heso 0.0.10

The agent-native web engine. No Chromium. No Node. One Rust binary.

heso — Playwright for agents, built around reproducibility. One Rust binary.

Site: heso.ca · Docs: heso.ca/docs · npm · PyPI · Releases

A headless web engine for agents: fetches a URL, runs the JavaScript, clicks, fills forms, scrapes — and returns JSON. The load-bearing part: every run can be stamped into a signed plat with an embedded network cassette, then byte-identically re-executed off-network with heso run. Anyone, any machine, same output. No Chromium. No Node.

Optimized for structured outputs and reproducibility, not for pixel-perfect rendering or anti-bot warfare. If your task lives in pixels or animations, use Chromium. If it lives in JSON, you're in the right place.

binary       9.6 MB
cold start   ~77 ms   (open https://example.com, network included)
engine only  ~28 ms   (no network)
batch        ~1.1 s   for 8 URLs in parallel

That's a real recording — Claude Code (claude -p from the repo root, with the heso skill loaded) discovering the verbs, navigating the page tree, and pulling the live top story off Hacker News. No edits, no replays.

Install

# Python (uv, pipx, or pip — any of them)
uv tool install heso          # or: pipx install heso  /  pip install heso

# Node
npm install -g @ixla/heso     # or one-shot: npx @ixla/heso open https://example.com

# Direct binary
# Windows:
powershell -c "irm https://github.com/blank3rs/heso/releases/latest/download/heso.zip -OutFile heso.zip; Expand-Archive heso.zip -DestinationPath ."

Shipping v0.0.10 for Windows-x64, Linux x64 + arm64, macOS x64 + arm64. cargo-dist builds every target on tag; npm/PyPI publish through the same workflow.

After install, heso is on $PATH:

heso open https://example.com
# → { url, title, description, tree, actions, plat_hash, ... }

You get JSON: title, description, a heading tree, and a list of clickable elements numbered @e0, @e1, and so on.

A note before you read further

Most of this codebase was written with help from Claude under one person's direction. The co-author tag is on basically every commit. It moved fast, which means the feature surface ran ahead of real usage. Treat this as working code that needs more eyes on real workloads, not a finished product.

What it can do

Find and read things.

• heso search "<query>" — searches the web (DuckDuckGo + Wikipedia, optional SearXNG). No API key.
• heso open <url> — fetches and returns a page summary: title, headings, actionable elements.
• heso read <url> — fetches, runs JS, returns the full picture: title, visible text, actions, forms, cookies, console output, framework detection. One call.
• heso read <url> --complete — same, but heso loops "fire pending observers + click load-more + wait for DOM to settle" until the page stops changing. For lazy-loaded sites.
• heso batch [open|read] <urls...> — runs many URLs in parallel. Shared cookie jar, JSON-Lines out.
• heso wait <url> --selector-exists ".foo" (also --text-contains, --url-matches, --network-idle, --time) — blocks until a condition is true. No polling loop.

Interact with sites.

• heso click <url> @e7 — click by element ref.
• heso click <url> --text "Sign in" — or by visible text, CSS selector, or aria-label.
• heso fill <url> @e3 "hello" — type into an input.
• heso submit <url> @e9 — submit a form.
• heso navigate — change URL within a session.
• heso eval-dom <url> "<js>" — fetch, run scripts, then run your JS against the resulting DOM.

Bundle, edit, replay, and re-execute action sequences.

