Services

Open-source AI, adapted to how you already work.

OpenEng Labs customizes open-source tools to an enterprise’s business needs, designs and supports AI automation, and builds custom AI solutions on its own governed platform.

Every engagement is scoped to the work in front of you. There is no published price list.

  • 4 live today
  • 1 in development
  • 1 planned for Q4 2026

Engagements

Six ways to work with us.

Four are live today, one is in development, and one will open with the Enterprise and on-prem plans in Q4 2026. Each card says which.

  • Live

    AI Automation Support

    We design, build, and support recurring automations on the OpenEng AI engine. Today global loops do the repeating while connectors, cells, and MCP servers reach your systems; schedulers and headless runners will arrive with the paid plans in Q4 2026.

    What you get

    • Global loops scoped to your work items and stopping rules
    • Connectors, cells, and MCP servers wired to the systems you already run
    • Support after go-live, with every side effect passing the governed gateway
  • Live

    Custom AI Solutions

    We build agents for your business as authored configs, drawing on catalogs of 3,300+ building blocks and writing the agents and templates your domain needs. Each release is published from the Console as one immutable id@version.

    What you get

    • Agents and templates authored in the Console for your domain, with skills adopted from the catalogs
    • A per-domain diff and release notes with every version you publish
    • Secrets held as references in the config, never as values
  • Live

    Open-Source and Open-Weight Customization and Support

    We adapt the open-source tools you already run and the open-weight models you want on your own hardware. We start from Ollama, the Hugging Face Hub, ONNX exports, and MCP servers.

    What you get

    • Open-weight models matched to your hardware, with no API key
    • Models on llama.cpp, ONNX Runtime, Ollama, or an OpenAI-compatible endpoint; tools through MCP servers
    • Customizations and support for the open-source tools in your stack
  • Planned for Q4 2026

    Private and On-Prem Agent Platforms

    The local-first architecture is live today: models, threads, memory, and secrets stay on the machine, and the engine binds no inbound port. Enterprise and on-prem plans will launch in Q4 2026.

    What you get

    • The OpenEng AI engine on your own Linux x86_64 machines today
    • Kernel-enforced isolation around each shell command and the per-user daemon
    • Enterprise and on-prem plans when they launch in Q4 2026
  • 0.1 in development

    Agent Engineering with AEL

    AEL is a native language whose compiler is first-party Rust: source to binary AEL IR to a native executable, with no LLVM and no generated C. Version 0.1 is in development and qualifies linux-x86_64 alone. Its agent features are planned.

    What you get

    • Agent designs reviewed against AEL’s published specification and its compact-source limits
    • A written plan for carrying those designs into AEL once typed components and the agent runtime land
    • Progress you can check against the dated public status record at ael.openeng.ai
  • Live

    Research Collaboration

    We partner on applied agentic-systems research: architecture surveys, code-verified gap analyses, sandboxing, and human-in-the-loop design. Findings are written against real code and dated.

    What you get

    • Architecture reviews against the 2025–2026 agentic-systems survey
    • Code-verified gap analyses with prioritized findings
    • Sandbox and human-in-the-loop design work you can act on

Process

How an engagement runs.

Four steps, each closed by a written output you keep. Durations depend on the systems involved, so we scope them with you.

  1. Discover

    We map the systems in play and the automation candidates: what repeats, what it touches, and who signs off.

    OutputAutomation map

  2. Design

    We choose the models, connectors, and governance policy, and decide what runs on which of your machines.

    OutputDesign note and policy

  3. Build

    We author the configs, agents, and customizations, and verify every run on your machines before it goes live.

    OutputConfig and verified runs

  4. Operate

    We support the loops in production and iterate on them as the work changes.

    OutputRunbook and change log

What you keep

Artifacts you keep, on machines you own.

An engagement ends with configs, agents, connectors, loops, and runbooks in your hands, verified on your own machines. Nothing lives only in ours.

  • Published configs

    Each release passes the Console’s validation gate and leaves as one immutable id@version with its release notes. A later release always takes a strictly higher version.

    id@version

  • Agents, skills, and templates

    Drawn from live catalogs of 1,092 agents, 885 skills, and 754 templates published as content-as-code, with agents and templates written for your domain where the catalogs stop.

    catalog.json

  • Connectors, cells, and MCP servers

    Cells and ecosystem connectors deployed into your own cloud account and wired into your config, plus the MCP servers your tools need. Thirteen ecosystem connectors are validated live today.

    your cloud account

  • Global loops

    A global loop is a request, a source (a plain repeat or a list of work items), and a stopping rule. Ready-to-run loops range from failing tests until green to a nightly PR review sweep.

    request · source · stop

Foundations

Built on open foundations.

Four model runtimes work out of the box; vLLM, SGLang, ExLlamaV2 and V3, MLX, TensorRT-LLM, and Transformers run once their toolchain is installed. Most of the systems we connect to are already in your stack.

What an engagement starts from.

Counts read from the live catalogs on September 25, 2026: starting material, not a promise of scope.

