282 orchestrators
Multi-level planners that decompose a goal and delegate. Drop one in and you inherit a whole team's worth of routing.
The ecosystem · compose, don't codeDon't start from a blank prompt.
Don't start from a blank prompt.
Start from 3,300+ building blocks.
Browse and snap together agents, skills, templates, models, plugins, cells, and loops — then publish a config your engine just runs. Everything is content-as-code on a CDN, consumed read-only. You assemble; you never start from zero.
1,092Agents
885Skills
754Templates
28On-device models
550Cells
44Loop strategies
2,433Engine tests passing
The catalogs
Eight catalogs. One composable platform.
Each catalog is a live, deep-linkable gallery with a true count. Pick what you need, drop it into a config, and the engine resolves it by name — no code to write.
Agents1,092
A 3-level org chart — 282 orchestrators delegating to 810 single-purpose agents.
agents.openeng.ai
Skills885
Named, invocable playbooks the engine runs on demand, across 18 categories.
skills.openeng.ai
Templates754
Prompt blueprints prepended above your persona, across 28 segments.
templates.openeng.ai
Cells550
AWS / Azure / GCP connectors — read-only by default, inside your own IAM.
cells.openeng.ai
Models28
On-device open-weight models, 1.3 GB → 70B — zero API keys.
models.openeng.ai
Plugins14
External decision, approval, and tooling endpoints layered above the security floor.
plugins.openeng.ai
Loops44
Agent loop strategies — react, reflection, plan-execute-replan, durable, and more.
loops.openeng.ai
LibraryNew
Shared connectors, bridges, and tools for the OpenEng product suite.
library.openeng.aiSoon
Counts are the verified, current platform numbers. Each tile opens its catalog subdomain in a new tab.
Agents · 1,092
A 3-level agent org chart — not a flat list of prompts.
Orchestrators plan and delegate; single-purpose agents do one thing well. Compose a top-level orchestrator and it routes sub-tasks to specialists, several levels deep.
282 orchestratorsplan & delegate
ResearchCodeReviewDataCloudDocs
810 single-purpose agents across 35 categories
810 single-purpose agents
Focused workers — research, code, review, data, cloud, docs — across 35 categories. Each is small, sharp, and recombinable.
Cells · 550
550 cloud cells, exposing 1,053 tools — inside your own IAM.
Cells give agents live signal from AWS, Azure, and GCP. Read-only by default, running under your own credentials. Write actions are opt-in and always pass the guardrail first.
AWS · Azure · GCP
One connector model across all three clouds — list, describe, query, and inspect resources from the same agent config.
- 550 cells
- 1,053 tools exposed
Read-only by default
Every cell starts read-only. Agents observe before they ever act, so an exploratory run can't mutate a thing.
Your IAM, your keys
Cells resolve credentials from
${ENV} in your local environment — secrets never touch the cloud control plane.Models · 28
28 on-device open-weight models, auto-matched to your hardware.
Open-weight models from 1.3 GB to a 70B — pulled from Hugging Face, the Ollama library, ModelScope, or a direct URL, in any major weight format. Zero API keys. The engine picks the right size for the machine it's on and runs them isolated in a sandbox.
Hardware-aware selection
The engine profiles your RAM and GPU and matches each role — chat, coding, planner, reviewer — to a model that fits. Runs fully offline.
Any format, the runtime that fits
GGUF, Safetensors, GPTQ/AWQ, EXL2/EXL3, MLX, ONNX, or TensorRT — the format determines which runtimes can run it, and the live compatibility matrix is published on the models catalog. All through one role system, from a tiny laptop model to a 70B.
- Open-weight, zero API keys
- Optional Metal / CUDA / Vulkan
Or bring your own
Point a role at any network endpoint — your own OpenAI-compatible API — and the engine routes to it. The same config, your cloud.
New catalogs
Loops and the Library — strategy and shared plumbing.
Two new catalogs round out the platform: agent loop strategies you select per task, and a shared Library of connectors and bridges for the OpenEng product suite.
Loops · 44
44 agent loop strategies
React, reflection, plan-execute-replan, durable, and more — swap the loop the engine runs without touching your agent. The control strategy becomes a composable choice, not hard-coded behavior.
Library · new
Shared connectors & bridges
A common Library of connectors, bridges, and tools shared across the OpenEng product suite — the reusable plumbing that lets the surfaces speak the same protocols.
library.openeng.ai · launching soon
The workflow
Compose your own agent. Freeze a version. Run it on your engine.
Assembling is the whole job — there's no glue code. You pick building blocks, publish an immutable version, and every machine you sign in to runs exactly that.
step 1 · compose
Snap blocks together
In the Console, combine an orchestrator, the agents and skills it should call, a model per role, the cells it can read, and a loop strategy. No prompt engineering from scratch.
step 2 · freeze
Publish id@version
Freeze the whole composition into an immutable, SemVer-versioned id@version. Publish-once with a strictly-higher version — you can always see exactly what will run.
step 3 · run
Run on your engine
Reference the version by name and your local engine pulls and runs it — same config on every machine you sign in to, no drift.
Run a frozen composition
openeng run my-research-team@2.1.0How it's served
Everything is content-as-code, consumed read-only.
The catalogs aren't a closed marketplace. Every entry is a published artifact on a CDN that the engine and Console only read — never a runtime your data passes through.
Published to a CDN
Each catalog generates from definitions into a static catalog.json on its own subdomain — versioned, cacheable, and independent of any live service.
Read-only consumption
The engine and Console pull catalog entries and resolve them locally. Nothing in the catalogs executes on a server; nothing of yours is uploaded to fetch them.
The Console mirrors it
The Console reads the same CDN catalogs to show you what's available — a mirror of the content, not a source you depend on at runtime.
Stop writing prompts from scratch. Start composing.
Open the Web Console, snap together your first agent from the catalogs, and freeze a version your engine just runs.