A plan is a JSON array of canonical actions (open, click, fill, submit). A plat is an observation, plus an embedded network cassette — every (method, URL, request-body) → (status, headers, response-body) tuple the engine touched during the run. Four verbs close the loop:

• heso stamp <plan.json> — executes the plan against the live web and mints a fresh plat that embeds the plan, the recorded cassette, and a per-step log. Accepts a bare Action[] array, a plat with a "plan" field, or a TraceFingerprint. Exit 0 on a clean run; 1 if any step failed (still prints the partial plat with error + steps).
• heso run <plat.json> — re-executes the plan against the embedded cassette. No network. For an unchanged cassette the output plat_hash equals the input's — byte-identical replay (ADR 0008). If the cassette has drifted (page changed since stamping), the failing step carries a structured cassette miss: METHOD URL not recorded error and run exits 1 — graceful, never silent.
• heso replay <plat.json> — pure observation. Reads the recorded step log from the plat and prints it. No engine, no JS, no cassette lookup, no network. Use run if you want to re-execute.
• heso unpack <plat.json> — extracts just the plan field. Edit it standalone and pipe back into stamp to re-mint a fresh plat (with a fresh cassette since the requests changed).

cat > plan.json <<EOF
[
  {"verb": "open",   "url": "https://news.ycombinator.com/"},
  {"verb": "click",  "ref": "@e3"},
  {"verb": "fill",   "ref": "@e7", "value": "claude"},
  {"verb": "submit", "ref": "@form1"}
]
EOF
heso stamp plan.json > plat.json          # plan → plat (records cassette)
heso run plat.json > plat-replay.json     # plat → plat (off-network, byte-identical)
heso replay plat.json                     # plat → step log (pure read, no execution)
heso unpack plat.json > plan-again.json   # plat → plan (edit, restamp)

The plat's plat_hash (BLAKE3 over canonical JSON via RFC 8785) commits to the plan, the observed content, AND the embedded cassette. Tamper with any of them and the hash no longer matches; heso plat-verify will say so. Two different <url> inputs always produce different plat_hash values — the URL is part of the hashed canonical bytes, and a regression test in crates/heso-engine-fetch/src/plat.rs::tests pins that invariant against future drift.

Replay a published plat in one command. Install heso (uv tool install heso / pipx install heso / npm install -g @ixla/heso), then:

curl -sL https://github.com/blank3rs/heso/releases/download/v0.0.10/replay-demo-1-goldfinger.plat.json \
  | heso run - \
  | jq -r .plat_hash
# → d93c08ba32b762dd6e47091a1d4bd4aa4d8308dbdbf44869f81146a3f5b8033a

That hash is BLAKE3 over the canonical bytes of the resulting plat. Anyone, any machine, any time — same hash. The cassette inside the plat carries every HTTP response the engine touched when it was stamped against the live Wikipedia article. No network is involved in heso run itself.

Three sample plats live as release assets on v0.0.10:

• replay-demo-1-goldfinger.plat.json — Wikipedia Goldfinger_(film) (1 MB plat, hash d93c08ba…)
• replay-demo-2-torvalds-bio.plat.json — Wikipedia Linus_Torvalds (1 MB plat, hash 27e66b0d…)
• replay-demo-3-rust-lang-rust.plat.json — github.com/rust-lang/rust (640 KB plat, hash 201e9410…)

Recover from broken sites.

• --best-effort on open / read / wait — exit 0 even when scripts crash. Output includes partial: true, partial_reason: "script_crash" | "wait_timeout" | "fetch_failed" | "parse_error", and failed_scripts: [...]. The agent sees what broke and decides what to try next.
• --inject-script "<inline-js>" or --inject-script @file.js — run JS before the page's own scripts. Use it to shim a missing global (the canonical window.lunr cascade kind of thing).

Detect cross-call state changes.

• heso read always returns a content_hash. Pass --since <prev_hash> to get a delta describing what changed (actions_added, actions_removed, forms_changed, text_changed, title_changed).

Honest about failure.

• Every open / read / fetch response carries http_status (200, 403, 503, ...) — captured pre-body-consumption so 4xx/5xx pages never come back wearing a 200 mask. Cloudflare-style "Just a moment..." interstitials are detected via __cf_chl_opt / challenge-token markers and surfaced as partial_reason: "bot_challenge". No more silent "I got something" when the server returned an error page.
• heso click @e7 on an <a href="..."> actually follows the link — the response carries the destination page's title, tree, actions, and http_status, not the source page.

