Core | Illustrator | Document Tools | Enterprise
Alex ACT AI Operations is an optional provider-routing plugin. It turns a task into an evidence-backed model plan, explains provider tradeoffs, obtains explicit approval, and then calls the selected provider tools.
It is installed separately. It is not part of Core brain health and is never included in the default constellation installation. Provider API keys are optional until a selected model requires one. Microsoft Foundry uses Microsoft Entra ID through its hosted MCP and does not require an API key.
Version 0.2.1. Installation remains optional and separate from the Alex ACT brain spine.
Retain this plugin provisionally through 2026-11-11. Re-evaluate whether its provider-neutral planning and consent workflow earns continued maintenance based on real use across at least two providers, absence of duplicated provider workflows, and whether live provider evidence produces decisions that the owning tools alone would not make clear. Retire it if that evidence does not materialize or its workflow becomes redundant with a provider-owned surface.
- Decomposes a request into model-sized operations.
- Captures modality, quality, privacy, region, license, latency, and budget constraints.
- Queries current provider tools instead of relying on a static model list.
- Removes infeasible candidates and reports unknown evidence honestly.
- Produces an executable plan with a primary choice and fallbacks.
- Displays provider, model, data boundary, and cost status for approval.
- Guides provider-native login, the host environment, or approved secret storage only for the selected provider/model when access is required.
- Executes approved calls, tracks asynchronous work, and records provenance.
| Provider | Connection | Initial role |
|---|---|---|
| Microsoft Foundry | First-party hosted MCP (preview) | Model discovery, metadata, and Foundry project tools |
| Hugging Face | First-party hosted MCP | Hub search, models, datasets, Spaces, jobs, sandboxes |
| ElevenLabs | Exact first-party local MCP 0.12.2 | Models, voices, speech, transcription, audio, music, agents |
Provider tools remain owned by their publishers. This plugin owns the workflow that decides when and how to use them.
| Surface | Role |
|---|---|
model-router |
Advisory planning and explainable provider/model selection |
model-task-execution |
Consent-gated execution and evidence capture |
setup-ai-operations |
Provider authentication and local runtime setup |
/alex-act-ai-operations choose-model |
Create an executable plan without running it |
/alex-act-ai-operations execute-model-task |
Review, approve, and execute a plan |
/alex-act-ai-operations setup-ai-operations |
Configure provider access |
Install from the Alex ACT Mall:
copilot plugin install alex-act-ai-operations@alex-mallThis is a separate action from installing Alex ACT Core or any optional specialization setup. See INSTALL.md for provider configuration.
Onboarding ends when installation is verified. During onboarding, do not ask for API keys, tokens, or other credentials, and do not initiate provider login. Defer provider login and authentication until an approved execution plan selects a provider and execution requires access.
Discovery and planning do not authorize execution. Before a paid or data-transmitting call, the user must see and approve:
- Provider and model
- Operation and expected output
- Data sent outside the workspace
- Region, retention, and license evidence when available
- Estimated cost or an explicit statement that cost is unknown
- Maximum approved cost
- Approved fallbacks
Changing the provider, model, data boundary, or cost ceiling invalidates prior approval and requires renewed consent.
- Provider catalogs and prices change. The router must query live evidence.
- Authentication remains provider-specific.
- Cross-provider benchmarks are not automatically comparable.
- A recommendation may be a set of Pareto choices rather than one winner.
- No paid provider canary runs as part of source validation.
The source boundary is defined by Alex ACT Steward ADR-026. Releases use semantic versioning and Mall origin delivery from immutable source tags.