Back to blog

Engineering · Jun 17, 2026

MCP Spotlight: Agent Config — The Universal AI Agent OS With 227 Skills, 79 Rules, 150 Commands & 6 Role-Shaped Profiles

Agent Config is the first AI Agent OS: 227 skills, 79 rules, 150 commands, 6 role-shaped profiles (developer, founder, creator, agency, finance, ops), projected into 12 host tools (Claude Code, Cursor, Augment, Windsurf, Copilot) from one npm package.

MCP ServerAgent ConfigAI Agent OSMulti-Tool ProjectionSkills & RulesAI Agents

MCP Spotlight: Agent Config — The Universal AI Agent OS With 227 Skills, 79 Rules, 150 Commands & 6 Role-Shaped Profiles

Server: @event4u/agent-config by event4u Bundle size: 227 skills · 79 rules · 150 commands · 78 guidelines · 24 personas · 5 advisors Targets: Claude Code, Augment, Cursor, Copilot, Windsurf, Cline, Gemini CLI, Codex, Aider, Continue, Claude Desktop, Roo Code MCP Tracker: glama.ai/mcp/servers/event4u-app/agent-config Docs: github.com/event4u-app/agent-config License: MIT

The MCP ecosystem has a discoverability problem. There are 37,000+ servers in the Glama registry. Most do one thing well. None of them give your agent a general operating system — a structured set of skills, rules, commands, and personas that turn a raw LLM into a reliable team member with predictable behavior across tasks.

Agent Config takes the opposite bet: one content layer, six role-shaped profiles, twelve host tools, zero model lock-in. 227 skills, 79 rules, 150 commands, 78 guidelines, 24 personas, 5 advisors — distributed as a single npm package and projected into the native config format of every supported AI tool via a condensation pipeline. The MCP server itself is a read-only bridge that serves prompts and resources from a release-pinned content bundle.

This is the "I just want my agent to be a good engineer / founder / writer / agency / finance lead / ops" product. And it's the only one that ships the entire stack in one place.

The Six Profiles: One Bundle, Six Entry Points

The architectural innovation: Agent Config doesn't ship one generic "developer profile" and ask you to customize. It ships six opinionated, role-shaped profiles — each with its own first commands, first skills, and use case:

ProfileAudienceFirst CommandsFirst Skills
👩‍💻 developerIC engineer/implement-ticket · /work · /review-changes · /fix · /commitdeveloper-like-execution · verify-completion-evidence · minimal-safe-diff · systematic-debugging · test-driven-development
✍️ content_creatorWriters, ghostwriters, marketers/work · /post-as · /ghostwriter · /optimize-prompt · /video:from-script · /video:storyboardvoice-and-tone-design · messaging-architecture · editorial-calendar · release-comms · character-consistency
🚀 founderSolo / early-stage founder/work · /feature · /challenge-me · /councilrefine-prompt · rice-prioritization · vision-articulation · fundraising-narrative · runway-cognition
🏛 agencyMulti-client delivery shop/work · /implement-ticket · /refine-ticket · /feature · /roadmapdoc-coauthoring · decision-record · refine-ticket · estimate-ticket · perf-feedback-craft
💼 financeCFO / fractional finance / FP&A/work · /council · /challenge-medcf-modeling · forecasting · scenario-modeling · unit-economics-modeling · runway-cognition
🛡 opsRevOps, support, SRE-adjacent/work · /threat-model · /review-changes · /fixincident-commander · dashboard-design · logging-monitoring · threat-modeling · launch-readiness

The profile is set by agent-config setup which writes profile.id to .agent-settings.yml. The wizard is browser-based (8-option role question), runs once, and persists. Beyond the six software profiles, user-types/ includes galabau, metalworking, and truck — a clear signal that Agent Config is being designed for every knowledge worker, not just developers.

The headline: "Get a focused command set, not a 500-artefact dump." Most "agent toolkits" dump every command into your context. Agent Config projects only the profile's commands into the active config, then layers additional packs on top.

Workflows, Not Raw Commands

You don't memorize 150 commands. You run a work journey. Four flows span the developer story end-to-end:

FlowStart WithThe Journey
🔍 Discovery/feature:plan · /researchexplore → plan → estimate → refine, before building
🔨 Implementation/work · /implement-ticketplan → implement → verify → commit
🔎 Review/review-changes · /judgeself-review → judge → quality-fix → threat-model
🚢 Delivery/commit · /pr:createcommit in chunks → open PR → answer review

Each flow names the command you type to start and the skills it composes. The pattern: the user invokes a high-level intent (/work "implement OAuth2 PKCE flow"), and the agent orchestrates the lower-level skills (developer-like-executionsystematic-debuggingtest-driven-developmentverify-completion-evidence) without the user having to know the skill names.

This is the same model as a good internal CLI: you say what you want, the tool figures out the steps. The difference is the steps are LLM-callable skills, not shell commands.

The Multi-Tool Projection: 12 Hosts, 1 Source

The killer feature: one content layer, projected into the native config format of 12 different AI hosts:

--tools=claude-code,cursor,augment,windsurf,cline,gemini-cli,copilot,roocode,aider,codex,claude-desktop,continue

Pick any subset. The condensation pipeline translates:

  • Claude Code~/.claude/commands/*.md, ~/.claude/skills/*, hooks in settings.json
  • Cursor~/.cursor/rules/*.mdc, command files in ~/.cursor/commands/
  • Augment.augment/rules/, context files
  • Windsurf~/.codeium/windsurf/memories/, rule files
  • Cline.clinerules/, command instructions
  • Gemini CLI~/.gemini/commands/, settings
  • Copilot.github/copilot-instructions.md, settings
  • Roo Code / Aider / Codex / Claude Desktop / Continue → each in their own native format

The same content, twelve different on-disk shapes. When you init once, the bundle projects to all selected hosts. Update once, every host gets the new version. CI smoke tests (smoke-public-install.yml) run across 3 OSes × 2 Node versions to confirm the projection works.

