Token Terminator

What Token Terminator does

Token Terminator works at the tool-result and provider-request boundaries of an agent runtime and removes token bloat while keeping the evidence recoverable. Its middleware combines several mechanisms: optional transparent terminal-command rewriting through RTK, temporal delta compression for repeated terminal observations after a command has run, and deterministic compression of large tool results. The project describes exact recovery and fail-open behavior, and documentation sits in the GitHub Wiki.

For Hermes Agent, it can be chosen as the ContextEngine through explicit engine selection, replacing the selected upstream engine with full-history JEV attention that has exact backing and rediscovery. A learned omission-risk policy is off by default, with shadow and active modes as opt-in. An Observatory shows per-bot input and output accounting at localhost:7474, and a Desktop bottom-bar counter is available. The explicit install-context-engine step may apply a backed-up, narrowly scoped change to a reviewed Hermes loader, and it refuses unknown host implementations rather than overwriting them.

Key features

• Selectable Hermes ContextEngine through explicit engine selection
• Temporal delta compression for repeated terminal observations
• Deterministic compression of large tool results
• Optional transparent terminal-command rewriting through RTK
• Observatory with per-bot input and output accounting at localhost:7474
• Learned omission-risk policy, off by default with shadow and active modes

When to use it

• Cutting token use in long Hermes Agent sessions with heavy terminal output
• Keeping removed tool output recoverable for later inspection
• Tracking per-bot input and output token accounting across several agents

Who it is for: Hermes Agent operators who run long, tool-heavy sessions and want context reduction that keeps the original evidence recoverable.

How it fits with Hermes Agent

Token Terminator is agent-agnostic, but it includes a first-party Hermes Agent adapter that is selected as the Hermes ContextEngine and tracks Hermes' live-compression behavior.

FAQ