Kuri – Zig based agent-browser alternative

Apr 22, 2026•sorcercode•View Original

TL;DR Highlight

A single Zig binary for agent browser control that cuts snapshot tokens 7-12%

Who Should Read

Backend developers running browser agents who are bothered by Playwright/Node dependencies and snapshot token costs

Core Mechanics

One Zig binary, no Playwright
- It runs as a single binary with no Node.js runtime, per the README.
- kuri is 1.04 MiB and kuri-agent is 615 KiB (Apple M4 Pro, ReleaseFast build).
Four modes in one project
- kuri is a CDP (Chrome DevTools Protocol, the remote-control protocol for Chrome) server.
- kuri-fetch fetches pages with QuickJS and no Chrome.
- kuri-browse is a terminal browser and kuri-agent is a scriptable CLI.
Shorter snapshot format saves tokens
- Refs are @e0 (3 chars) instead of [ref=e0] (9 chars), and the per-line '-' prefix is dropped.
- Versus agent-browser, Hacker News is 12% fewer tokens and Google Flights SIN→TPE is 7% fewer (chars/4 approximation).
/page/state observation is 44x lighter
- On the same Google Flights page, /page/state is 48 tokens and a full snapshot is 2,124.
- It returns only URL, title, scroll %, and counts of forms and links.
POST /batch cuts round-trips
- It sends N commands in one HTTP call, removing N-1 round-trips and LLM turns.
- Token count stays the same, but HTTP calls drop from 5 to 1.
Versus libretto it's a mixed result
- Per-call latency is 4–117 ms for kuri and 1,344–1,500 ms for libretto (same Chrome and tab, tiktoken o200k_base).
- Default snapshot of a large list is 4,424 tokens for kuri and 813 for libretto, so kuri loses there.
- For repeat runs libretto compiles to a Playwright script for 0-token replays.

Community Reactions

Snapshot diffs made a 9-click run 49x cheaper
- A 9-click feed trajectory with naive full re-snapshots costs 44,285 tokens.
- A truncated base plus diffs costs 898 tokens, versus libretto's 939.
A commenter who used brow.sh for page fetching found this interesting but asked if it's just a Chrome DevTools wrapper
- They question whether kuri-fetch is really standalone if Chrome must run somewhere.
A commenter reported that the bun and sh install instructions return a 404
- It suggests the install commands may not work as written.
A commenter noted the README benchmark is self-published
- It claims 16% lower workflow token cost than Vercel's agent-browser on a Google Flights task.

How to Apply

Swap your observation step for /page/state
- Find where your agent loop takes a full snapshot at every step.
- Replace only the post-click check with /page/state (about 48 tokens) and measure the difference.
Request snapshots as interactive+compact
- Call /snapshot?filter=interactive&format=compact to get only clickable elements.
- Compare token counts against your current tool on the same page.
Set limit on long list pages
- Use /snapshot?limit=N to truncate repeated items.
- Re-capture the part you need with scope=@ref.
Rerun the benchmark in your own environment before adopting
- The README says to rerun cross-tool numbers before quoting a percentage.
- Compare tokens and latency against your current tool on 3 real pages of your service.

Try It

copy & run

# Run in a shell. Installs Kuri, opens Hacker News, and compares snapshot vs page/state size (Chrome required)

curl -fsSL https://kuri.trilok.ai/download | sh
kuri &
sleep 3
BASE=http://127.0.0.1:8080
SESSION=demo
curl -s -H "X-Kuri-Session: $SESSION" "$BASE/tab/new?url=https%3A%2F%2Fnews.ycombinator.com"
echo "--- snapshot chars:"
curl -s -H "X-Kuri-Session: $SESSION" "$BASE/snapshot?filter=interactive&format=compact" | wc -c
echo "--- page/state chars:"
curl -s -H "X-Kuri-Session: $SESSION" "$BASE/page/state" | wc -c

Terminology

CDPChrome DevTools Protocol. It's the remote-control signal spec that DevTools uses to talk to Chrome.

A11y snapshotA text summary of the page from the accessibility tree. Like a table of contents with only buttons and links instead of full HTML.

ref (@e0)A short ID tag on each snapshot element. The agent says 'click @e3' to act on it.

