Model Sherpa
nordicnode/model-sherpa
Hermes Agent plugin that repairs bad tool calls, breaks tool loops and redacts secrets from logs
Model Sherpa is a plugin for Hermes Agent that works as middleware between the model and the Hermes tool registry. It repairs malformed tool calls, detects looping behavior and masks secrets in logs.
What Model Sherpa does
Model Sherpa is a Hermes Agent plugin that sits between the model and the Hermes tool registry as middleware. Its Arg Guard validates tool calls against their JSON schema, repairs curly quotes, maps wrong key names such as file_path to path, and has dedicated repair logic for the patch tool. Sequence Loop Detection spots simple repeats, ping-pong loops and longer cycles, then injects a nudge so the model changes approach.
A range-aware Read Damper lets the model page through a file but blocks a read already contained in an earlier range in the same turn. Logs mask any key containing api_key, token, password or secret, and non-UTF8 tool results become a compact placeholder. It is zero-config, with guide-rails switched using commands such as /sherpa doctor and /sherpa feature read_damper off. CI runs a 194-test suite on Ubuntu and Windows across Python 3.9 to 3.13. It hooks the Hermes tool registry and is not framework-agnostic today.
Key features
- Schema validation, smart-quote repair and fuzzy key normalization for tool arguments
- Sequence Loop Detection for repeats, ping-pong loops and longer cycles
- Range-aware Read Damper that allows paging but blocks contained re-reads
- Substring redaction of api_key, token, password and secret values in logs
/sherpa doctordiagnostic and per-feature toggles- Linux, macOS and Windows support
When to use it
- Stop a model from burning tokens by repeating the same failing tool call
- Fix curly-quoted or misnamed tool arguments without a failed run
- Keep API keys out of session logs
Who it is for: Hermes Agent users who see hallucinated tool parameters, looping behavior or secrets in their logs.
How it fits with Hermes Agent
It is built as a Hermes Agent plugin that hooks the Hermes tool registry. The README says it is not framework-agnostic today, and an MCP server port is only on the roadmap.
How to install Model Sherpa
These commands are copied from the project's README. Check the repository for the latest steps before you run them.
hermes plugins install nordicnode/model-sherpaRequirements: Hermes Agent; CI covers Python 3.9 to 3.13 on Ubuntu and Windows
Note: It hooks the Hermes tool registry and is not framework-agnostic; an MCP server port is only on the roadmap.
FAQ
What is Model Sherpa?
Model Sherpa is a Hermes Agent plugin that acts as guide-rails for tool use. It repairs bad tool arguments, breaks repetitive loops and redacts secrets from logs.
How do I install Model Sherpa?
Run hermes plugins install nordicnode/model-sherpa, restart Hermes and run /sherpa doctor to check the tool registry. It is not on PyPI, and a manual install means cloning into ~/.hermes/plugins/model-sherpa.
Is Model Sherpa free and open source?
Yes, Model Sherpa is open source under the MIT license.
Similar security for Hermes Agent
All securityProof-carrying defensive security reviews for AI-assisted code with deterministic policy and portable evidence
dafka007 Hermes Security AuditApproval-gated Hermes plugin that runs Gitleaks, OSV-Scanner and Semgrep CE on a workspace
chchchadzilla GitHub Safe PushHermes skill that keeps secrets out of git and checks a project is properly packaged before shipping
xielevi Hermes Dashboard Auth for FeishuFeishu and Lark OAuth sign-in for the Hermes Agent web dashboard, limited to an allow-list of users
cybertecla Hermes Telemetry DashboardDashboard tab that shows what telemetry Hermes Agent would send to Nous before any sending is enabled
angel12 hermes-ldap-authLDAP and Active Directory password login for the Hermes Agent web dashboard, as a Hermes plugin
Related guides: How to run Hermes Agent securely