Vendor-neutral curation
The selection points to official repositories and authoritative sources relevant to public-sector AI. The cards are starting points for your own assessment — not ready-made procurement or security recommendations.
Curated knowledge
Search by name, type, publisher or topic. Every entry links directly to an official repository or primary source.
54 matches
A technical selection spanning agent workflows, data, local model operations, evaluation, observability and development. Open-source and commercial options are labelled so their operating models are not conflated.
Building blocks and integrations for developing LLM and agent applications in Python or JavaScript.
Points to check: Review the licence, data processing and security of every selected integration and model.
Open official repositoryA low-level framework for modelling stateful, long-running agent workflows with explicit control of flow and state.
Points to check: Orchestration is not a security boundary; tools, networks and data access still require external enforcement.
Open official repositoryA terminal-based coding agent that can connect to different models and development workflows.
Points to check: Run only with constrained permissions, approved models and protection for repositories, secrets and production environments.
Open official repositoryA small, extensible terminal harness for agentic coding with extensions, skills, prompts and multiple model providers.
Points to check: Pi intentionally has no built-in sandbox and runs with the user’s permissions. Use a container, VM or other isolation for untrusted work.
Open official repositoryA framework for connecting agentic applications to documents, data sources, retrieval and indexing.
Points to check: Integrations and associated services may have different licences and data paths from the core; assess each separately.
Open official repositoryConverts PDF, DOCX, PPTX, HTML and other formats into a common structured document representation for downstream use.
Points to check: Individual models may have licences that differ from the codebase. Test quality, OCR and sensitive-document handling in your own environment.
Open official repositoryA local runtime and distribution mechanism for downloading and running language models on your own machine or infrastructure.
Points to check: The runtime licence does not automatically cover models. Check model licences, hardware needs, access control and update practices.
Open official repositoryA high-throughput LLM inference and serving engine, relevant when operating models on your own GPU infrastructure.
Points to check: Requires specialist capability for GPU operations, scaling, patching, monitoring and API security.
Open official repositorySupports tracing, evaluation, prompt management, datasets and debugging for LLM and agent applications.
Points to check: Some enterprise features require a licence. Define retention, access, masking and processing location before collecting traces containing content.
Open official repositoryAn open protocol for connecting AI applications to tools, data sources and other external capabilities.
Points to check: MCP servers are real integration points. Permissions, inputs, outputs and supply chain require threat modelling and enforcement outside the model.
Open official specificationA protocol for interoperability between agents, including discovery, tasks and exchanges across system boundaries.
Points to check: Protocol support does not replace identity, authorisation, data classification, audit or control of actions.
Open official repositoryMicrosoft’s framework for building, orchestrating, evaluating and operating agents and multi-agent workflows in Python and .NET.
Points to check: Define identity, tool permissions, approval points and traceability separately for each workflow.
Open official repositoryA code-first Python framework for developing, evaluating and deploying agents with tools, sessions and multi-agent patterns.
Points to check: Review model, hosting and data choices even though the framework can be used across providers.
Open official repositoryModular pipelines for retrieval, routing, memory, evaluation and agent workflows with explicit control of context.
Points to check: External components and model integrations can change data flows, licences and operating responsibilities.
Open official repositoryA compact Python library from Hugging Face for tool-using and code agents across different models.
Points to check: Run code-generated actions in a real sandbox with constrained data, file and network permissions.
Open official repositoryA C/C++ runtime for local inference across many model architectures on CPUs and GPUs, including models in GGUF format.
Points to check: Quality, security and usage rights also depend on the chosen model and its licence; benchmark on your own hardware.
Open official repositoryA local, OpenAI-compatible API for text, audio and image models, among others, on your own infrastructure.
Points to check: API compatibility does not mean functional equivalence. Validate model behaviour, isolation, patching and capacity before production.
Open official repositoryA platform for tracing, evaluating, registering and monitoring agent, LLM and conventional machine-learning systems.
Points to check: Configure access, retention and redaction before prompts, responses or personal data are collected in traces and experiments.
Open official repositoryA CLI and library for reproducible evaluations, model comparisons and testing for issues such as prompt injection and jailbreaks.
Points to check: Local hooks and providers can execute code with the user’s permissions; isolate tests and keep production secrets out.
Open official repositoryAn extensible framework for detecting, masking, redacting and pseudonymising personal information in text, images and structured data.
Points to check: Automated detection may miss sensitive information; combine it with other controls and measured quality assurance.
Open official repositoryA library for programmable input, output, dialogue and tool controls around LLM-based applications.
Points to check: Guardrails reduce selected risks but do not guarantee protection from prompt injection, incorrect outputs or misuse.
Open official repositoryA general-purpose policy engine that keeps authorisation and control rules outside an agent and evaluates them as code.
Points to check: Policies only work when they are enforced consistently across every relevant API, tool and infrastructure layer.
Open official repositoryVendor-neutral collection, processing and export of traces, metrics and logs from applications and agent workflows.
Points to check: Telemetry can contain prompts, responses and identifiers; filter, minimise and protect data before export.
Open official repositoryA library providing evaluation metrics, test-data generation and feedback loops for RAG and other LLM applications.
Points to check: LLM-based judges can be unstable and biased; combine them with fixed test cases, human review and domain metrics.
Open official repositoryA graphical platform for building, connecting, evaluating, administering and publishing agents and agent workflows in the Microsoft ecosystem.
Points to check: Assess environment strategy, DLP, connectors, identity, logging, capacity consumption and exit options before rollout.
Open official documentationCode suggestions, chat and agentic development workflows across IDEs, the command line and GitHub.
