# Integrations — What Cromus Works With

Cromus fits where you already build. It is the pre-execution layer for AI workflows: it governs and scores your workflows before tokens are spent, then compiles them into whatever runtime you ship to. Because the governance lives in portable open specs, your workflows stay yours — moving across agents, frameworks, and providers without a rewrite.

Below: the platforms Cromus compiles to, the tools it connects with, and the providers it reasons about. Every item here is a shipped capability.

## At a glance

- **Agent runtimes & platforms** — compile one governed skill to six targets.
- **Model context (MCP)** — a hosted, zero-LLM MCP server for Claude, Cowork, and Codex.
- **Model providers** — 62 active models across 11 providers in a verified registry.
- **Workflow frameworks** — export to LangGraph, CrewAI, AutoGen, n8n, and more.
- **Coding-agent platforms** — credit-pool governance for Replit Agent, Cursor, Lovable, v0, Bolt.
- **Automation platforms** — trigger-to-action templates for Zapier, Make, n8n.
- **Open specs** — SKILL.md, ETHOS.md, MEMORY.md, the portable contract any compliant runtime can read.

## Compile targets: one governed skill, six platforms

Skill Studio's Platform Selector compiles a single Cromus SKILL.md to any of six targets, each with platform-specific validation. The governance travels with the artifact; Cromus never sits between your agent and the model.

- **Cromus Native** — the default Cromus SKILL.md format.
- **Replit Agent** — emits the `.agents/skills/` layout aligned to Replit's current mode taxonomy (Lite / Economy / Power), with governance-lock awareness and a copyable compiled config.
- **AgentSkills.io** — open-spec registry publishing, with compatibility, metadata, and allowed-tools validation.
- **OpenClaw** — ClawHub community registry, with semver, security, license, and credential-scan validation, plus a preview card.
- **Claude Managed Agents** — Anthropic Agent endpoint config (the `managed-agents` beta), with tool whitelisting and a JSON preview matching Anthropic's Agent endpoint shape.
- **Google AI Studio** — Gemini API config plus the Interactions API typed-steps template, with Gemini-family model validation and Google tool whitelisting.

## Model context: the hosted MCP server

Cromus runs a hosted, streamable MCP server with six deterministic tools and **zero LLM calls in the diagnostic path**. It plugs Cromus's scoring and governance directly into the tools your team already uses:

- **Claude / Claude Cowork / Claude Code** — install via the Cowork connector page; score and validate skills without leaving the assistant.
- **OpenAI Codex CLI** — install via the Codex plugin page.
- **Any MCP client** — connect with OAuth 2.1 (PKCE-mandatory) for consumer clients, or a legacy API key. Three scopes (`read`, `score`, `write`) gate exactly what a connected client can do.

The same scoring engine powers the MCP tools, the public validator, and the in-app readiness tile, so a skill scores identically everywhere.

## Model providers: 62 models, 11 providers, verified pricing

Cromus reasons about cost across a verified, admin-maintained registry of 62 active models (81 total entries) from **OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Meta, Cohere, Amazon, Microsoft, and Qwen.** Real per-token pricing, context windows, and quality tiers, with a "last verified" timestamp.

You reach those providers through whichever channels your team has enabled, across four ecosystem tiers:

| Tier | Examples | Setup |
|------|----------|-------|
| Mainstream | OpenAI, Anthropic, Google AI | Paste an API key |
| Managed | Mistral, Cohere, AWS Bedrock, Google Vertex AI | Cloud account + IAM |
| Aggregator | OpenRouter, Together | One key, many providers |
| Infra | Ollama, vLLM, your own GPUs | Self-hosted, compute-billed |

Provider Scope lets a workspace declare which providers it's actually willing to run, and every cost estimate and recommendation honors that declaration.

## Workflow frameworks & automation

- **Open Workflow Ecosystem** — a registry of 24 open-source AI tools across 7 categories (agent frameworks, orchestration, local inference, workflow engines, search & retrieval, vector databases, developer tools), with export to **LangGraph, CrewAI, AutoGen, n8n,** and more.
- **Automation platforms** — a dedicated skill category for trigger-to-action automations on **Zapier, Make, and n8n,** with platform detection built into the SOP compiler and 10 seeded templates.
- **Coding-agent platforms** — Layer 3 platform governance models credit-pool platforms (**Replit Agent, Cursor, Lovable, v0, Bolt**) so you can govern budgets where billing is opaque, with honest "prediction unavailable" handling rather than fabricated numbers.

## The agentic cost stack

Beyond the inference meter, Cromus's simulator models vendor agentic costs where they exist: **Salesforce** Flex Credits / Agentic Work Units and **Microsoft Copilot** credits (application meter), and **ServiceNow** Action Fabric (governance meter). Vendor rates that aren't yet verified are shown as "Incomplete" rather than guessed, so a total never silently misleads.

## Open specs: the portable contract

The deepest integration isn't a connector at all, it's a standard. Cromus authors and validates three open specs, each with a free public validator:

- **SKILL.md** — capability definition (inputs, outputs, steps).
- **ETHOS.md** — behavioral governance (what an agent should and shouldn't do).
- **MEMORY.md** — portable memory (identity, context, compression, handoff).

Any runtime that reads these specs works with Cromus by definition. That's what makes a governed workflow portable: the rules travel inside the artifact, into any compliant runtime.

## Where Cromus sits

Cromus is **upstream of your runtime.** It compiles, scores, and governs before execution, then hands a portable, validated artifact to whatever runs it. It does not proxy your model calls, add latency, or sit in the hot path. `scope: "enforce"` is rejected by design; runtime enforcement belongs to your runtime.

Don't see your tool? Because Cromus governs over open specs, many runtimes work with it without a formal connector. If you'd like to confirm a specific setup, contact us at https://cromus.ai/contact.