How is this different from giving an AI database access?

It is not general database access. CanXP MCP exposes a read-only window onto tables you nominate and returns bounded results. The model cannot write, cannot run SQL, and cannot see a table you did not list.

Does our data go to CanXP?

No. The server runs on your infrastructure and connects directly to your database or files. The result becomes context for the question you asked, exactly like opening a file. Credentials are never transmitted to CanXP.

What is available today?

PostgreSQL is available for search, fetch, and list across allowlisted tables. Markdown, text, CSV, and code are available on local or mounted storage. Snowflake, BigQuery, S3, Azure Blob, FHIR, HL7, MySQL, PDF, and DOCX support are planned.

What happens when a question would return too much data?

Every source has a bounded result limit that is capped in configuration. A caller cannot raise it beyond that cap, so one broad question cannot pull a million rows.

Can our security team audit it?

Yes. Published binaries include SHA256 checksums so your team can verify exactly what they run, and the codebase is deliberately small enough to review in an afternoon. Source access for security review is available under NDA — ask us.

What if we already operate MCP servers?

Keep using them. CanXP Research works with any compatible MCP server. CanXP MCP exists for the more common case where you have a database or file share but no server yet.

How are database results ranked beside documents?

CanXP Research uses reciprocal rank fusion to combine each source by position rather than pretending their scores are comparable. Every result retains provenance, and failed sources are reported rather than silently treated as empty.

CanXP MCP

Databases, documents, object storage and FHIR available today

Your data,wherever it lives.

Turn a database or file share into an answerable source with one config file and one binary. No integration project. No credentials sent to CanXP.

Download for Linux

Read-only interfaceExplicit table allowlistFive native buildsVerified checksums

CanXP data cloud

One query · every allowed source

PostgreSQL

ready

File share

ready

Warehouse

planned

Ranked answer

[postgres#1]

[files#2]

[postgres#3]

[files#4]

❯ Which active patients have an incomplete intake?

PostgreSQL + documents searched in parallel

configuration file

read-only methods

runtime dependencies

native platform builds

From database to answer in 60 seconds

No SDK. No connector project.

Before: weeks of integration work before the first question. After: a ten-line file and a binary that runs where your data already is.

Describe the boundary

Name the source, reference the credential, and allowlist only the tables the agent may reach.

canxp-mcp.yaml

sources:
  - name: clinic
    type: postgres
    dsn: env:PG_DSN
    tables: [patients, visits]

Inspect before serving

Validate the file and see exactly what the server would expose before a client connects.

terminal

canxp-mcp --check

✓ source clinic connected
✓ 2 tables allowlisted
✓ configuration valid

Start the server

Run one binary over stdio. CanXP Research—or any MCP client—can now use the source.

terminal

canxp-mcp

CanXP MCP ready
transport: stdio
methods: search, get, list

Connector roadmap

All your major data sources, one boundary.

Relational databases, documents, object storage and FHIR all ship today. Warehouses are in preview. One configuration model, and one allowlist, across every one of them.

Every card says exactly what is available now. Planned connectors are shown as roadmap—not as features you can use today.

available

Relational databases

PostgreSQL · MySQL · MariaDB

Search and retrieve bounded results across only the tables you explicitly allowlist.

available

Documents & file shares

PDF · DOCX · Markdown · text · CSV

Search local or mounted storage without uploading the corpus to another service. PDF and DOCX are extracted to text on the way through.

available

Object storage

Amazon S3 · Azure Blob · MinIO · R2

Search governed object stores from the infrastructure that already has access. Azure uses a read-only SAS token, scoped and expiring.

available

Healthcare

FHIR R4 · Azure Health Data Services

Reach clinical systems through explicit, bounded, read-only source definitions. Only GET is ever issued.

preview

Cloud warehouses

Snowflake · BigQuery

Make curated warehouse tables answerable without exporting a new copy of the data. Built and tested against recorded API responses; not yet exercised against a paid account.

planned

More on the roadmap

Azure SQL · Synapse · OCR

T-SQL engines and optical character recognition for scanned documents are next on the connector roadmap.

Hybrid search

One question. Every source answers together.

CanXP Research searches files, a local vector index, notes, and every configured MCP source in parallel—then uses reciprocal rank fusion to produce one honest ranking.

Provenance stays attached to every result.

Scores from incompatible systems are never compared directly.

A failed source is reported, not quietly treated as empty.

research_search "Where do we handle patient consent?"

index · local semantic search4 hitsfiles · workspace text search11 hitsvault · Obsidian notesemptymcp · clinic PostgreSQL7 hits

1. src/consent.ts [files#1, index#2]

2. patient_consent:8472 [mcp#1]

3. docs/policy.md [index#1]

Parallel retrieval

Rank-level fusion

Source provenance

A security boundary you can inspect

Not a promise. A deliberately small interface.

The product does not ask you to trust a broad claim. Each security property maps to a configuration value, an interface constraint, or code your team can read.

Credentials stay in your environment

Connection strings use references such as env:PG_DSN. They are never written into the config and never sent to CanXP.

dsn: env:PG_DSN

Read-only by construction

The interface exposes search, get, and list. The model cannot submit SQL, and there is no write method to disable.

search · get · list

The allowlist is the boundary

A table that is not listed cannot be searched, read, or discovered. Wildcards are not accepted.

tables: [patients, visits]

Bounded and verifiable

Result limits are capped by your configuration. Every binary ships with a SHA256 checksum so you can verify exactly what you run.

limit: capped · SHA256

Software that holds your database credentials should be short enough to read.

Every build ships with SHA256 checksums, and we make the source available for security review under NDA.

Request a source review

Download · version 0.2.1

Run it where your data already is.

One native binary, no runtime. Put it on the machine or network that already has access to the source.

Download for Linux

SHA256SUMS

macOS · Apple silicon

Native binary · no runtime

macOS · Intel

Native binary · no runtime

Linux · x64

Native binary · no runtime

Detected

Linux · arm64

Native binary · no runtime

Windows · x64

Native binary · no runtime

Bring your own server

Already speak MCP? Keep what works.

CanXP Research is an MCP client. Connect an existing local or remote server, nominate the read tool, and include it in the same ranked search as every other source.

Authenticatecanxp-research mcp auth <name>

Diagnosecanxp-research mcp debug <name>

~/.config/canxp/canxp.jsonc

{
  "research_search": {
    "sources": {
      "index": true,
      "files": true,
      "vault": true
    },
    "mcp": [{
      "server": "warehouse",
      "tool": "search_encounters",
      "queryArg": "query"
    }],
    "timeoutMs": 8000
  }
}

Questions

The details security and data teams ask first.

Direct answers about boundaries, deployment, roadmap, and verification.

The first question should not take a quarter.

Download one binary, define the boundary in one file, and make the data you already have answerable.

Download CanXP MCP Explore CanXP Research