ToolRecall Reference

$ toolrecall setup
  Creates config, starts daemon, begins caching

$ toolrecall status
  Shows cache stats: hit rate, unique files, context savings

$ toolrecall shim --install
  Installs .pth shim for transparent open()/subprocess interception

$ toolrecall shim --remove
  Removes the .pth shim

$ toolrecall reset
  Clears all caches and resets daemon state

$ toolrecall config set <key> <value>
  Set config (path allowlist, terminal TTLs)

$ toolrecall config show
  Print current config

Cache Layers

File cache: mtime-based invalidation. Read once from disk → served from memory forever. 99.3% hit rate.

Terminal cache: TTL-based, configurable allowlist. ~8 static commands cached.

MCP cache: Result caching for slow/idempotent external calls (search, fetch, docs).

Forward proxy: Request-body hash. Byte-identical requests skip the API entirely.

Context Tracker

In-memory dirty/clean tracking. Files read via cached_read() tracked as clean. Files written via cached_write()/cached_patch() marked dirty. Checkpoint-based — resets between sessions.

Does ToolRecall replace provider prefix caching?

No — complementary. ToolRecall reduces what you send. Provider caching discounts what you're billed. Together: lower latency + lower cost.

Does it work offline?

Yes. File cache, terminal cache, MCP bridge — all offline. Only the forward proxy needs an API endpoint.

Does it slow down first reads?

First read is ~2× slower than bare subprocess (IPC + SQLite write). Every repeat read is ~0.6ms — ~1000× faster than a subprocess fork.

How do I share one daemon across agents?

Run toolrecall setup once. All agents on the same machine connect to the same UDS. MCP multiplexer pools servers across sessions.

Is there a warm-up command?

No. First read after daemon start is always a cache miss. Benefit materializes on repeat reads within the session.