Managed compute
Choose CPU or GPU resources for a private endpoint. Review the price before you start, and set an idle timeout so an experiment does not bill all night.
Hosted endpoint · scale to zero when idleExplore endpoints
The platform
Everything the work needs, in one place.
Pick an area to see what it does. Each one is a real screen in the product, not a capability claim.
Go from a prompt to a reviewed, versioned, running system.
- ChatCompare answers across models, search your own knowledge, and build a small app from the conversation.
- Projects & IDEA browser editor with a terminal and checkpoints, where a coding agent works on your repository under review.
- AgentsGive an agent its prompt, its tools and its knowledge, then inspect every run it produces.
- WorkflowsChain agents into asynchronous pipelines, with retries and an approval step where a human decides.
- BlueprintsPackage an agent or workflow as a portable version you can review, publish and roll back.
- Prompt LibraryKeep reviewed prompts in one place so teams reuse what already works.
Model catalogue
Open-weight models your infrastructure can actually run.
55 releases, each importable through the same path. Curated entries carry their verified size and context.
Serving engines
- vLLM
- SGLang
- TGI
- Ollama
55 models
- Moonshot AIVerifiedKimi K32.8T MoE, 104B active, 1M contextKimi K3 LicenseOpen the upstream weights on Hugging Face
- Alibaba QwenVerifiedQwen3.8 Max2.4T MoE, 95B active, 262K contextQwen3.8-Max LicenseOpen the upstream weights on Hugging Face
- DeepSeekVerifiedDeepSeek V4 Pro1.6T MoE, 49B active, 1M contextMITOpen the upstream weights on Hugging Face
- Z.aiVerifiedGLM-5.2753B MoE, 40B active, 1M contextMITOpen the upstream weights on Hugging Face
- MiniMaxVerifiedMiniMax-M3428B MoE, 23B active, 1M contextMiniMax Community LicenseOpen the upstream weights on Hugging Face
- DeepSeekVerifiedDeepSeek V4 Flash284B MoE, 13B active, 1M contextMITOpen the upstream weights on Hugging Face
Model links open the upstream weights on Hugging Face. Being listed here does not mean a hosted endpoint is already running.
Deployment
Run it where the workload belongs.
Start on managed compute to try something out, then move it to your own hardware. The deployment artifact is the same either way.
Your own infrastructure
Download a portable bundle, read its configuration, and apply it to a Kubernetes cluster, a Compose host or a bare Linux server. Nothing is hidden in a vendor console.
Bundle · Kubernetes · Docker · LinuxExplore bundles
Supported targets
- Kubernetes
- Docker
- Google Cloud
- Microsoft Azure
- AWS
- OVHcloud
- Scaleway
- Bare metal Linux
Getting started
Three steps, and the tools you already use.
The setup script finds the coding tools installed on your machine and writes the gateway endpoint, token and model profiles into each one.
- 01
Read how a run is put together
Every agent run keeps its trace, its token usage and its outcome, and you can read the whole thing before it reaches anyone else.
See a run - 02
Point your editor at the gateway
One command writes the endpoint, token and model profiles into the coding tools already on your machine. Scoped keys decide what each tool may reach.
Issue a key - 03
Decide where the model runs
Bring a model you already run, or let Edelux start one on managed compute. Either way the endpoint is the same shape.
Start an endpoint
8 tools configured
- Oh My Pi~/.omp/agent/models.yml
- Claude Code~/.claude.json
- Codex CLI~/.codex/config.json
- OpenCode~/.config/opencode/opencode.json
- Edecode~/.config/edecode/edecode.json
- Hermes Agent~/.hermes/config.yaml
- Continue.dev~/.continue/config.json
- Aider~/.aider.conf.yml
Pricing
Pay for the month, spend the credit on any model.
A personal plan buys monthly inference credit for Build and Chat. Unused credit does not roll over, and renewal is manual.
Starter
Monthly€10/ month
Inference credit
€20
For lighter Build and Chat use.
- €20 of inference credit each paid month
- Build and Chat across the model catalogue
- Manual crypto renewal, no automatic charges
Plus
Monthly€100/ month
Inference credit
€200
For more frequent Build and Chat use.
- €200 of inference credit each paid month
- Build and Chat across the model catalogue
- Manual crypto renewal, no automatic charges
- API reference rate
- 0.25× the official published price
- Marketplace fee
- 8% of a settled sale
- Billing interval
- Per second, while a workload runs
Business plans, reserved capacity and wholesale commitments are quoted per contract. Talk to us.
Governance
Give teams access, keep the record.
Control who can run workloads, what they may spend, where they may connect, and which steps need a human decision.
Scope every surface
Roles and scoped API keys decide who reaches which workspace, checked on each request rather than in the interface.
Manage team & rolesCap the spend
Budgets, quotas and per-team chargebacks, with alerts that fire before a limit is passed.
Set budgetsKeep a human in the loop
Require an approval inside a workflow, assign the reviewer, and store the decision with the run.
Add an approvalDetect tampering
Consequential actions append to a per-organisation SHA-256 hash chain that a verification endpoint can replay.
Open the audit trailClassify under the AI Act
Register a system, get its risk tier and obligations, and track the reporting windows the regulation sets.
Register a systemFence the network
An egress policy restricts the domains a workload may reach, in audit mode first, then enforced.
Configure egress
Market
Spare capacity is an asset, not a sunk cost.
Most clusters are either saturated or half empty, rarely correctly sized. Foundry puts both sides in the same book and settles the result in your own ledger.
- models listed
- 55
- serving engines
- 4
- surfaces, one ledger
- 29
Buy what you lack
Discover MCP servers, blueprints and compute packs published by other organisations, with the contract terms attached to each listing.
Browse the marketplaceSell what sits idle
List spare capacity on the order book. Matching, commission and settlement are recorded against your own wallet, so the position is inspectable.
Open the exchangeServe your own clients
Resell inference behind per-client keys, or sign take-or-pay wholesale commitments whose endpoints are probed against the contract.
See resale
Start here
Open a workspace, or connect your editor first.
Signing in gives you the assistant, the agents, the endpoints and the cost record. Connecting an editor first is the smaller step, and it needs nothing but a key.