sovereign-agent-logbook
Use whenever an agent is about to take an action with side effects (write, send, pay, delete, change configuration) and a Sovereign Agent Logbook MCP server is connected. Teaches when to log, what to put in each field, when to ask a human, and what must never go in the log.
Sovereign Agent Logbook — how to use it well
You have a tamper-evident logbook available as MCP tools: log_event,
verify_logbook, recent_events, export_logbook. This skill tells you
how to use them so that a human can later reconstruct what you did and why.
Setup
The tools come from a local MCP server, installed from PyPI as
sovereign-agent-logbook (Python 3.11+). Connect it to your agent as a
stdio server with the command sovereign-agent-logbook. In OpenClaw:
openclaw mcp add logbook --command sovereign-agent-logbook
openclaw mcp doctor logbook --probe
The server makes no network calls of its own. The log is a file on your
own machine (LOGBOOK_PATH).
The one rule
Log before and after every action that changes something outside this conversation. Writing a file, sending a message, making a payment, changing a record, deleting anything, calling an external system that does something. Reading is optional to log; acting is not.
Never skip a log because the action feels small. The small ones are the ones people ask about later.
Before the action
Call log_event with:
agent_id: your stable name as the operator configured it (e.g.invoice-bot).model: the model you are running on, as precisely as you know it.action: the tool or operation, e.g.send_email,update_crm_record.arguments: the actual arguments. Secrets are masked automatically, but do not put passwords, tokens or personal identifiers anywhere else.reason: one plain sentence a colleague would accept. "Invoice 2026-118 matches purchase order 4471 and is within the auto-approval limit."approval:"human"+approved_byif a person said yes in this conversation."auto"+approval_ruleif a configured rule allows it (name the rule)."none"if neither — and then stop and ask a human unless the operator has explicitly allowed unlogged autonomous actions.
outcome:"pending".run_id: the same id for every event in this task, so they group.
After the action
Call log_event again with the same run_id and action, now with
outcome set to "success", "error" or "escalated", and
result_summary in one line. If it errored, say what the error was.
When to escalate
Set outcome: "escalated" and hand over to a human when:
- an action would move money, delete data, or contact someone outside the organisation, and no human has approved it in this conversation;
- an argument looks wrong (an amount an order of magnitude off, an address you have never seen);
- the previous event for this
run_iderrored and you would be retrying something with side effects.
What must never be in the log
- Passwords, API keys, tokens (masked if under a recognised key; do not rely on it — leave them out).
- Full personal identifiers (BSN, IBAN, passport numbers) in
reasonorresult_summary. Refer to "the customer's IBAN on file" instead. - Free-text copies of documents or emails. Log that you read them and their reference, not their content.
Checking yourself
At the end of a task, call verify_logbook. If it reports ok: false,
tell the human immediately — it means the file was altered outside your
control, which is itself an incident.
Works with any model
This skill contains no model-specific instructions. It has been written to read the same for Mistral, GPT-NL, Llama, Claude, GPT or any other model that can follow instructions and call tools.
- Downloads
- Security audit
- Pass
- Last updated
- 5d ago
- Current version
- v0.1.1
- License
- MIT-0
