@tokz/compress

Compressed context your agent can still edit against

A local MCP server that hands your agent the exact part of a file it asked for — checked byte-for-byte against disk before it's returned.

MIT, and local by default — no account, no API key, no network.

Why verbatim is the whole point here

An agent editing against paraphrased code produces an old_string that does not match disk, so the edit fails. Verbatim is not a nicety here — it is the difference between an edit landing and an agent looping.

// what a lossy context tool handed the agent
function handleRateLimit(retry) {
  // wait before retrying the call
}

// the exact bytes tokz-compress returned
function handleRateLimit(retryAfterMs) {
  await sleep(retryAfterMs);
}

Four strategies, picked by what the file is

Code

Tree-sitter parse, BM25 ranking over the symbols that match the query.

Logs

Collapses identical-shaped lines, keeps the ones that differ.

JSON

Extracts the schema, keeps representative samples.

Prose

Block selection, offline-first.

86%

recall

37 of 43 required spans found verbatim across the eval set.

~1 ms

retrieval

tokz_retrieve(hash, lines) returns any elided range byte-for-byte from a local cache.

12 / 29 ms

median / p95

Compression latency, local, no network.

0

verbatim failures

Across the eval set. Every emitted span matched the file on disk.