Example

MCP Grounding

How MCP-aware agents should resolve PersonaKit context before choosing host-local tools.

Persona	mcp-grounded-agent
Directive	resolve-before-tools
Kits	mcp-grounding-guardrails
Authorized Skills	read-only-mcp-grounding
Forbidden Skills	autonomous-agent-loop, mcp-writeback

Adopt This Root

Copy the closest example root into your project.
Rename the copied folder to .personakit.
Revise the session, persona, directive, kits, and skills for your lane.
Run personakit validate before relying on the root.
Inspect the resolved contract before exporting handoff context.

Inspect The Contract Pieces

Start with the session file; it shows which persona, directive, kits, and skills compose this lane.

Session Sessions/mcp-grounding.session.json Persona Packs/personas/mcp-grounded-agent.persona.json Directive Packs/directives/resolve-before-tools.directive.json Kit Packs/kits/mcp-grounding-guardrails.kit.json Authorized Skill Packs/skills/read-only-mcp-grounding.skill.json Forbidden Skill Packs/skills/mcp-writeback.skill.json

This example shows how PersonaKit fits when an agent can access the PersonaKit MCP server. The agent should ground itself first, then choose tools only if the resolved contract allows them.

The important boundary: MCP is read-only grounding. It does not authorize file mutation, command execution, agent launch, or workflow orchestration.

Unsafe Interpretation This Prevents

An MCP-aware agent might otherwise treat “PersonaKit context is available” as permission to choose tools, mutate files, or launch a worker. This contract says the opposite: resolve context first, then stop unless the needed capability is explicitly authorized.

Key Contract Signal

authorizedSkillIds: read-only-mcp-grounding

Runnable Commands

cd Site/public/examples/mcp-consumer-agent
personakit validate --root personakit-root
personakit contract --root personakit-root --session mcp-grounding

Expected Validation Shape

Validation summary: personas=1 kits=1 directives=1 intents=0 references=0 skills=3 essentials=1 errors=0

What To Inspect

The session resolves to mcp-grounded-agent and resolve-before-tools.
read-only-mcp-grounding is authorized.
mcp-writeback and autonomous-agent-loop are forbidden.
The directive requires grounding before host-local tool selection.
MCP remains a context and provenance surface, not an execution path.

Agent Behavior This Should Produce

An MCP-aware agent should ground itself with PersonaKit before selecting host-local skills, treat MCP resources and tools as read-only context, trace the session when provenance matters, and stop when a needed capability is undeclared or unauthorized.