Points to check: Set organisation policies for data, models, repositories, agent permissions, code review and measurement of actual outcomes.
Open official documentationA terminal- and IDE-based coding agent that can read repositories, edit files and run development tools.
Points to check: Constrain file, shell, network and MCP access; use approvals, isolation and human review before releasing changes.
Open official documentationDanish and international primary sources for law, public-sector practice, governance, security, risk assessment, research and skills.
The European Commission’s single entry point for the AI Act Explorer, compliance checker, frequently asked questions and Service Desk contact.
Open the EU platformGuides for citizens, businesses and public authorities on the responsible and secure use of generative AI.
Open the guidance (Danish)Guidance, decisions, an impact-assessment template and information on the AI regulatory sandbox, focused on data protection.
Open AI guidance (Danish)Knowledge, tools, a legal AI toolbox, technology radar and updates on joint-municipal AI initiatives in Denmark.
Open the knowledge centre (Danish)Ten principles and practical guidance for the safe, effective and responsible use of AI in government and public organisations.
Open the playbookA cross-sector framework and playbook for systematically managing AI risks through Govern, Map, Measure and Manage.
Points to check: AI RMF 1.0 is being revised; check the current edition and applicable local law.
Open AI RMFPractical threat and mitigation resources, including Top 10 guidance for LLM applications and agentic systems.
Open security resourcesAn annual, data-driven overview of AI’s technical progress, economy, societal impact, responsible AI and policy.
Open the AI IndexShort explainer videos covering AI agents, RAG, models, data, cloud and cybersecurity, among other topics.
Points to check: The channel is vendor-owned; use it as a technical introduction and verify product-adjacent claims across multiple sources.
Open the YouTube channelA self-paced introduction to AI, problem solving, machine learning, neural networks and societal implications.
Open the courseA searchable database of AI systems and projects in Denmark’s public sector, including work in development and discontinued projects.
Points to check: The database includes submitted information; use it for orientation and follow up with the organisation responsible for each entry.
Open the database (Danish)The taskforce’s target vision and shared ambitions for scaling AI across central, local and regional government towards 2035.
Open vision and background (Danish)The Danish Agency for Digital Government’s overview of strategies setting the political direction for responsible AI development and use in Denmark.
Open the strategy overviewThe authoritative entry point to Regulation (EU) 2024/1689, including chapters, annexes, definitions and application dates.
Points to check: Read the regulation alongside subsequent guidance, standards and national supervisory practice; obtain legal advice for concrete decisions.
Open the legal text on EUR-LexThe European Data Protection Board’s analysis of anonymity, legitimate interest and consequences of unlawfully processed personal data in AI models.
Open the EDPB opinionA lifecycle-based mapping of assets, threats and security challenges across AI systems and their supply chains.
Points to check: The report dates from 2020; pair it with current threat intelligence and ENISA’s more recent publications.
Open the ENISA reportUNESCO’s shared framework covering human rights, human oversight, fairness, the environment, data governance and other policy areas.
Points to check: The recommendation is normative and does not replace applicable law, sector requirements or a specific impact assessment.
Open the UNESCO recommendationThe first binding international treaty on AI, human rights, democracy and the rule of law, opened for signature in 2024.
Open the convention and explanationRequirements for establishing, operating, maintaining and continually improving an AI management system with responsibilities, risks and controls.
Points to check: The ISO page provides an overview; the full standard is paid. Certification does not in itself demonstrate compliance with all applicable law.
Open the official standard overviewPolicies, indicators, principles, country comparisons and tools for responsible AI, including the use of AI in government.
Open OECD.AIA companion to the NIST AI RMF covering risks and suggested actions specific to generative AI systems across the lifecycle.
Points to check: The profile is US-based and voluntary; map it to EU law, national requirements and your organisation’s own risk context.
Open the profile as PDFA template and field-by-field guide for publishing why and how a public-sector organisation uses algorithmic tools.
Open the ATRS hubA questionnaire-based tool for assessing impact and control levels for automated decision systems in Canada’s federal government.
Points to check: The tool is designed for Canadian government requirements; use its structure as inspiration and map it to Danish and European law.
Open the assessment toolExamples of training, onboarding and other AI literacy initiatives from public and private organisations, supporting learning around AI Act Article 4.
Points to check: The examples do not in themselves create a presumption of compliance; tailor content and evidence to roles, risks and context.
Open the practice repositoryThe European Commission’s forum and voluntary pledges covering AI governance, system mapping and AI literacy ahead of full AI Act application.
Points to check: Participation and pledges are voluntary and do not replace binding obligations under the AI Act.
Open the AI PactA searchable collection of AI risks drawn from many existing frameworks and organised by causal factors and risk domains.
Points to check: The database is broad, not a prioritised risk assessment for your system; assess likelihood, impact and local controls separately.
Open the risk databasePractical control questions across governance, data, performance and ongoing monitoring, designed for public bodies and auditors.
Points to check: The framework dates from 2021 and is US-based; supplement it with current technical risks and applicable Danish and European requirements.
Open the GAO frameworkOpenRouter is included as a concrete reference for the questions raised by a hosted model gateway. Inclusion is neither a recommendation nor an approval.
Provides access to many model providers behind one API, with routing, fallbacks, usage tracking and credit-based billing.
Points to check: Assess contract, data routes, logging, subprocessors, model choice, price changes and exit options before public-sector use.
Open official documentationArchived projects can remain useful as inspiration, but should not be presented as active recommendations.
LangChain’s former web interface for building and working with LangGraph agents. The repository is read-only and marked as deprecated.
Points to check: Use only as a historical or architectural reference; do not expect active maintenance or security fixes.
Open archived repository