The Caveman MCP server: five compression tools — caveman_compress,
caveman_retrieve, caveman_stats, caveman_toon_encode, and
caveman_toon_decode — over stdio, for any MCP host. It links the
Caveman Engine in-process. Local-only (opens no
network connection) and inferred-only (it never claims verified savings).
The MIT launcher downloads matching BSL-1.1 binary on first run, verifies
key-signed checksum manifest plus artifact SHA-256, and caches it under
~/.caveman/bin. No Go toolchain or global Caveman install is required.
To use an existing reviewed binary instead:
CAVEMAN_MCP_BIN=/path/to/caveman-mcp npx caveman-mcpMIT applies to npm launcher. Downloaded binary follows BSL-1.1 terms named
in BINARY_LICENSE.md.
| Tool | Input | Returns |
|---|---|---|
caveman_compress |
input (string) |
compressed text, inferred ratio, recovery_handle (null on pass-through) |
caveman_retrieve |
recovery_handle (string), optional query |
byte-exact original without a query; ranked complete records with a query; error on unknown handle |
caveman_stats |
— | this MCP process's compression-call totals, including repeated inputs and pass-throughs: tokens before/after, requests, ratio, basis:"inferred", scope:"session" |
caveman_toon_encode |
input (JSON string) |
explicit JSON→TOON result with sizes; pass-through plus note when not encodable |
caveman_toon_decode |
input (TOON string) |
decoded JSON; error on invalid TOON |
Compression is lossy (S4) but reversible: every drop is recoverable via
caveman_retrieve. Incompressible or malformed input passes through unchanged
with ratio:0 — never an error.
Repeated recovery calls remain available after host compaction. An empty query returns the exact stored original on every call; a nonempty query keeps the same record-narrowing behavior regardless of how many earlier calls were made.