Web platform coverage.

• XMLHttpRequest (sync + async, backed by the same reqwest client as fetch), performance.mark / performance.measure, document.getElementsByClassName / getElementsByName / getElementsByTagName, 60+ HTMLElement subclass constructors (new HTMLDivElement() works, instanceof HTMLScriptElement works), element.style = "color: red" string-coercion setter, data: URL fast path in <script src>.
• MutationObserver + IntersectionObserver fire on real DOM mutations and viewport intersections; setTimeout / setInterval accept the 1-arg form per WHATWG HTML; classic <script> runs sloppy-mode per spec (so sites like Apple and Wikipedia that use var = ... at the top level work); ES modules (<script type="module">) stay strict per ECMA-262.

Stateful sessions.

• heso serve — JSON-RPC over stdin/stdout. Cookies, DOM mutations, listeners, and history persist across calls. Useful for login → navigate → scrape flows.

What it can't do

• No rendering. No canvas, WebGL, CSS layout, or video. If the meaning is in pixels, use a real browser.
• CAPTCHAs and hard bot-detect. Hits one, stops. The default user-agent is heso/<version> so anything fingerprinting will see us coming. We detect Cloudflare interstitials and surface them as partial_reason: "bot_challenge" rather than pretending the page loaded.
• Pages built on tech we don't simulate. Service Workers, WebRTC, WebUSB, WebBluetooth — not supported.
• Sites whose JS we can't run. QuickJS isn't V8. Most works; some doesn't.
• Sibling-script cascades we haven't shimmed. When script A sets window.X and script B reads it, and X doesn't exist on first load, heso surfaces the crash and the agent can --inject-script a stub.

Use as a library

The Python (heso) and Node (@ixla/heso) packages each ship two faces of the same bundled binary: a CLI on $PATH and a programmatic API that spawns that binary under the hood and gives you back parsed JSON as native objects. No FFI, no Python extension module, no N-API addon — subprocess + JSON is the contract.

# Python
import heso

page    = heso.open("https://example.com")              # -> dict
results = heso.search("rust web scraping", limit=5)     # -> dict
content = heso.read("https://example.com", complete=True)

# Stateful flow over one long-lived `heso serve` process:
with heso.session() as s:
    s.open("https://example.com")
    s.click(text="More information...")
    page = s.read()

// Node
import { open, search, read, session } from "@ixla/heso";

const page    = await open("https://example.com");
const results = await search("rust web scraping", { limit: 5 });
const content = await read("https://example.com", { complete: true });

await session(async (s) => {
  await s.open("https://example.com");
  await s.click({ text: "More information..." });
  const page = await s.read();
});

Per-language idioms: Python is snake_case + sync, Node is camelCase + Promises. Full API at heso.ca/docs.

Examples

Search the web, then read the top hits in parallel:

heso search "rust web scraping" --limit 5
heso batch read url1 url2 url3 --parallel 2

Read everything from one page in one call:

heso read https://nextjs.org/
# → { title, text, actions, forms, cookies, console, framework,
#     content_hash, lazy_hints, partial: false, ... }

Find by visible text, click, follow:

heso click https://news.ycombinator.com --text "More"

Wait for an SPA condition:

heso wait https://app.example.com/ --selector-exists ".dashboard" --timeout 5s

Rescue a broken site with a polyfill:

heso open https://shoelace.style --best-effort \
  --inject-script "window.lunr = (() => ({ Index: { load: () => ({}) } }))()"

Multi-step session over stdio:

heso serve
# → JSON-RPC. Page state, cookies, DOM all persist across requests.

