ToolRush

What ToolRush does

The README argues that tool calls, not token generation, have become the bottleneck, because every read_file, search_files and terminal call pays for process spawns, shell round-trips and wrapper layers. ToolRush adds native file reads, a persistent warm shell, direct ripgrep execution for search, and a parallel helper (from hermes_tools import parallel) that runs a batch of 1 to 16 read operations on up to four workers through one RPC. A fail-closed classifier sends anything that looks like a hidden write back to the normal sequential path.

The README reports measured gains on its own harness: native reads from 255.23 ms to 4.44 ms, a warm terminal from 285 ms to 12.1 ms, and search transport from 183 to 455 ms down to 27 to 97 ms. Compatibility patches live outside the upstream checkout and are restored in memory after a Hermes update. A doctor.py --smoke command and per-lane TOOLRUSH_*=0 kill-switches support diagnosis and rollback.

Key features

• Native file reads that skip process spawns
• Persistent warm shell with preserved exit status, cwd and exports
• Direct ripgrep execution for search
• Parallel batches of 1 to 16 read operations on up to four workers
• Fail-closed scheduler admission for hidden writes
• Per-lane kill-switches and a doctor.py --smoke check

When to use it

• Speed up tool-heavy Hermes turns such as repository search and file reading
• Run several independent read operations in one call
• Keep performance patches working across Hermes harness updates

Who it is for: Hermes Agent power users whose sessions spend most of their time waiting on tool calls.

How it fits with Hermes Agent

Built for Hermes Agent: it is described as an execution layer for Hermes and patches its tool transport, restoring those patches after harness updates.

Note: The README states the speedups are tool-operation wall times, not full-turn speed, and discloses one regression: trivial native reads do not benefit from threading.

FAQ