pureyq

yq, as a pure Python library. Run jq programs over YAML, TOML, XML, CSV and JSON — no yq binary, no jq binary, no C extension required: if Python runs, pureyq runs. Pyodide/WASM, sandboxes, Lambda, anywhere pip install is all you get.

pip install pureyq

pureyq -i '.spec.replicas = 3' deploy.yaml          # edit YAML in place
pureyq -o json '.services | keys' compose.yaml      # YAML in, JSON out
pureyq '.dependencies' pyproject.toml               # TOML works the same way

import pureyq

pureyq.apply(".spec.replicas = 3", manifest_text)   # text -> text, one call
data = pureyq.load(manifest_text)                   # YAML 1.2 -> Python
pureyq.first(".spec.template.spec.containers[].image", data)

The expression language is jq — the real one, not a dialect: the engine is purejq, which passes 96.2% of jq's own test suite. Everything you know from jq works on your YAML: select, map, group_by, paths, assignment operators, reduce, try/catch, string interpolation, regexes.

Why pureyq

• No binaries, anywhere. kislyuk/yq needs a jq binary on the system at runtime; mikefarah/yq is a Go binary. In sandboxed or pip-only environments (Pyodide, Lambda layers, locked-down CI images, agent sandboxes) neither is an option. pureyq is plain Python wheels all the way down.

• Embedding in Python. Transforming a manifest in-process with pureyq.apply() takes a fraction of a millisecond; spawning a yq binary per call costs milliseconds. For agent/automation loops that edit many small configs, in-process wins by an order of magnitude.

• YAML 1.2 correctness by default. PyYAML-based tools (including kislyuk/yq) speak YAML 1.1, with famous consequences:

  input	YAML 1.1 loaders read	pureyq (1.2 Core Schema)
  country: NO	false (!)	"NO"
  version: 010	8 (octal)	10
  time: 1:30	90 (sexagesimal)	"1:30"
  date: 2026-06-11	a datetime object	"2026-06-11"

  On output, strings that either YAML generation would misread are quoted automatically, so emitted files are safe for downstream 1.1 parsers too. Merge keys (<<:) still work — real-world configs depend on them.

Formats

Input format is detected from the file extension (-p to force); output defaults to the input format (-o to convert). pureyq -o json . config.toml and pureyq -o yaml . data.json are complete format converters.

CLI

pureyq [options] '<jq filter>' [files...]

-p FMT   input format: auto|yaml|json|toml|xml|csv|tsv (default: by extension)
-o FMT   output format (default: same as input)
-i       edit files in place (atomic; preserves permissions)
-n -r -j -c -s -e -f --arg --argjson    the flags you know from jq
--indent N    output indentation

Multi-document YAML streams behave like jq input streams: each document is one program run, --slurp collects them into an array, and input/inputs consume the rest.

Benchmarks

Measured with tools/bench.py: M-series MacBook, CPython 3.12, mikefarah yq v4.53.3 (native arm64 binary), kislyuk/yq 3.4.3 over jq 1.8 — three independent rounds, each the median of 7 runs, and every workload's outputs verified equal across all three tools before timing. The numbers below are cross-round medians. Reproduce: python tools/bench.py --verify.

Embedded in Python — editing a k8s manifest, per call:

In-process beats shelling out ~36x. For agent/automation loops that touch many configs, this is the number that matters.

Command line, small file — 40-line k8s manifest, startup included:

Command line, big file — 15 MB YAML, 100k objects, end to end:

On big files pureyq is 15–25% faster than the jq-wrapper approach, while also removing its jq-binary requirement. Where the Go binary wins: big-file throughput, by 2–5x; if you can install binaries and that is your workload, use mikefarah/yq. (One caveat on the Go side: its compact-JSON mode -I0 is quadratic on large arrays — converting a 20k-row file takes it 104 s vs pureyq's ~1 s — so agents asking for compact JSON from big YAML hit a wall pureyq doesn't have.)

Correctness, measured

• TOML: the official toml-test suite is vendored in this repo (tests/conformance/toml-test) and runs in CI on every commit: 704/704 of the TOML 1.0 cases pass (209 valid documents match the typed expectations, 495 invalid documents are rejected).
• jq semantics: inherited from purejq, which vendors jq's own test suite (751/781 passing, every difference documented).
• YAML 1.2 schema: a directed test set covers the 1.1/1.2 divergences (booleans, octals, sexagesimals, timestamps, .inf/.nan, quoting on output), and the libyaml fast path is asserted to agree with the pure Python fallback on every case.

Limitations (honest ones)

• Comments and exact formatting are not preserved through an edit, same as kislyuk/yq. (mikefarah/yq can preserve them because its engine operates on the YAML node tree; a jq engine works on values.) Anchors/aliases are resolved on load and not re-emitted.
• TOML output requires a single object result (that's what a TOML document is); CSV output requires flat rows.
• When PyYAML carries its libyaml C extension (standard wheels do), pureyq uses it for parsing speed — with a pure Python fallback that behaves identically, asserted by tests. "Pure" means required by, not faster with.

License

MIT