Reproducibility (same seed → same output across machines):

heso eval-js --seed 42 'Math.random()'   # 0.5140492957650241
heso eval-js --seed 42 'Math.random()'   # 0.5140492957650241

Signed receipts

Every heso open / heso read call can emit a signed receipt alongside its JSON output — an Ed25519-signed envelope describing what was run, what came back, and the BLAKE3 trace hash. The recipient verifies the signature against an allowlist of trusted public keys (or rejects the receipt). Per ADR 0005 + ADR 0008.

One-time setup — generate a local Ed25519 identity:

heso identity init
# → {"path": "heso-local-data/identity.key", "public_key": "fdibx2...IE=", "algorithm": "Ed25519"}

Sign a receipt on every call by passing --receipt PATH:

heso open https://example.com/ --receipt receipt.json
# stdout: the normal page JSON
# receipt.json (sibling file):
# {
#   "trace": [{"op": "cd", "target": {"kind": "url", "url": "https://example.com/"}}],
#   "results": [{"op": "cd", "url": "https://example.com/"}],
#   "trace_hash": "7e501fac...",
#   "seed": 0, "mode": "deterministic", "cost": {...},
#   "signature": {"algorithm": "Ed25519", "public_key": "fdibx2...IE=", "signature": "bNBb...Cg=="}
# }

Verify the receipt — bind it to a trusted signer with --trusted-keys:

# trusted.json is a JSON array of base64 pubkeys you accept signatures from.
echo '["fdibx2rLqGfrIf+duGbRKlM1iPwVSynHUq+nEisjwIE="]' > trusted.json

heso receipt-verify --trusted-keys trusted.json receipt.json
# → OK fdibx2rLqGfrIf+duGbRKlM1iPwVSynHUq+nEisjwIE=
# exit 0

Or via the HESO_TRUSTED_KEYS=<path> env var if you'd rather not pass the flag every call.

Verify enforces three rejections:

# 1. Tampered receipt — any byte change invalidates the signature
sed -i 's/"seed": 0/"seed": 999/' receipt.json
heso receipt-verify --trusted-keys trusted.json receipt.json
# → INVALID: signature verification failed       (exit 1)

# 2. Wrong signer — receipt is well-formed but the pubkey isn't allowlisted
heso receipt-verify --trusted-keys other_keys.json receipt.json
# → INVALID: signing pubkey `...` is not in the trusted-keys allowlist   (exit 1)

# 3. `mode: live` — live runs use real time + real network and aren't
#    replay-safe, so the signature has no replay value (ADR 0008)
heso open https://example.com/ --receipt live.json --mode live
heso receipt-verify --trusted-keys trusted.json live.json
# → INVALID: receipt `mode: live` is not replay-safe — per ADR 0008 ...   (exit 1)

Verify without an allowlist still works for backwards compatibility, but emits a stderr warning so the missing trust anchor isn't silent:

heso receipt-verify receipt.json
# stderr: warning: no pubkey allowlist configured (pass --trusted-keys PATH or set HESO_TRUSTED_KEYS ...)
# stdout: OK fdibx2...IE=
# exit 0

Exit codes: 0 valid + (allowlist empty OR pubkey allowlisted), 1 invalid (tampered, wrong signer, or mode: live), 2 missing/malformed receipt or --trusted-keys load failure.

Error handling

Both libraries throw a structured error (HesoError in Python, HesoError extends Error in Node) when the binary exits non-zero. Fields on the error tell you what to retry:

import heso
try:
    page = heso.read("https://shoelace.style")
except heso.HesoError as e:
    print(e.returncode, e.stderr[:200])  # exit code + first 200 chars of stderr

import { read, HesoError } from "@ixla/heso";
try {
  const page = await read("https://shoelace.style");
} catch (e) {
  if (e instanceof HesoError) {
    console.error(e.code, e.stderr.slice(0, 200));
  }
}

For sites that crash some scripts, use best_effort / bestEffort instead — heso exits 0 with a partial: true envelope so you handle the failure as data, not an exception:

page = heso.read("https://shoelace.style", best_effort=True)
if page["partial"]:
    print("got partial:", page["partial_reason"], page["failed_scripts"])

