier, preferably an official publication or application number. Do not deduplicate solely by title because translations and formatting differ.
Retrieval
For each shortlisted record, retrieve the authoritative page or document. Parse bibliographic data first, then claims and description only if required. Large PDFs should be downloaded once and processed asynchronously where possible.
Evidence
Every material statement returned to an agent should have evidence. For PDF extraction, include page numbers. For HTML records, include the field label and source URL. If OCR was used, mark the text as OCR-derived and assign a lower confidence score.
A useful evidence object contains:
{
"source_url": "https://...",
"retrieved_at": "2026-09-03T10:00:00Z",
"document_identifier": "IN...",
"page": 3,
"field": "priority_date",
"quoted_text": "...",
"extraction_method": "pdf_text",
"confidence": 0.96
}This makes the agent’s answer inspectable and helps patent professionals validate important details quickly.
Use IPC and CPC Classifications for Better Results
Keyword-only search performs poorly for patent research because the same concept can be described with different terminology. International Patent Classification (IPC) codes provide a valuable second retrieval axis. CPC codes may add granularity where available, but should not be assumed to exist in every Indian record.
A practical pipeline is:
1. Ask the agent to identify candidate technical concepts.
2. Map concepts to likely IPC classes using a curated reference table.
3. Search keywords and classifications separately.
4. Intersect or rank results based on textual and classification relevance.
5. Present the classification used and the reason for inclusion.
Do not let a language model assign definitive classifications without validation. Use the model for candidate generation, then check codes against a maintained taxonomy. Store mapping versions so a result can be reproduced later.
Build Ranking Without Creating False Legal Confidence
Ranking is useful for triage, not for legal conclusions. A transparent relevance score can combine:
- Title and abstract term matches
- Claim-term matches, if claims were retrieved
- IPC or CPC overlap
- Applicant or inventor match
- Date proximity
- Document-family relationships
- Source completeness
For example:
score = 0.35 × text_relevance
+ 0.25 × claim_relevance
+ 0.20 × classification_match
+ 0.10 × entity_match
+ 0.10 × metadata_completenessExpose the score components or at least the ranking explanation. Avoid labels such as “anticipates the invention” or “safe to use.” Safer language includes “potentially relevant,” “requires claim-level review,” and “metadata could not be fully verified.”
Handle Names, Dates, and Indian Identifiers
Indian patent records commonly contain variations in applicant names, corporate suffixes, punctuation, transliteration, and historical entities. Normalize names for search while retaining the original form for display.
Recommended normalization fields include:
- Lowercase comparison value
- Unicode-normalized value
- Removed punctuation version
- Alias table for known corporate names
- Original source value
Dates require similar care. Maintain separate fields for filing, priority, publication, grant, and event dates. If a date is shown only in an ambiguous format, preserve it as text and set a parsing warning.
Patent numbers also need normalization. Store:
- Original identifier
- Canonical identifier
- Identifier type
- Source-specific identifier
Never assume that a publication number, application number, and patent number are interchangeable.
Security and Agent-Safety Controls
A web-connected patent tool is an attack surface. Treat all retrieved content as untrusted input. Patent documents may contain prompt-injection-like text, hidden instructions, scripts, or malicious links. The agent must not follow instructions found inside a patent document.
Implement these controls:
- Allowlist domains and URL patterns
- Block arbitrary navigation and downloads
- Sanitize HTML before model exposure
- Scan PDFs and limit file size
- Disable active content where possible
- Enforce request timeouts and concurrency limits
- Restrict outbound network access from parsers
- Validate every tool argument server-side
- Keep secrets outside prompts and browser pages
- Log user, tool, source, query, and outcome
- Separate retrieved text from system instructions
Use a service account or backend proxy rather than exposing portal credentials to an agent. If authentication is required, obtain explicit authorization and comply with the portal’s terms and applicable Indian law.
