Build & CI
Scan the agent's own code for secrets and personal data, give coding agents a privacy check they can call, and make a boundary crossing fail the merge like any other failing test.
Pre-runtime scanning
Static code scan reads an agent's own source — a local directory, an uploaded zip, or a GitHub repo — for hardcoded secrets, PII in log statements and sensitive values sent to third parties, before a single trace is captured.
Point the scanner at a local directory, a zip file, or a GitHub repo and branch — it reads the same source your agent runs, not a sandboxed copy.
Findings use the identical L1 to L4 severity scale as trace analysis, plus code-specific layers, so a hardcoded API key and a leaked account_id are directly comparable.
POST /api/agent/code is part of the same scoped-key API an agent uses to self-test and improve, so a code scan can run automatically before every deploy.
GitHub-native privacy automation
Run AgentLeak static analysis on pull requests and merge queues before sensitive code lands.
Fail CI automatically when the code privacy score falls below the threshold set for the repository.
Apply stricter secret, PII and third-party-send rules to production and release branches.
Keep findings, reviewers and the current privacy gate attached to the exact commit in GitHub.
Scan a GitHub repo for hardcoded secrets
agentleak scan --repo acme/support-bot --branch main --fail-under 80
# Code privacy score: 74/100 — Conditional pass
# [L3] app/memory_adapter.py:42 hardcoded_api_key (secret, high-entropy)MCP privacy preflight
agentleak mcp exposes four local tools for code scanning, trace checks, redaction and change-aware preflight. Before an agent says the work is done, it can see what is new, what it fixed and what remains open.
$ privacy_preflight workspacehardcoded_secretsrc/client.pyMove the credential to an environment variable before completing this change.
The MCP surface runs the same analyzer as the CLI, directly inside Claude Code, Cursor or any compatible client.
Preflight compares the current scan with the previous check and separates new, fixed and still-open findings instead of returning an isolated score.
History stays in the project as capped, redacted JSONL. Without a project key, no result leaves the machine.
Register the local MCP server
{
"mcpServers": {
"agentleak": {
"command": "agentleak",
"args": ["mcp"]
}
}
}CI policy gate
Set a policy per project and wire AgentLeak into CI. When an agent crosses its boundary, the check fails and the pull request is blocked, with the offending channel and severity attached to the run.
Define what counts as a failure (a channel, a severity level, a score threshold) and the gate enforces it on every run.
The official GitHub Action scans code, traces, scenarios or benchmark packs, annotates the pull request and writes a job summary. The same CLI still works in GitLab CI and every other runner.
A blocked merge comes with the trace, the offending channel and the severity, so the author knows exactly what to fix.
Gate a merge in CI
# .github/workflows/agentleak.yml
- name: AgentLeak privacy gate
uses: yagobski/agentleak@v0.14.1
with:
trace: traces/latest.json
config: agentleak.yaml
fail-under: '80'Built for autonomous agents
llms.txt discovery, one-call onboarding, scoped project keys and machine-readable remediation hints. An agent can find AgentLeak, audit itself and fix its own leaks in a bounded loop, with no browser and no human in the middle.
A machine-readable llms.txt and an A2A agent card let an agent find AgentLeak and learn how to use it without a human.
An agent submits its own trace and gets back a full report: passed, compliant, failed frameworks and the exact remediation hints.
The improve endpoint returns a delta versus the previous run and priority-sorted next steps, so an agent can converge on a clean score safely.
Onboard an agent in one call
curl -sX POST https://www.agentleak.org/api/agent/onboard \
-H 'content-type: application/json' \
-d '{"email":"agent@example.com","agent_name":"SupportBot"}'FAQ