Tools & Skills

1. Skills Live

Reusable instruction sets that agents (e.g. in Cursor, Claude Projects) can load to apply AAF when designing or reviewing agentic systems. Copy a skill into your project or reference it from the repo.

• Architecture Review aaf-architecture-review — Pillar checklist, design-time spec and pre-production review.
• Security aaf-security — Boundaries, tool actuation, epistemic gatekeeping, supply chain.
• Epistemic Gates aaf-epistemic-gates — When and where to place validation and authority gates.
• Cost & Context aaf-cost-context — Budgets, model routing, token economics, context discipline.
• Cross-Cutting Foundations aaf-cross-cutting — Context optimization and autonomy & outcome governance.
• ACC Implementation aaf-acc-implementation — Agent Control Contract template, placement, and usage.
• Orchestration & OCC aaf-orchestration-occ — Orchestrator Capability Contract and multi-agent governance.

Skills live in tools/skills/ on GitHub.

2. MCP Server Live

A Model Context Protocol server that exposes the framework as tools. AI assistants that support MCP can call these tools to get accurate, framework-grounded answers.

• aaf_guideGuide — Call first when unsure — ordered tool plan by intent.
• aaf_list_skillsCore — List skill ids and purposes.
• aaf_list_docsCore — List framework doc filenames and titles.
• aaf_get_docCore — Load a full doc or section by filename.
• aaf_lookupCore — Search the framework docs for a term or topic.
• aaf_checklistCore — Architecture design or review checklist.
• aaf_pillars_summaryCore — Summary of the six pillars and cross-cutting foundations.
• aaf_get_skillCore — Retrieve the full content of any skill by ID.
• aaf_list_workloadsWorkload — Catalogue Common Agentic Workloads (optional requirements ranking).
• aaf_workload_guidanceWorkload — Rubrics, dominant trades, failure modes for a workload.
• aaf_tradeoff_catalogDesign — List pillar and workload trade-off ids.
• aaf_design_questionsDesign — Design questionnaire for a given autonomy level.
• aaf_tradeoff_analysisDesign — Deterministic trade-off analysis (optional workloadId).
• aaf_generate_accDesign — Generate an Agent Control Contract from design answers.
• aaf_scaffold_specBuild — File manifest for agent code scaffold generation.
• aaf_posture_interpretReview — Interpret posture CLI JSON (run local aaf-posture first).
• aaf_review_against_accReview — Gap analysis: ACC vs posture report (CLI JSON accepted).
• aaf_pillar_guidanceCross-cutting — Targeted design guidance for a specific pillar.
• aaf_security_scanSecurity — CIA-oriented heuristic lint on snippets (optional declared_tree absence checks).

Using the MCP server in different IDEs

The hosted AAF server speaks HTTP (Streamable MCP). Editors differ in how they connect:

• Direct URL — Some clients let you register the server by URL only (no local process). Fastest when supported.
• Local bridge (mcp-remote) — Many IDEs only launch stdio MCP servers. Use npx mcp-remote … so the editor talks to localhost while the bridge forwards to agenticaf.io. Requires Node.js 18+.

Option A — Direct URL (Cursor and clients that support remote HTTP MCP). Replace the key with yours from purchase or Manage access:

{
  "mcpServers": {
    "aaf": {
      "url": "https://www.agenticaf.io/api/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_AAF_LIVE_KEY>"
      }
    }
  }
}

Option B — mcp-remote bridge (Antigravity, Claude Desktop, Windsurf, and any stdio-only MCP host). Auth is required:

{
  "mcpServers": {
    "aaf": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://www.agenticaf.io/api/mcp",
        "--transport",
        "http-first",
        "--header",
        "Authorization:${AAF_MCP_AUTHORIZATION}"
      ],
      "env": {
        "AAF_MCP_AUTHORIZATION": "Bearer <YOUR_AAF_LIVE_KEY>"
      }
    }
  }
}

After saving, restart the app or reload MCP. Ask the assistant to call aaf_guide — you should see AAF tools. Without a valid Bearer key the server returns 401. See api/README.md.

Note: Some clients rename the server (e.g. user-aaf). That is normal. Windsurf MCP availability may depend on your plan — check Codeium’s docs.

---
description: Use AAF MCP for agentic architecture decisions
alwaysApply: true
---

# AAF MCP — architecture authority

When designing, reviewing, or changing **agentic architecture**:

1. Use the connected AAF MCP tools before inventing guidance.
2. If unsure which tool: call `aaf_guide`.
3. For new systems: `aaf_list_workloads` → `aaf_workload_guidance` → design → trade-offs → ACC.
4. For codebase review: run `node tools/aaf-posture/cli.js . --format json --output ./aaf-posture.json` in the terminal, then `aaf_posture_interpret` (hosted MCP cannot scan disk).
5. Ground with `aaf_get_doc` / workload tools. Prefer Common Agentic Workloads over inventing hybrids.
6. Do not skip AAF for “quick” architecture choices.

3. AAF Posture Report Live

A CLI that scans a codebase and produces an AAF posture report: how well the project aligns with the six pillars and two cross-cutting foundations. Output is human-readable HTML/markdown and optional JSON for CI. Findings are heuristic signals — not a production certification.

CLI lives in tools/aaf-posture/ on GitHub.