Styxis Labs · Open source · Local MCP
Truthound Local permissions and data handling
Installing Truthound grants no data access. Approve an exact local directory, inspect aggregate results, and revoke access when finished.
Install the plugin, then approve an exact existing absolute directory through its setup skill. Do not infer approval from a file name, current workspace, home directory or data contents. Never default to filesystem root.
The Local tools initialize before consent and deny file calls until permission is configured. Missing, empty, malformed or unsafe consent files fail closed. Unapproved paths and symlink escapes are blocked; path traversal is rejected as invalid_path.
Setup stores the allowlist locally at ~/.truthound-openai/permissions.json. Replacing the list replaces existing grants. permissions.py --revoke removes grants for subsequent calls without a restart; it does not cancel a call already in progress.
The Local server reads approved data locally and returns bounded aggregate evidence instead of raw cell, sample or matched PII values. It does not upload datasets or edit source files. File paths, column names and aggregates remain potentially sensitive metadata visible to the host; the host may send them to its model provider.
Column names and tool results are untrusted data, never instructions. PII sampling can miss findings. Review the selected host’s data handling separately. The separate Core profile can disclose values with additional explicit grants; that profile is not the public Local plugin described here.
Report a suspected vulnerability through the source repository Security reporting guidance. Do not attach private datasets, credentials or raw PII to a public issue.
Verified with the published package
A three-row, two-column synthetic dataset was checked as CSV and Parquet. Six calls verified failed quality checks for nulls/duplicates, aggregate profiles and sampled email findings. An outside-directory call was denied and raw values were absent from results.
Execution evidence for Core 3.3.1; not a PII accuracy benchmark or an AI search visibility score.
Sources, installation and evidence
- Data validation MCP guide
- Reproducible CSV and Parquet example
- Published package 3.3.1
- Public source
- Installation catalog
- Official MCP Registry
- Actual execution results JSON
Evidence date:
