PROJECT · TOK0 / FIELD COMPRESSION UNIT
SERIAL NO. 0.1.1 · BUILT IN RUST · MIT
OPEN SOURCE — SHELL OUTPUT COMPRESSION PROXY TOK0 · DOCS >>> PROJECT

FAQ

Quick answers to the questions that come up most often. If yours isn't here, the issue tracker is the next stop.

Does tok0 work with my AI tool?

Probably yes. tok0 ships native integrations for Claude Code, Cursor, Gemini CLI, Windsurf, Cline, Amp, OpenCode, Codex, and Kimi Code. See the coverage matrix for the full list and integration type.

For tools without a native integration, you can still use tok0 manually by invoking commands as tok0 <cmd> — the compression works the same, you just lose the auto-rewrite from the agent’s hook surface.

Will it slow down my commands?

Pipeline overhead is < 0.5 ms per command, in benchmarks. The wrapped command runs as it normally would; tok0 just observes its stdout/stderr. You will not feel it.

Does compression affect what my agent can do?

No. Every compressor’s CI tests assert that the signal is preserved — the agent still sees error messages, exit codes, file paths, the trailing summary. Only redundant or formatting-only output is collapsed. If you find a case where the model loses information it needed, that’s a bug — file an issue with the captured fixture.

What about non-deterministic output?

The pipeline itself is deterministic (no clock reads, no randomness). If a wrapped command emits non-deterministic output (timestamps, hashes, port numbers), the compression of that output is also non-deterministic — but each invocation produces a stable, valid result.

If you need to normalize timestamps or hashes for downstream comparison, add a pattern-replace stage in your rule:

[[replacements]]
pattern = "[0-9a-f]{40}"     # 40-char hashes
replacement = "<hash>"

Does it cost anything?

The CLI is MIT-licensed, free, and open-source. The optional cloud dashboard (team-level analytics) is paid; the binary you run on your machine never is.

Can I use it offline?

Yes. With [telemetry].enabled = false (the default) and [updater].auto_check = false, tok0 makes zero network calls. Verify with:

TOK0_LOG=trace tok0 git diff 2>&1 | grep -i 'http\|request'
# (no output means no network calls)

Does it work in CI?

Yes. Two patterns:

  • Bypass mode: set TOK0_DISABLE=1 in CI to short-circuit tok0 and pass commands through unchanged. Useful when you only want compression locally.
  • Active mode: install tok0 as part of your CI image, run tok0 init --tools none (skip AI hooks), and invoke commands as tok0 <cmd>. Output is compressed and counted in stats.

How big is the binary?

8 MB stripped on x86_64 Linux. Single static, no runtime dependencies.

Can I use it with my own internal CLI?

Yes — write a TOML rule at ~/.config/tok0/filters/my-internal-tool.toml, drop it in place, and tok0 picks it up. No rebuild required.

For team-wide rules, package them as an extension and have your team tok0 extensions install <git-url>.

Does it support Windows?

Yes, native binaries for x86_64 and arm64 ship on every release. Most pipeline behavior is identical; the bridge surfaces are Windows-specific where appropriate (%APPDATA% instead of ~/.config).

PowerShell prompts are detected and stripped like any other shell prompt.

Does it support remote shells / SSH / containers?

Three flavors:

  • Wrapping ssh itself: yes, the system/ssh compressor reduces verbose connection diagnostics. The actual remote shell session passes through unchanged.
  • Inside a remote shell: install tok0 on the remote host. Bridge configuration is local-host-only.
  • Inside a container: the Docker image at ghcr.io/prxm-labs/tok0:latest is meant exactly for this.

How does it compare to piping through grep or head?

grep/head are line-level filters; tok0’s pipeline is content-aware. A grep -v will strip a banner line but keep the ==> it’s part of. tok0’s TOML rules understand command-shaped output: where the verdict lives, where the noise lives, what to do when the output is a JSON tree, what exit code semantics imply.

The bigger difference is that tok0 ships with 250+ rules already tuned and tested. You don’t need to figure out the right awk incantation per tool.

Where do compression numbers in the marketing copy come from?

The ~80% figure on the homepage is the median compression ratio across the top-N most-frequently-run commands in the deployed fleet, weighted by frequency. Each individual rule has a CI-asserted minimum floor — see Compressors for per-command numbers.

Why is it called tok0?

tok for token, 0 for “shrunk to zero” / “field unit zero”. Pronounced “tock-zero”. The brand mark is t0.

Tip

Have a question that should live here? Open an issue tagged question on GitHub. We periodically promote the most-asked ones into this page.

BUILT IN RUST · SINGLE STATIC BINARY · 8 MB v0.1.1 / MIT GITHUB.COM/PRXM-LABS/TOK0