The scope_guard pre-flight refuses a second, conflicting scope on the same machine — no two installs fighting over ~/.claude/. Cleanup script provided for when you need to recover.

The MCP Server: Read-Only Bridge

The Agent Config MCP server is intentionally read-only:

"Read-only MCP bridge serves prompts and resources from a release-pinned content bundle."

The MCP exposes:

  • list_commands — every command in the current profile
  • list_skills — every skill available to the active agent
  • lint_skills — frontmatter validation against JSON-Schema
  • (more tools for skill/rules/guidelines lookup)

The read-only design is deliberate. Agent Config is a content layer, not an agent runtime. The agent loop, the LLM dispatcher, and the tool orchestration stay with the host tool (Claude Code, Cursor, etc.). Agent Config is the playbook and style guide for those tools, not a replacement.

This is the right scope boundary. Too many "agent toolkits" try to be a full runtime and end up competing with their host. Agent Config stays in the content lane and lets the host tool handle the loop.

Audit-Disciplined by Construction

For teams that care about governance, the audit story is real:

"Every memory consult, decision key, and hook concern lands in agents/runtime/state/ so you can replay it."

The agents/runtime/state/ directory is the source of truth for:

  • Which skills were loaded for a given task
  • Which rules were consulted
  • Which commands were invoked
  • Which memory entries were added or promoted

Replaying an agent's work means opening the state directory and walking the events. The four core invariants — citation discipline, scope guards, scope safety, cross-tool projection — are documented in the Core Principles section and enforced by hooks.

For teams shipping production agents, this is the "we can show you exactly what the agent did and why" story — without the agent having to write a separate audit log. It's built into the runtime.

Facio Integration

npx -y @event4u/agent-config init
# Pick profile: developer
# Pick tools: claude-code
# Run: /work "implement OAuth2 PKCE flow in the auth module"

The agent runs the work flow inside Facio's HITL-first runtime. Each skill citation, each rule consult, each command invocation flows through Facio's audit trail. The MCP tools (list_skills, list_commands, lint_skills) are exposed as read-only tools that Facio can call without destructive risk.

For HITL gating, the destructive operations are explicit in the profile skills:

  • /commit — soft confirm (review the staged diff)
  • /pr:create — soft confirm (review the title and body)
  • /fix — autonomous (read-only investigation)
  • /work — autonomous (multi-step implementation)

The agent runs the high-level flow autonomously, surfaces the destructive step for review, and Facio captures the full chain.

Quickstart

# 1. Install — browser wizard launches automatically
npx -y @event4u/agent-config init

# 2. Pick profile + tools in the wizard, click Finish
#    Writes ~/.event4u/agent-config/, ~/.claude/, ~/.cursor/, etc.

# 3. First real task — agent refines, plans, verifies
/work "implement input validation on the /api/orders endpoint"

# Or pick a specific AI:
npx -y @event4u/agent-config init --tools=claude-code,cursor

# Headless / CI:
CI=1 npx -y @event4u/agent-config init \
  --profile=developer \
  --tools=claude-code,cursor \
  --no-ui

# Verify hook coverage:
npx @event4u/agent-config hooks:status --strict

Use Cases

Solo founder shipping fast: agent-config setup → profile=founder. Run /work "draft a Q3 OKR doc with measurable outcomes", /feature "user feedback widget", /council "should we raise now or in 6 months?". The skills layer brings the right framework (RICE, runway cognition, fundraising narrative) without the founder having to remember the methodology names.

Agency team with 5 clients: agent-config setup → profile=agency. Run /refine-ticket "build a campaign landing page" to turn client-speak into a ticket, /implement-ticket to build, /estimate-ticket to scope, /roadmap to plan the quarter. The decision-record and doc-coauthoring skills ensure every client interaction has a written record.

CFO running scenarios: agent-config setup → profile=finance. Run /work "model the impact of a 20% price increase on MRR", /council "should we extend runway or invest in growth?". The dcf-modeling, scenario-modeling, and unit-economics-modeling skills bring the right analytical framework.

Content team with consistent voice: agent-config setup → profile=content_creator. Run /post-as to write in a specific persona's voice, /ghostwriter to draft a piece in someone else's style, /optimize-prompt to refine the brief. The voice-and-tone-design and character-consistency skills enforce brand voice across all output.

Developer on a regulated codebase: agent-config setup → profile=developer. Run /work "implement the data export endpoint", /review-changes before commit, /fix when something breaks. The verify-completion-evidence, test-driven-development, and systematic-debugging skills enforce engineering rigor.

Bottom Line

Agent Config is the first "AI Agent OS" that takes content layer portability seriously — 227 skills, 79 rules, 150 commands, 6 role-shaped profiles, 12 host tools, one source of truth. The MCP server is a read-only bridge, the install is a 5-minute browser wizard, the audit discipline is built in.

The profile architecture is the right pattern. Most agent toolkits dump everything on the user. Agent Config projects only what's relevant to your role, then layers packs on top. The multi-host projection is the engineering pattern every "AI for X" team should copy: write the skill once, ship to 12 tools.

npx -y @event4u/agent-config init and pick your profile. Your agent becomes a team member.


MCP Spotlight is a series covering servers that give AI agents real capabilities. Every server is evaluated for architecture, content quality, and integration fit with Facio's HITL-first agent runtime.