Skills & Tools

Tools are the actions an agent can take. Moneypenny supports multiple tool sources and makes them all available through a single registry with unified policy, audit, and discovery.

Because Moneypenny exposes its full API through MCP, you can manage tools and skills conversationally from any MCP-compatible client (Claude Desktop, Cursor, etc.) or through the CLI.

Tool Sources

Source	Description	Examples
Built-in	Compiled into the binary	`file_read`, `file_write`, `shell_exec`
Runtime	Operate on the agent’s own database	`memory_search`, `fact_add`, `job_create`
MCP	Discovered from external MCP servers	Any MCP-registered tool
JS	User-defined JavaScript in QuickJS sandbox	Custom tools created at runtime

Tool Execution Lifecycle

Every tool call follows the same pipeline:

Tool call requested
  → Policy check (allow / deny / audit)
  → Pre-hooks (argument transformation or abort)
  → Dispatch (runtime → MCP → JS → builtin)
  → Post-hooks (output transformation)
  → Secret redaction
  → Audit log write
  → Result returned to agent

Runtime Tools

These tools let the agent operate on its own state:

Tool	Purpose
`memory_search`	Search across all memory stores
`fact_add`	Store a new long-term fact
`fact_update`	Update an existing fact
`fact_list`	List active facts
`scratch_set` / `scratch_get`	Session-scoped working memory
`knowledge_ingest`	Ingest a document
`knowledge_list`	List ingested documents
`job_create`	Create a scheduled job
`job_list` / `job_pause` / `job_resume`	Manage jobs
`policy_list`	List active policies
`audit_query`	Search the audit trail
`web_search`	Search the public web

MCP Tools

If MCP servers are configured, tools are discovered automatically at startup:

[[agents.mcp_servers]]
name = "github"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]

Discovered tools appear in the agent’s tool list alongside built-in tools.

JS Tools (Agent-Created)

Agents can create their own tools at runtime using JavaScript:

CLI
Natural Language

echo '{"op":"js.tool.add","args":{
  "name":"celsius_to_fahrenheit",
  "description":"Convert Celsius to Fahrenheit",
  "source":"function run(args) { return String(args.celsius * 9/5 + 32) + \" °F\"; }",
  "parameters_schema":"{\"celsius\":\"number\"}"
}}' | mp sidecar

JS tools:

Execute in a QuickJS sandbox (not Node.js)
Are policy-gated and audited like any other tool
Persist in the agent’s database
Survive restarts
Sync across agents via CRDT

echo '{"op":"js.tool.list","args":{}}' | mp sidecar
echo '{"op":"js.tool.delete","args":{"name":"celsius_to_fahrenheit"}}' | mp sidecar

Skills

Skills are reusable procedures stored in the agent’s database. They surface in retrieval when the agent encounters a relevant query:

CLI
Natural Language

echo '{"op":"skill.add","args":{
  "name":"incident-triage",
  "description":"Triage production incidents",
  "content":"1) Check dashboards. 2) Escalate: on-call → lead → VP. ..."
}}' | mp sidecar

CLI
Natural Language

mp skill list

Skills track usage and success. High-performing skills surface more prominently in context assembly.

Tool Governance

Every tool call is subject to policy. Common patterns:

CLI
Natural Language

# Block shell access entirely
mp policy add --name "no-shell" --effect deny --action "call" --resource "shell_exec"

# Audit all web searches
mp policy add --name "audit-web" --effect audit --action "call" --resource "web_search"

# Allow only read-only tools for low-trust agents
mp policy add --name "read-only" --effect deny --actor "intern-bot" --action "call" --resource "file_write"