Works with the AI Agents you already use
Judikt MCP Security
Your AI agent can call every tool an MCP server exposes, including the ones that delete, overwrite or deploy.
Free
Judikt MCP Security
Example session with this skill installed
Put Judikt in front of my filesystem MCP server for ~/project so Claude Code can only read files there, then prove it works.
- Read your context and instructions
- Compiled the judikt mcp security
Judikt is now guarding your filesystem server. Claude Code can read files under ~/project, every write tool is hidden and blocked, and the audit trail verifies.
1. Environment check (read-only)
python3 scripts/diagnose.py found Python 3.13 and
judikt 0.3.0. It reported "network_calls": false and "config_contents_read": false, so nothing was changed.
2. Starter policy, generated from the server's real tools
With your approval I ran
judikt init --name filesystem --output ~/project/judikt-policy.json -- \
npx -y @modelcontextprotocol/server-filesystem ~/project
Judikt starter policy (read-only): read=10 write=1 destructive=3
Excluded tools: write_file, edit_file, create_directory, move_file
Approval-gated tools: none
Annotation/name conflicts: none
3. Client configuration (the change you approved in .mcp.json)
{
"mcpServers": {
"filesystem": {
"command": "judikt",
"args": ["wrap", "--policy", "/Users/you/project/judikt-policy.json", "--name", "filesystem", "--",
"npx", "-y", "@modelcontextprotocol/server-filesystem", "/Users/you/project"]
}
}
}
4. Smoke test
| Check | Result |
|---|---|
tools/list | 10 tools, all read-only (read_text_file, list_directory, search_files, ...) |
read_text_file docs/runbook.md | ALLOW, returned the file's contents |
write_file docs/runbook.md (the denial check you authorized) | DENY before reaching the server: tool 'write_file' is not allowlisted for 'filesystem' (rule: allowlist). The file is unchanged. |
judikt audit verify | "status": "valid", "signed": true, 2 events checked |
5. Assessment
| Control | Status | Evidence |
|---|---|---|
| Least-privilege tool exposure | verified | 10 of 14 tools exposed; all 4 write tools absent from tools/list |
| Pre-execution denial | verified | write_file refused with rule allowlist; file unchanged |
| Tool-definition pinning | verified | Pin store active; the 4 excluded tools await review with judikt pins accept |
| Signed, hash-chained audit | verified | judikt audit verify: valid, signed |
| Response inspection | configured-not-verified | On by default; no malicious content was tested |
| Approval tokens | not-applicable | Read-only profile, so no approval-gated tools |
| Remote authentication | not-applicable | Local stdio only |
| Tamper evidence against a local attacker | not-configured | No external checkpoint destination such as S3 Object Lock or a SIEM |
Next steps
- Restart Claude Code and confirm the
filesystemserver shows 10 tools. - If the agent should edit files, rerun
judikt initwith--profile approve-writesso every write needs a signed, single-use approval. - For audit evidence that survives a compromised machine, export checkpoints with
judikt audit checkpoint --output s3://....
Connects securely to your tools. The creator never sees your data.
About this skill
Your AI agent can call every tool an MCP server exposes, including the ones that delete, overwrite or deploy. This skill puts Judikt, an open-source security gateway, between your agent and that server, and walks your agent through setting it up safely, asking for your consent at every step that changes anything.
What it does
- Generates a least-privilege policy from the server's real tool list. Read-only by default: tools that write, edit, move or delete are hidden from the agent. The
approve-writesprofile exposes them behind human-signed, single-use approvals. - Connects Claude Code or Cursor by proposing a minimal change to
.mcp.jsonor.cursor/mcp.jsonand leaving your other servers untouched. - Smoke-tests the gateway with
tools/listand one read-only call you authorize. It never uses a mutating tool just to prove connectivity. - Produces an evidence-based assessment that marks each control as verified, configured-not-verified, not-configured or not-applicable, so configuration is never mistaken for proof.
- Troubleshoots policy, identity, approval, pinning and audit failures without weakening security to make a test pass. When Judikt holds a call, the skill tells you the
judikt approve <id>command to run yourself; it never approves on your behalf.
What Judikt enforces once it's in place
- Allowlists and argument checks before any tool runs
- New in 0.4.0: a provenance gate that holds a write reusing an address, account, link or ID that an instruction in tool output asked for, until you approve it with one command (
judikt approve <id>) - New in 0.4.0: task scopes (
judikt wrap --task-writes send_email) that hold any write the task didn't declare - Tool-definition pinning: a server that silently changes a tool is refused at startup
- Kill switches, rate limits and circuit breakers
- Redaction of secrets in results and audit logs
- A hash-chained, HMAC-signed audit trail you can check with
judikt audit verify
Honest limits
All figures below come from replaying each benchmark's own correct tool calls, assuming an agent that obeys every injection. No language model ran, so they are not end-to-end agent results.
- In an InjecAgent replay, 0 of 1,054 harmful actions executed under the read-only or approve-writes profiles.
- On AgentDyn, a benchmark Judikt was never developed against, the provenance gate held 67.8% of attacks and provenance plus a task scope held 94.8%. 27 of 60 legitimate tasks would stop for one approval, mostly tasks that act on instructions in emails or web pages.
- Prompt-injection detection is a tripwire, not the main defence: it caught 54.8% of BIPIA and 7.6% of InjecAgent attacks with no false positives.
The skill does not bundle or silently install Judikt, edit client configuration without your consent, or invent approval tokens.
Requirements
Python 3.11+ and the Judikt runtime, installed separately with pip install "judikt==0.4.0". Upgrading from 0.3.0? The provenance gate is on by default; read the upgrade notes: https://github.com/gaurav-gs7/Judikt/blob/v0.4.0/CHANGELOG.md. Open source under Apache-2.0: https://github.com/gaurav-gs7/Judikt
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
Requires Python 3.11+ and a separately installed Judikt runtime (pip install judikt). Never installs software, edits client configuration, or handles credentials without your consent. Smoke tests use read-only tools only. Only Claude Code and Cursor are qualified clients. Judikt's prompt-injection scanning is a heuristic tripwire: on independent benchmarks it caught 0.8% (BIPIA) and 0% (InjecAgent) of attacks, so the real protection comes from policy, approvals and pinning.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 19 days ago
- Free with an Agensi account
- Software
Python 3.11+judikt 0.4.0MCP server to protect
Install with pip install judikt==0.4.0. The server can be stdio or a remote Streamable HTTP endpoint.
Permissions
File Scopes
Terminal runs the read-only diagnostic and, only with your consent, judikt init and judikt wrap. Read Files inspects candidate client-configuration paths and the generated policy. Write Files creates the policy file and, after you approve a shown diff, edits .mcp.json or .cursor/mcp.json. Network access is used only if you agree to install Judikt with pip or to test a remote Judikt endpoint. The bundled diagnostic makes no network calls and reads no environment values.