  • 3,300+

    Building blocks across six catalogs: agents, skills, templates, models, cells, and plugins

    Source: six catalog.json files, 3,323 entries

  • 28

    On-device open-weight models, from 1B to 70B parameters

    Source: models.openeng.app

  • 1,053

    Cloud tools exposed by 550 cells for AWS, Azure, and GCP

    Source: cells.openeng.app

  • 6

    Ready-to-run global loops for tests, backlogs, tickets, inboxes, alerts, and pull requests

    Source: loops.openeng.app

Four runtimes out of the box.

  • Embedded GGUF

    In-process llama.cpp with weights fetched from the Hugging Face Hub, checksum-verified.

  • Ollama library

    Any tag in your local Ollama library; public tags are pulled for you at provisioning.

  • ONNX exports

    Standard Optimum and onnx-community exports on ONNX Runtime; TensorFlow models run via ONNX export.

  • Network endpoint

    Any OpenAI-compatible endpoint, as a secondary, per-role option.

llama.cppONNX RuntimeOllamavLLMSGLangExLlamaV2MLXTensorRT-LLMTransformersHugging Face HubMCP serversKubernetesPostgreSQLRedisMongoDBNeo4jApache KafkaAWSAzureGCP

The OpenEng platform itself is not open source; it runs open-weight models and builds on open-source tooling.

Where it runs

Your machines today. Fleet and on-prem plans next.

The OpenEng AI engine is live on Linux x86_64, and the cloud pieces are source you deploy yourself. Plans for per-fleet licensing and a whole-platform on-prem license will launch in Q4 2026.

  • Live

    Your Linux x86_64 machines

    The OpenEng AI engine runs on your own Linux x86_64 machines today, installed with one checksum-verified command. It binds no inbound port.

  • Live

    Your own cloud account

    Cells and ecosystem connectors ship as source-only Python that you deploy into your AWS, GCP, or Azure account, behind your own API gateway. No binary is shipped to your cloud.

  • Coming

    macOS and Windows engines

    Only the Linux x86_64 engine is published today. macOS and Windows builds are coming, and general availability of the platform is planned for Q4 2026.

  • Planned for Q4 2026

    Enterprise and on-prem plans

    Priced per fleet rather than per seat: one account will run your configs, schedulers, and headless runners at fleet scale, or the whole platform on-prem. Pricing will be announced at launch.

Status as of September 25, 2026. The Enterprise and on-prem plans cannot be purchased yet; we scope work on your machines now and carry it over when the plans launch.

Guarantees

Governed by construction.

The automations and agents we build for you run inside the OpenEng AI engine, so the guarantees come from its architecture, not from a promise in a contract.

Inside the OpenEng AI engine the model emits typed intents; every side effect passes one governed gateway with a fail-closed policy floor before it reaches your systems. The engine sends no telemetry.OpenEng AI engine · no telemetryModeltyped intentsGoverned gatewayfail-closed policy floorYour systemsshell · Kubernetes · data · MCP
  • One gateway, fail-closed

    Every side effect passes one gateway with a fail-closed policy floor: a configuration can add policy above it but never remove it. A build-time gate enforces the no-bypass rule.

  • Kernel sandbox by default

    Shell commands run in a kernel sandbox by default wherever the OS provides one: a read-only root, writes only inside the workspace, and the network off unless you allow it.

  • Secrets by reference only

    The Console stores ${ENV} references and never a value. Each reference resolves on your machine at run time, so no secret leaves it.

  • No telemetry

    The OpenEng AI engine sends no telemetry and binds no inbound port. Models, threads, memory, and secrets stay on the machine.

These guarantees describe the OpenEng AI engine, which is live on Linux x86_64 as of September 25, 2026. They will apply the same way under the Enterprise and on-prem plans that will launch in Q4 2026.

Questions

Before you write to us.

All questions

What services does OpenEng Labs provide?

Six: AI automation support, custom AI solutions, open-source and open-weight customization and support, private and on-prem agent platforms (the plans will launch in Q4 2026), agent engineering with AEL (version 0.1 in development; agent features planned), and research collaboration. Every engagement is scoped; there is no published price list.

Do you work with open-source tools?

Yes. We adapt the open-source tools an enterprise already runs, such as Ollama, the Hugging Face Hub, ONNX Runtime, and MCP servers, and the OpenEng AI engine runs open-weight models on your own machines. The OpenEng platform itself is not open source.

Can the platform run entirely on our own machines?

The OpenEng AI engine runs on your own Linux x86_64 machines today: models, threads, memory, and secrets stay there, it sends no telemetry, and it binds no inbound port. Running the entire platform inside your own infrastructure will come with the on-prem plan in Q4 2026.

How do we start?

Write to hello@openeng.ai or use the Contact Us page. We begin with a discovery conversation about your systems and the automation candidates, then scope a first engagement in writing.

Start a conversation about AI automation.

Tell us which systems are in play and what repeats. We will come back with a scoped first step.