Skip to main content
Glama

Lucairn SDKs

Official client libraries for Lucairn — an EU-based privacy-preserving AI gateway. Lucairn sits between your application (or AI agent) and the upstream LLM provider you choose, removes personal data from prompts before the model ever sees them, and returns a signed Lucairn Certificate proving what was redacted, when, and by which sanitizer layer.

This monorepo hosts four packages at parity:

  • @lucairn/mcp-server — Model Context Protocol server (one-line npx install for Claude Desktop, Cursor, Cline, Continue, …)

  • @lucairn/sdk — TypeScript / Node SDK

  • lucairn — Python SDK

  • github.com/declade/lucairn-sdks/go — Go SDK

Quick start (MCP)

For most agent use cases, the fastest path is the MCP server. No build step, no install — npx runs it on demand:

npx -y @lucairn/mcp-server

Add it to your MCP client config (Claude Desktop's claude_desktop_config.json, Cursor's mcp.json, Cline's cline_mcp_settings.json, Continue, etc.):

{
  "mcpServers": {
    "lucairn": {
      "command": "npx",
      "args": ["-y", "@lucairn/mcp-server"],
      "env": {
        "LUCAIRN_API_KEY": "<your_lucairn_api_key>",
        "ANTHROPIC_API_KEY": "<optional_byok_anthropic_key>",
        "OPENAI_API_KEY": "<optional_byok_openai_key>"
      }
    }
  }
}

Restart your client. The chat_via_lucairn tool becomes available immediately. See mcp-server/README.md for full details.

Related MCP server: ArkForge Trust Layer

What it does

Each request through any Lucairn SDK follows the same pipeline:

  1. PII detection runs on every user message in three layers:

    • Layer 1 — Known-entity matching (your tenant's named entities)

    • Layer 2 — Presidio NER (names, emails, IBANs, addresses, phone numbers, customer IDs, …)

    • Layer 3 — GPU-hosted custom-trained PII shield (Enterprise tier only, optionally trained on your domain corpus)

  2. Detected PII is replaced with placeholders ([PERSON_1], [EMAIL_2], [IBAN_3], …) before the request reaches the upstream LLM.

  3. The selected upstream model sees only the sanitized text. It never receives raw personal data.

  4. The response is returned with a signed compliance certificate (Ed25519 witness signature + RFC 3161 timestamp + Sigstore Rekor inclusion proof).

  5. Response handling depends on tier:

    • Developer (free) — placeholders are returned verbatim. Useful for testing the redaction surface.

    • Pro / Enterprise — placeholders are re-linked back to the originals on the gateway before the response reaches your application.

For Lucairn-hosted Developer-tier callers, on-gateway pseudonymization happens before your LLM sees the request. Enterprise self-host deployments can run the entire stack inside the customer environment, in which case no raw identity data leaves that environment at all.

Provider routing

The gateway picks the upstream provider from the model parameter you send:

Model prefix

Upstream provider

BYOK env var

claude-*, anthropic-*

Anthropic

ANTHROPIC_API_KEY

gpt-*, openai-*, o1-*, o3-*, o4-*

OpenAI

OPENAI_API_KEY

Cross-provider BYOK shipped in @lucairn/mcp-server@1.1.0 — set one or both keys in the same MCP config and the server forwards the matching one as X-Upstream-Key per request, so your provider account is billed directly.

Per-language SDKs

Language

Package

Version

README

MCP server

@lucairn/mcp-server

1.2.7

mcp-server/README.md

TypeScript

@lucairn/sdk

1.1.1

ts/README.md

Python

lucairn

1.4.1

python/README.md

Go

github.com/declade/lucairn-sdks/go

v1.3.1

go/README.md

All SDKs are at parity at the observable level. Cross-language byte-equivalence is locked via shared Go-assembler-generated fixtures, so a certificate signed via one SDK verifies identically via the other two.

Get an API key

Sign up at https://lucairn.eu/account/signup. Free Developer tier: 500 requests/month, no credit card required.

