Skip to content

Agent skills

riskgate ships two Agent Skills: folders of instructions and scripts that coding and chat agents load when a task matches. Install them, then ask your agent in plain words. For example: "put a gate in front of my refund tool", or "here are my LangSmith traces and chargebacks: how often could this agent run unsupervised?"

Skill What the agent does with it
riskgate Wraps one consequential tool with a gate. It names the worlds, adds read-only lookups and hard rules, calibrates on labelled episodes, replays on held-out ones, and wires the gate into your framework or MCP server. Includes a runnable quickstart (scripts/quickstart_gate.py).
riskgate-from-traces Takes your agent's traces plus an outcome signal. It checks the data (scripts/inspect_traces.py), seals a fit / cal / eval split, learns the worlds on fit only, calibrates once, and writes an honest replay report. See Start from your traces.

Both skills tell the agent the rules that protect the guarantee:

  • never label outcomes with an LLM or with approvals;
  • hold out data, and calibrate once;
  • count groups, not tool calls;
  • set doer_id;
  • never weaken fail-closed behaviour.

They install the library from PyPI (pip install --pre riskgate). The agent needs Python 3.10+ and network access.

Install

/plugin marketplace add reallysaurabh/riskgate
/plugin install riskgate@riskgate

From a shell, use claude plugin marketplace add reallysaurabh/riskgate and then claude plugin install riskgate@riskgate. The skills appear as riskgate:riskgate and riskgate:riskgate-from-traces.

Without the plugin system, copy a skill folder to ~/.claude/skills/ (all projects) or .claude/skills/ (one project).

  1. Download riskgate.zip and/or riskgate-from-traces.zip from the latest release.
  2. Go to Customize → Skills → + → Upload a skill and pick the zip.

Code execution must be enabled. Each zip contains the skill folder at its root.

Messages API: upload the folder with POST /v1/skills. In Python:

import anthropic
from anthropic.lib import files_from_dir

client = anthropic.Anthropic()
skill = client.skills.create(files=files_from_dir("skills/riskgate"))

Agent SDK: copy the skill folders to .claude/skills/ in the agent's working directory.

To enforce the gate in an agent built with the Claude Agent SDK, use the PreToolUse hook as well:

from claude_agent_sdk import ClaudeAgentOptions, HookMatcher
from riskgate.adapters.claude_agent_sdk import pretooluse_hook

options = ClaudeAgentOptions(hooks={"PreToolUse": [
    HookMatcher(matcher="mcp__payments__.*", hooks=[pretooluse_hook(gate, context=snapshot)])]})

gemini extensions install https://github.com/reallysaurabh/riskgate

The repository is a Gemini CLI extension, and its skills/ folder is loaded with it. Without the extension, copy a skill folder to ~/.gemini/skills/ or .gemini/skills/, or run gemini skills install https://github.com/reallysaurabh/riskgate.

Copy the skill folders to .agents/skills/ in your repository, or to ~/.agents/skills/. Codex CLI and other agents that follow the Agent Skills standard read them from there.

Skills guide; the gate enforces

A skill shapes what the agent writes and suggests. It does not stop a tool call at run time. To enforce the verdicts, put the gate in the call path:

  • @rg.guard for any Python tool;
  • the framework adapters: Integrations, including pretooluse_hook for the Claude Agent SDK;
  • riskgate serve-mcp in front of an MCP server, for any MCP client: Claude Desktop, Claude Code, Gemini CLI, IDEs.

Updating

The skills are versioned with the package (metadata.version in each SKILL.md).

  • Claude Code: claude plugin update riskgate@riskgate.
  • Gemini CLI: gemini extensions update riskgate.
  • Copied folders: re-copy them from the new release.