Download the PHP package netresearch/nr-llm without Composer

nr-llm — The Shared AI Foundation for TYPO3

One LLM setup. Every extension. Full admin control.

nr-llm is shared infrastructure for AI in TYPO3 — like the caching framework, but for language models. Administrators configure providers once; every AI-powered extension on the site uses them automatically.

The Problem

Every TYPO3 extension that wants AI capabilities today has to:

• Build its own provider integration (HTTP calls, auth, error handling, streaming)
• Store API keys in its own way (often plaintext in extension settings)
• Create its own backend configuration UI
• Leave administrators with no central overview of AI usage or costs

When a site runs three AI extensions, that's three separate API key configurations, three places to check when something breaks, and no way to switch providers globally.

The Solution

nr-llm provides the missing shared layer:

For Extension Developers

Add AI to your TYPO3 extension in 5 minutes — no API key handling, no HTTP client code, no provider-specific logic.

1. Require the package

2. Inject the service you need

That's it. Provider selection, API keys, caching, error handling — all managed by nr-llm.

What you get for free

Available services

Why not call the OpenAI API directly?

You can — but then your extension only works with OpenAI. Your users can't switch to Anthropic, use a local Ollama instance, or route through Azure. And every extension on the site manages its own API keys, its own error handling, its own caching.

nr-llm solves this once for the entire TYPO3 ecosystem.

Developer Guide — Full API reference, custom provider registration, typed options, response objects

Integration Guide — Step-by-step tutorial for building your extension on nr-llm

For TYPO3 Administrators

One dashboard for all AI

The Admin Tools > LLM backend module gives you full control:

• Providers — Register API connections (OpenAI, Anthropic, Gemini, Ollama, …)
• Models — Define which models are available and their capabilities
• Configurations — Create use-case presets (temperature, system prompts, token limits)
• Tasks — Define reusable prompt templates for editors
• User Budgets — Cap per-backend-user spending across every preset (requests / tokens / cost × daily / monthly). See the Administration guide.
• Analytics — Estimated cost and usage trends with breakdowns per provider, model, and service, plus per-user consumption against monthly budgets. See the Administration guide.
• Skills — Ingest SKILL.md instruction files from GitHub (SHA-pinned, admin-reviewed) and attach them to tasks and configurations. See the Administration guide.
• Tools — 41 built-in read-only function-calling tools in 8 toggleable groups — content (search, page content, record history: "who changed this?"), structure (TCA/FlexForm schema, URL resolution, TCA validation), configuration (TypoScript/TSconfig, syntax check, site config), code (last exception, source read/search), files (FAL browse/search/references), system (URL probe, diagnostics), accounts (BE users/groups) and rag (cited site-content evidence from the installed search index — EXT:solr, ke_search, indexed_search or a database fallback). See the Administration guide.

Tool Playground — a glass box for agent runs

The admin-only Playground runs the bounded agent loop against any configuration and streams the whole nr_llm ↔ LLM dialog live: every request (messages + tools offered), every response (structured, raw JSON, extracted thinking), every tool execution with arguments, result and duration — plus a dry-run mode that shows the exact prompt without calling the model.

Resilience

• Fallback chain — Each configuration can list other configurations to retry against on a connection error, HTTP 5xx, or rate-limit. Streaming requests are excluded (chunks can't be replayed). See ADR-021.
• Per-capability permissions — Every AI capability (chat, vision, tools, embeddings, …) is a native TYPO3 customPermOptions entry. Check a box per BE group; admins bypass. See ADR-023.
• Dashboard widgets — When typo3/cms-dashboard is installed, "AI cost this month" and "AI requests by provider (7d)" show on the dashboard sourced from the existing usage table. See ADR-024.

Security by default

• Encrypted API keys — All keys stored as vault identifiers (UUIDs) via nr-vault envelope encryption; nr-llm never stores raw keys
• Admin-only access — Backend module restricted to administrators
• No plaintext secrets — Keys never stored or logged in plain text

Setup in 2 minutes

The Setup Wizard auto-detects your provider type from the endpoint URL, discovers available models, and generates a ready-to-use configuration. Paste your API key and go.

AI-powered wizards

• Task Wizard — Describe what you need in plain language; the AI generates a complete task with configuration, system prompt, and model recommendation in one step
• Configuration Wizard — AI-assisted configuration generation with system prompts, parameters, and model selection
• Model Discovery — "Fetch Models" button on model fields queries the provider API and auto-fills capabilities and pricing

Supported providers

For Agencies & Solution Architects

• Reduce integration effort — AI capabilities across client projects without per-project plumbing
• No vendor lock-in — Switch from OpenAI to Anthropic (or a local model) without code changes
• Compliance-friendly — Encrypted keys, admin-only access, SBOM and SLSA provenance on every release
• Local-first option — Ollama support means AI features work without sending data to external APIs
• Production-proven — Powers t3x-cowriter, the CKEditor 5 AI writing assistant for TYPO3

Built on nr-llm

Building on nr-llm? Open a PR to add your extension here.

Architecture

The three-tier configuration hierarchy separates concerns cleanly:

Benefits:

• Multiple API keys per provider type (prod/dev/backup)
• Custom endpoints (Azure OpenAI, Ollama, vLLM, local models)
• Reusable model definitions across configurations
• Clear separation of concerns

Packaging: one extension for now

nr-llm ships as a single extension today, even though it bundles several independently useful subsystems — the provider core, specialized services (translation / image / speech), the tool & agent system, the guardrail safety layer, and the backend UI.

This is deliberate. While the extension is under rapid pre-1.0 development the internal contracts between these subsystems still change often, and splitting now would freeze them into public cross-extension APIs prematurely and add coordinated multi-repo release friction. We keep the architecture split-ready instead — the phpat architecture tests enforce the vertical layering (Controller → Service → Provider), and the horizontal module seams are kept clean by review (with phpat rules for those seams a planned split-readiness step) — and will revisit a split with or before the 1.0 release, once the seams have stabilized.

The anticipated seams, if/when we split:

See ADR-090 for the full rationale and the criteria that would trigger a split.

Requirements

• PHP 8.2+
• TYPO3 v13.4 LTS or v14.3 LTS
• PSR-18 HTTP Client (e.g., guzzlehttp/guzzle)

Installation

Then activate in Admin Tools > Extensions and run Admin Tools > LLM > Setup Wizard.

Documentation

• Introduction — Overview and use cases
• Configuration — Backend module setup
• Developer Guide — API reference, services, custom providers
• Integration Guide — Build your extension on nr-llm
• Architecture — Design decisions and ADRs

Supply Chain Security

This project implements supply chain security best practices:

SLSA Level 3 Provenance

All releases include SLSA Level 3 build provenance attestations, providing:

• Non-falsifiable provenance: Cryptographically signed attestations
• Isolated build environment: Builds run in GitHub Actions with no user-controlled steps
• Source integrity: Provenance links artifacts to exact source commit

Verifying Release Artifacts

Verifying Signatures (Cosign)

All release artifacts are signed using Sigstore Cosign keyless signing:

Software Bill of Materials (SBOM)

Each release includes SBOMs in both SPDX and CycloneDX formats:

• nr_llm-X.Y.Z.sbom.spdx.json - SPDX format
• nr_llm-X.Y.Z.sbom.cdx.json - CycloneDX format

Checksums

SHA256 checksums for all artifacts are provided in checksums.txt, which is also signed.

License

GPL-2.0-or-later

Author

Netresearch DTT GmbH https://www.netresearch.de