Agents

An agent in Moneypenny is a self-contained unit of intelligence backed by a single SQLite file. That file contains the agent’s memory, policies, skills, knowledge, audit trail, sessions, and tool definitions.

Because Moneypenny exposes its entire surface area via MCP (Model Context Protocol), every operation that works from the CLI also works as a natural language request through any MCP-compatible client — Claude Desktop, Cursor, or anything else that speaks the protocol.

Agent Anatomy

Every agent has:

Component	Storage	Purpose
Facts	facts table	Long-term structured memory
Messages	messages table	Conversation history
Knowledge	documents + chunks tables	Ingested documents
Skills	skills table	Reusable procedures and tool definitions
Policies	policies table	Governance rules
Jobs	jobs + job_runs tables	Scheduled tasks
Audit	policy_audit table	Decision log
Scratch	scratch table	Session-scoped working memory

Creating Agents

Initialization creates a default agent named main:

CLI

mp init

Natural Language

Ask your MCP-connected agent:

Initialize a new Moneypenny project.

Create additional agents:

CLI

mp agent create research
mp agent create ops-bot

Natural Language

Ask your MCP-connected agent:

Create a new agent called research.

Agent creation goes through the canonical operation pipeline — policy-checked, hooked, and audited like any other mutation.

Agent Configuration

Each agent is configured in moneypenny.toml:

[[agents]]
name = "main"
trust_level = "standard"
persona = "You are a helpful engineering assistant."

[agents.llm]
provider = "anthropic"       # or "local" for GGUF, "http" for other APIs
# api_base = "https://api.openai.com"
# api_key = "sk-..."
# model = "gpt-4o"

[agents.embedding]
provider = "local"
model = "nomic-embed-text-v1.5"
dimensions = 768

Update config at runtime:

CLI

mp agent config main persona "You are a senior SRE."

Natural Language

Ask your MCP-connected agent:

Update the main agent’s persona to “You are a senior SRE.”

Trust Levels

Agents have a trust level that policies can reference:

• standard — default, subject to all policy rules
• elevated — can be granted access to restricted resources
• admin — full access (still audited)

Trust levels are labels — their meaning is defined by the policies you write.

Listing and Inspecting

CLI

mp agent list
mp agent status
mp agent status research

Natural Language

Ask your MCP-connected agent:

List all agents.

Show me agent status.

Show me the status of the research agent.

Status shows memory stats: fact count, session count, document count, and skill count.

Deleting Agents

CLI

mp agent delete research --confirm

Natural Language

Ask your MCP-connected agent:

Delete the research agent.

This removes the agent from the registry and deletes its database file.

Multi-Agent Architecture

When running mp start, each agent runs as a separate worker process. The gateway routes messages to the correct worker based on agent name. Agents can delegate to each other with depth-limited recursion.

Gateway
  ├── Worker: main      (main.db)
  ├── Worker: research  (research.db)
  └── Worker: ops-bot   (ops-bot.db)

Facts, skills, policies, and jobs can sync between agents via CRDT. Conversations and scratch stay local to each agent.

Portability

An agent is its database file. To move an agent:

cp mp-data/main.db /backup/main.db        # backup
scp mp-data/main.db server:/agents/main.db # transfer

Open with any SQLite client to inspect:

sqlite3 mp-data/main.db "SELECT pointer, confidence FROM facts WHERE status='active';"