sosa-governor — Claude Code plugin: MCP call control
Sorts every MCP tool call by impact level, holds risky actions until you approve, and keeps an audit trail. Useful when an agent sends mail or deletes data.
/plugin marketplace add anthropics/claude-plugins-community
/plugin install sosa-governor@claude-communityWhat sosa-governor does
Three hooks go in and run on their own. SessionStart prints the governance status plus a summary of today's audit, PreToolUse tags each MCP tool call as low, medium or high impact, and PostToolUse records finished actions while updating trust for high-impact tools.
Sending a WhatsApp message, an email, a tweet or deleting data is blocked until you approve it by hand. High-impact tools begin at a trust score of 0. Every approval adds 0.1, and at 10 approvals the tool reaches 1.0 and gets auto-approved. An incident you flag costs 0.5, and the skill can reset a score.
The skill answers phrases like “Run SOSA audit”, “Show trust scores” or “Check compliance”. Config sits in config/: impact-registry.json maps tools to levels, trust-state.json holds current scores, budgets.json defines the token budget hierarchy from monthly down to per-agent. Logs are daily JSONL files under audit/, for instance audit/2026-03-31.jsonl.
What it helps with
- stop an email or WhatsApp message from going out without your sign-off
- pull a report of every tool call made today
- see which tools have earned autonomy
- scan a project for hardcoded secrets
- change the impact level of one tool
Example prompts for the agent
- Run SOSA audit and show me today's tool calls
- Show trust scores for all tools
- Reclassify
mcp__notion__update_pageas medium
How to install sosa-governor
Plugin marketplace
/plugin marketplace add anthropics/claude-plugins-community/plugin install sosa-governor@claude-communityStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about sosa-governor
How do I install sosa-governor in Claude Code?
Add the marketplace first: /plugin marketplace add anthropics/claude-plugins-community, then install the plugin: /plugin install sosa-governor@claude-community.
What does sosa-governor do?
It adds a governance layer to Claude sessions following the SOSA methodology. MCP tool calls get an impact level, risky ones need approval, and everything lands in an audit log.
How does the trust gradient work?
A high-impact tool starts at 0. Each approval you give raises it by 0.1, and at 1.0 the tool is auto-approved. A flagged incident drops the score by 0.5.
Source text
sosa-governor · MIT
# SOSA Governor
**Supervised, Orchestrated, Secured, Agents — Governance for Claude Code**
A plugin that implements the SOSA methodology as a real-time governance layer for Claude Code sessions. Every MCP tool call is classified by impact level, high-impact actions are gated with approval requirements, and all actions are audit-logged.
## About the publisher
[MSApps is Certified in the Claude Partner Network.](https://claude.com/partners)
## How It Works
### Automatic (Hooks)
The plugin installs three hooks that run automatically:
- **SessionStart** — Displays SOSA governance status and today's audit summary
- **PreToolUse** — Classifies every MCP tool call as low/medium/high impact. High-impact actions (sending WhatsApp messages, posting tweets, sending emails, deleting data) are **blocked** until the user explicitly approves
- **PostToolUse** — Logs completed actions and updates the trust gradient for high-impact tools
### On-Demand (Skill)
Say any of these to trigger the governance skill:
- "Run SOSA audit" — Full audit report of today's tool calls
- "Show trust scores" — See which tools have earned autonomy
- "Check compliance" — Evaluate against SOSA Level 1/2/3
- "Security scan" — Check for hardcoded secrets and suspicious patterns
- "Token budget report" — Review budget allocation and usage
- "Reclassify [tool] as [level]" — Change a tool's impact level
## Impact Levels
| Level | Action | Examples |
|-------|--------|----------|
| **Low** | Auto-approved | Read Notion, search Gmail, list WhatsApp chats |
| **Medium** | Logged | Create Monday item, update Notion page, create calendar event |
| **High** | Blocked until approved | Send WhatsApp message, send email, post tweet, delete data |
## Trust Gradient
High-impact tools start with a trust score of 0. Each time a user approves a blocked action, the tool's trust score increases by 0.1. After 10 successful approvals (score reaches 1.0), the tool is auto-approved — it has earned autonomy.
Trust can be reset manually via the skill ("reset trust for [tool]").
An incident (flagged by the user) reduces trust by 0.5.
## Configuration
All config lives in `config/`:
- `impact-registry.json` — Tool-to-impact-level mapping
- `trust-state.json` — Current trust scores per tool
- `budgets.json` — Token budget hierarchy (monthly → weekly → daily → per-agent)
## Audit Logs
Daily JSONL files in `audit/` (e.g., `audit/2026-03-31.jsonl`). Each line:
```json
{
"timestamp": "2026-03-31T14:22:00Z",
"session_id": "abc123",
"tool": "mcp__whatsapp__send_message",
"impact_level": "high",
"reason": "External communication to real people",
"decision": "denied"
}
```
## SOSA Methodology
Full spec: https://github.com/MSApps-Mobile/claude-plugins/blob/main/docs/SOSA.md
## License
MIT — MSApps Research, 2026
© MSApps-Mobile · MIT · MSApps-Mobile/claude-plugins