Privacy, Copyright, and Compliance in India
Public availability does not eliminate compliance obligations. A responsible implementation should consider:
- The Digital Personal Data Protection Act, 2023, where personal data is processed
- Data minimization and retention limits
- Access logging and incident response
- Website terms and acceptable-use rules
- Copyright and database-rights considerations for document storage and redistribution
- Contractual restrictions on commercial reuse
Inventor and applicant data may be public, but your system should still avoid unnecessary enrichment, profiling, or bulk exposure. Store only what the workflow needs, define retention periods, and provide deletion or access processes where applicable.
For legal and patent-service deployments, clearly disclose that automated outputs are research assistance, not legal advice. Route high-impact decisions to a qualified patent agent or attorney registered to practice in the relevant jurisdiction.
Testing and Observability
Test the tool against a fixed evaluation set containing:
- Exact publication-number searches
- Common spelling and punctuation variants
- Applicant-name variants
- No-result queries
- Queries with IPC filters
- Records with missing fields
- PDFs requiring OCR
- Duplicate and family-related records
- Portal downtime and CAPTCHA responses
Measure more than search accuracy. Track:
- Precision of the top results
- Recall on known records
- Identifier extraction accuracy
- Date parsing accuracy
- Citation completeness
- Duplicate rate
- Median and tail latency
- Source error rate
- Percentage of responses marked partial
Create contract tests for the normalized schema so a portal redesign does not silently corrupt records. Use canary runs and alerts when selectors fail or field completeness drops sharply.
A Practical MVP Roadmap
A realistic first version can be built in stages:
Phase 1: Read-only lookup
Implement publication-number lookup, basic keyword search, normalized metadata, source URLs, and retrieval timestamps.
Phase 2: Evidence and PDF extraction
Add document retrieval, page-level citations, OCR fallback, warnings, and human-review links.
Phase 3: Relevance and classification
Add IPC filters, synonym expansion, transparent ranking, deduplication, and family grouping.
Phase 4: Production controls
Add caching, observability, policy enforcement, domain allowlists, access controls, and versioned parsers.
Phase 5: Professional workflow integration
Provide exports to CSV or JSON, review queues, saved searches, annotation, and collaboration features—without presenting automated results as final legal determinations.
Common Mistakes to Avoid
- Giving the agent unrestricted browser access
- Scraping without checking current portal terms
- Treating a search result as proof of legal status
- Confusing application, publication, and grant numbers
- Returning uncited summaries
- Silently guessing missing dates or applicants
- Ignoring OCR errors in scanned PDFs
- Using only English keywords for Indian technology searches
- Failing to distinguish Indian records from international family records
- Storing bulk personal or document data indefinitely
- Building around fragile CSS selectors without monitoring
FAQ
Can an AI agent automatically determine whether an invention is patentable in India?
No. It can support prior-art discovery and organize public records, but patentability requires claim interpretation, legal analysis, complete evidence, and professional judgment.
Should I scrape the Indian Patent Office website directly?
Only when permitted and technically appropriate. Prefer official APIs or approved data access. Respect terms, rate limits, robots guidance, authentication requirements, and CAPTCHA controls.
What should the tool return when a patent record is incomplete?
Return available fields with null values for missing data, a completeness label such as partial, explicit warnings, and the authoritative source URL. Never fill gaps with model-generated guesses.
Is WebMCP the same as giving an agent a browser?
No. A WebMCP tool should expose constrained, typed operations. It can use browser automation behind the scenes, but the agent should not receive unrestricted navigation or arbitrary web actions.
How can Indian AI startups use this technology?
Start with a narrow read-only workflow, add citations and human review, validate against known records, and expand only after the source adapter and compliance controls are stable.
Apply for AI Grants India
Building a secure WebMCP tool for patent intelligence can become a valuable AI product for Indian researchers, founders, and IP professionals. Apply through AI Grants India to explore support and funding opportunities for your AI venture.