Plug into agent harnesses

heso is harness-agnostic. The same package serves five integration patterns:

Harness style	How heso fits
Python frameworks (LangChain, Pydantic AI, LangGraph, smolagents, AgentScope)	import heso. Each function returns a dict. Wrap with @tool / Tool(...) / a function schema.
Node / TS frameworks (Mastra, Vercel AI SDK, LangGraph.js, Stagehand, Browser Use TS)	import { open, search } from "@ixla/heso". All async; TypeScript types ship in index.d.ts.
Skill-markdown harnesses (Claude Code, Cursor, Aider, Cline, Continue, Windsurf)	Drop the manifest in the "Use as an agent skill" block below into ~/.claude/skills/heso/SKILL.md (or the harness's skills dir). The harness auto-discovers; heso on PATH does the rest.
CLI-spawning harnesses (Aider, shell-script agents, homegrown loops)	Same heso <verb> ... CLI used by both libraries. JSON on stdout. No special integration.
Long-running JSON-RPC harnesses	heso serve is a JSON-RPC 2.0 server over stdin/stdout. Cookies + DOM state persist across calls.

The verbs are the contract (see ADR 0017) — no heso-specific framework dependency, no adapter layer.

Use as an agent skill

heso is built to be a tool an agent calls, not a library a human drives. The cleanest integration is the skill markdown pattern that Claude Code, Cursor, Aider, Cline, and similar harnesses use:

---
name: heso
description: Use the heso headless browser (one Rust binary, no Chromium, no Node) to search the web, fetch pages, run their JavaScript, extract content, navigate, fill forms, or click links. Prefer this over WebFetch when you need a DOM, stateful clicks, or framework-rendered content.
---

## Verbs

- `heso search "<query>" [--limit N]` — web search via DDG + Wikipedia
- `heso open <url>` — page summary
- `heso read <url> [--complete]` — full content + actions + forms (use --complete for lazy-loaded sites)
- `heso wait <url> --selector-exists ".x"` — block until a condition is true
- `heso batch [open|read] <urls...> [--parallel N]` — parallel scrape
- `heso click <url> --text "..." | --selector "..." | @eN` — click
- `heso fill <url> @eN "value"` — type into input
- `heso submit <url> @eN` — submit form
- `heso eval-dom <url> "<js>"` — run JS against the page
- `heso serve` — multi-step JSON-RPC session
- `--best-effort` on open/read/wait — exit 0 on partial failures, surface what broke
- `--inject-script "<js>" | @file` — inject a polyfill before page scripts run

The verbs are the contract. Same shape works in any harness that does tool or skill markdown.

Stats

Measured on Windows 11, AMD x86_64, with the release binary:

Thing	Number
Binary size	9.2 MB
Cold start (open https://example.com, network included)	~80 ms
Engine-only (no network, local fixture)	~35 ms
Batch (8 URLs, --parallel 8)	~1.1 s total
Search (DDG, 5 results)	~1 s

No comparisons to other tools — different tools have different tradeoffs and "X is faster than Y" framing rarely survives contact with a real workload.

Building from source

If you're on Linux/macOS today (v0.0.2 will ship prebuilt binaries) or want to hack on heso itself:

git clone https://github.com/blank3rs/heso
cd heso
cargo build --release -p heso-cli
./target/release/heso search "rust web scraping" --limit 5

Requires Rust 1.80+ (rustup from https://rustup.rs).

Status

Pre-alpha. v0.0.1 is on every registry. Worth trying if the use case fits; not worth depending on in production yet. Next (v0.0.2) ships Linux + macOS binaries and the library APIs above.

License

MIT or Apache-2.0, your choice.

---

Full docs: heso.ca/docs · Site: heso.ca · npm: @ixla/heso · PyPI: heso