verdant

oss

Content-addressed cache for AI agent loops. Returns exact bytes from prior executions instead of re-running tools or re-calling LLMs.

status: active
started: 2025
role: creator
stack: Rustblake3MCP

github ↗

Think of it like make: make doesn’t recompile foo.o if foo.c hasn’t changed, because the inputs haven’t changed so the output can’t have changed. Verdant applies the same idea to agent tool calls and LLM completions.

A normal cache keys on the string "cargo test" and guesses with a TTL. Verdant keys on actual content:

key = blake3(
    command  + cwd
  + hash(src/lib.rs) + hash(src/main.rs) + hash(Cargo.lock) + ...
  + env(RUSTFLAGS, CARGO_TARGET_DIR)
)

A hit is a proof (inputs identical), not a gamble (time-based expiry).

Numbers

~23% of tool time saved, measured on one 9-task agent suite (31 cache hits, single run)
LLM replay turned two recorded generations (16.7s and 24.5s) into single-digit-millisecond hits, each a deterministic replay of that one prior generation, not an average across prompts
Zero divergences observed under verify mode (dual-runs every hit, byte-compares)

License: Apache-2.0.