Myra AI Workspace vs Scaleway

Colour-coded multi-model chat and projects

Employees work in a familiar chat and shared project workspace with model switching. Green marks Myra-hosted EU models, orange marks frontier models with PII masking, and red marks unmasked frontier use where administrators permit it.

PII masking and two-tier guardrails

Before prompts reach external providers, configurable guardrails can detect and replace personal data (names, emails, IBANs and similar) and block jailbreak or injection patterns. Tier 1 runs in-process; tier 2 optional sidecars add NLP PII and prompt-guard checks. Masking is skipped on wholly Myra-local legs because data does not leave Myra.

Myra EU model fleet with residency controls

Open-weight models such as Qwen (docs cite the current default Myra on-prem chat model) run on Myra's EU fleet without a BYOK key. Gateways can enable EU-region routing that returns HTTP 403 data_residency_blocked for non-EU routes, including post-dispatch host re-checks when the Myra router would overflow off-EU.

OpenAI-compatible AI gateway and BYOK

A unified compatibility endpoint resolves twenty-plus providers (OpenAI, Anthropic, Gemini/Vertex, Azure OpenAI, AWS Bedrock, Mistral and others). Admins store provider keys in an AES-256 vault, set fallbacks, exact-match response caching, and organisation/team/user budgets with hard enforcement.

Agents, workflows and MCP connectors

Documented agents support versioning, scheduling, webhooks and required approval before tool runs. Workflows include human-approval steps. MCP connectors attach external tools and data sources; the connector catalog documents Microsoft SharePoint, OneDrive and Graph with Entra OAuth.

Enterprise identity: OIDC, SAML and SCIM

SSO authenticates existing users via OIDC or SAML 2.0 (including Microsoft Entra ID) with domain allowlists. SCIM 2.0 on the admin API provisions and deactivates users and groups. Access can be limited by IP allowlist (IPv4).

Elastic Metal Bare Metal Servers

Elastic Metal provides on-demand bare metal instances like Aluminium, Iridium, and GPU-equipped Titanium models. Deployment takes minutes via API, with models tailored for general compute, storage-intensive tasks, or AI. Users avoid virtualization overhead for better performance in databases or gaming servers. Models start around €0.02 per hour, with full root access and custom images.

GPU Instances and AI Tools

High-end GPUs such as NVIDIA L40S, H100, and B300-SXM support AI training and inference. Custom GPU clusters scale for large models. Additional AI services include generative APIs and managed inference. These integrate with Kubernetes for orchestration, serving needs from startups to research labs without vendor lock-in.

Managed Kubernetes, Databases, and Storage

Kapsule offers managed Kubernetes clusters, while Kosmos enables multi-cloud setups. Databases cover PostgreSQL, MySQL, Redis, MongoDB, and ClickHouse as managed services. Storage options include object (S3-compatible), block, and Glacier for archiving. Serverless Functions, Containers, and Jobs handle event-driven workloads. All scale automatically.

Renewable Data Centers and Networking

Nine data centers consume renewable energy, hitting a 1.15 PUE. Locations ensure low latency in Paris (multi-AZ), Amsterdam, and Warsaw. Networking features private VPCs, load balancers, WAF, and global anycast IPs. Tools like CLI and Terraform simplify management.