Pro adds response re-linking, programmatic certificate JSON access, audit-event export, and higher quota. Enterprise adds self-host, BYOK with provider-side billing isolation, and the optional custom-trained PII shield (priced per scope).

See https://lucairn.eu/pricing for the full tier comparison.

Verify a response

Every response through any SDK gets a signed Lucairn certificate. Two surfaces:

  • HTML summary — DPO-friendly, available on every tier including Developer (free). Use getCertificateSummary (TS) / get_certificate_summary (Python) / GetCertificateSummary (Go), or paste the certificate URL into https://lucairn.eu/verify.

  • JSON certificate + local Ed25519 verify — Pro tier and above. Use getCertificate + verifyCertificate (and language equivalents). The verifier is in-tree — see ts/src/verify-certificate/, python/src/lucairn/verify_certificate/, and the internal/verify package under go/.

  • Every claim, not only the witness signature — verifyCertificateChain / verify_certificate_chain / VerifyCertificateChain check each claim's own signature against per-service keys you pin, the claim bytes, the typed copies, the signed upstream-request hashes and the unsigned cert_tier. All three return identical results on the shared 98-case corpus in testdata/parity-corpus/. Render and decide only from the returned verified values; pass the expected request id so another turn's certificate is refused.

External RFC 3161 + Sigstore Rekor anchor verification is currently surfaced as pass-through metadata; full external anchor verification lands in a follow-up release.

Status

Production packages are versioned independently and tagged per the table above. Cross-language byte-equivalence is locked via shared fixtures. Follow CHANGELOG.md for release notes.

Contributing

See CONTRIBUTING.md. Security reports: SECURITY.md.

License

MIT — see LICENSE.

Available Tools

1 tool
chat_via_lucairnA

Send a chat request through the Lucairn privacy gateway with cross-provider BYOK (Anthropic + OpenAI). PII is detected and replaced with placeholders before reaching the upstream LLM. The gateway picks the upstream provider based on the model parameter: claude-* / anthropic-* use ANTHROPIC_API_KEY; gpt-* / openai-* / o1-* / o3-* / o4-* use OPENAI_API_KEY. Wire format follows the Anthropic Messages API. Developer-tier responses contain raw placeholders; Pro and Enterprise tiers can enable automatic re-linking back to the original values.

Inputs: model (required), max_tokens (required), messages (required), system, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.4/5.0

    Scored across 1 tool

    Disambiguation5/5

    With only one tool, there is no possibility of ambiguous selection. The single tool clearly defines its action and scope, so agents cannot confuse it with anything else.

    Naming Consistency5/5

    The sole tool name, `chat_via_lucairn`, follows a clean snake_case verb-phrase convention. There are no other names present, so there are no conflicting naming styles to navigate.

    Tool Count4/5

    One tool is slightly below the typical 3–15 tool range, but it is a substantial operation rather than a trivial placeholder. For a narrow privacy-gateway chat use case, the single-tool scope is reasonable.

    Completeness5/5

    The tool fully covers the server's apparent purpose: sending chat requests through the Lucairn gateway with provider routing and PII handling. There are no dead ends in the core workflow, and configuration concerns are handled through parameters and environment keys.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      C
      maintenance
      Provides cryptographic signing and verification for AI decisions to generate verifiable, Ed25519-signed receipts for compliance and auditing. It automatically maps AI actions to regulatory frameworks like HIPAA and SOX with high-performance, sub-3ms signing.
      4
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Third-party certifying proxy — sign any HTTP call (AI agents, webhooks, microservices) with an independent Ed25519 signature, RFC 3161 timestamp, and Sigstore Rekor anchor.
      4
      96 PyPI
      3
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      Security middleware for LLM apps and AI agent pipelines. Detects prompt injection attacks (22 signatures, 7 languages) and anonymizes PII (17 entity types). Deterministic, sub-25ms, GDPR Art.30 compliant.
      -
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides a pre-flight/post-flight firewall for LLM calls with comprehensive detection, classification, policy enforcement, reversible redaction, output safety, and immutable audit logging.
      1
      MIT