The control plane for AI agents just got a local address — and it's bringing 1,987 pre-built skills you can audit before your agent touches production.
The Summary
- AAS Core v15.5.1 launches as a local, agent-first control plane with a catalog of 1,987+ pre-built skills that coding agents can search, select, and validate before execution
- The architecture is read-only by design: agents browse and choose, but AAS validates selections locally and generates an immutable plan before any code changes
- Ships with CLI tooling, local MCP server, and a browser-based Workbench for human review of agent-selected skill stacks
The Signal
Here's what makes this different from the dozens of agent framework repos flooding GitHub: AAS Core isn't trying to be the agent. It's infrastructure for agents that already exist.
The workflow inverts the typical pattern. Instead of a framework that wraps your agent's behavior, AAS Core gives Claude, Codex, or any other coding agent a local catalog to search, a validation layer to check its choices, and a planning stage before touching your codebase. The agent inspects your project, searches the full catalog through a local MCP server, picks exact skills by ID, and AAS validates the stack in memory. Only after validation does anything persist as an aas-stack.json file.
The read-only constraint matters. AAS Core doesn't rank, recommend, or execute. It's a discovery and validation layer, not a decision layer. The agent owns the selection. AAS just makes sure the pieces fit together before the plan ships to a human for review.
"Agent-owned selection with human-reviewable plans before any target change."
The 1,987 skills span the full stack of what coding agents actually do: Git operations, Docker workflows, cloud deployment patterns, testing frameworks, CI/CD integrations. Each skill is a reusable SKILL.md playbook. Think of them as Lego blocks with assembly instructions. The catalog is local, the MCP server runs on stdio, and the entire validation happens in memory. No API calls. No external dependencies for the core loop.
Key architectural pieces:
- Local catalog search via MCP server (read-only stdio)
- Agent-owned skill selection by exact ID
- In-memory stack validation through compose_stack tool
- Immutable plan generation before execution
- Optional Workbench UI for human review
This sits at the intersection of two trends. First, the MCP (Model Context Protocol) standard is turning into plumbing for agent tool discovery. Second, enterprises are waking up to the fact that agents making unchecked code changes is a governance nightmare. AAS Core threads that needle: give agents rich tooling, but make every decision auditable before it runs.
The project explicitly marks apply and recovery as experimental and outside the supported path. Translation: you get the planning and validation now. Autonomous execution is your problem. That's the right call for a v15 release trying to be production-adjacent.
The Implication
If you're building coding agents or trying to deploy them inside a company with actual compliance requirements, this is worth a fork and a weekend. The value isn't in the 1,987 skills, though those are useful. The value is in the pattern: local catalog, agent choice, validation layer, immutable plan, human gate.
Watch for other agent frameworks to adopt this architecture. The read-only MCP pattern with local validation solves the "how do we let agents be useful without letting them be dangerous" problem that's blocking enterprise adoption. This repo might be the first clean implementation of that pattern in the wild.