HARA file format recording a browser's network requests and responses. Like a black-box recording of traffic.

QuickJSA lightweight JavaScript engine. Used to run page JS without Chrome.

Related Resources

Kuri – Zig based agent-browser alternative

Apr 22, 2026•sorcercode•View Original

TL;DR Highlight

A single Zig binary for agent browser control that cuts snapshot tokens 7-12%

Who Should Read

Backend developers running browser agents who are bothered by Playwright/Node dependencies and snapshot token costs

Core Mechanics

One Zig binary, no Playwright
- It runs as a single binary with no Node.js runtime, per the README.
- kuri is 1.04 MiB and kuri-agent is 615 KiB (Apple M4 Pro, ReleaseFast build).
Four modes in one project
- kuri is a CDP (Chrome DevTools Protocol, the remote-control protocol for Chrome) server.
- kuri-fetch fetches pages with QuickJS and no Chrome.
- kuri-browse is a terminal browser and kuri-agent is a scriptable CLI.
Shorter snapshot format saves tokens
- Refs are @e0 (3 chars) instead of [ref=e0] (9 chars), and the per-line '-' prefix is dropped.
- Versus agent-browser, Hacker News is 12% fewer tokens and Google Flights SIN→TPE is 7% fewer (chars/4 approximation).
/page/state observation is 44x lighter
- On the same Google Flights page, /page/state is 48 tokens and a full snapshot is 2,124.
- It returns only URL, title, scroll %, and counts of forms and links.
POST /batch cuts round-trips
- It sends N commands in one HTTP call, removing N-1 round-trips and LLM turns.
- Token count stays the same, but HTTP calls drop from 5 to 1.
Versus libretto it's a mixed result
- Per-call latency is 4–117 ms for kuri and 1,344–1,500 ms for libretto (same Chrome and tab, tiktoken o200k_base).
- Default snapshot of a large list is 4,424 tokens for kuri and 813 for libretto, so kuri loses there.
- For repeat runs libretto compiles to a Playwright script for 0-token replays.

Community Reactions

Snapshot diffs made a 9-click run 49x cheaper
- A 9-click feed trajectory with naive full re-snapshots costs 44,285 tokens.
- A truncated base plus diffs costs 898 tokens, versus libretto's 939.
A commenter who used brow.sh for page fetching found this interesting but asked if it's just a Chrome DevTools wrapper
- They question whether kuri-fetch is really standalone if Chrome must run somewhere.
A commenter reported that the bun and sh install instructions return a 404
- It suggests the install commands may not work as written.
A commenter noted the README benchmark is self-published
- It claims 16% lower workflow token cost than Vercel's agent-browser on a Google Flights task.

How to Apply

Swap your observation step for /page/state
- Find where your agent loop takes a full snapshot at every step.
- Replace only the post-click check with /page/state (about 48 tokens) and measure the difference.
Request snapshots as interactive+compact
- Call /snapshot?filter=interactive&format=compact to get only clickable elements.
- Compare token counts against your current tool on the same page.
Set limit on long list pages
- Use /snapshot?limit=N to truncate repeated items.
- Re-capture the part you need with scope=@ref.
Rerun the benchmark in your own environment before adopting
- The README says to rerun cross-tool numbers before quoting a percentage.
- Compare tokens and latency against your current tool on 3 real pages of your service.

Try It

copy & run

# Run in a shell. Installs Kuri, opens Hacker News, and compares snapshot vs page/state size (Chrome required)

curl -fsSL https://kuri.trilok.ai/download | sh
kuri &
sleep 3
BASE=http://127.0.0.1:8080
SESSION=demo
curl -s -H "X-Kuri-Session: $SESSION" "$BASE/tab/new?url=https%3A%2F%2Fnews.ycombinator.com"
echo "--- snapshot chars:"
curl -s -H "X-Kuri-Session: $SESSION" "$BASE/snapshot?filter=interactive&format=compact" | wc -c
echo "--- page/state chars:"
curl -s -H "X-Kuri-Session: $SESSION" "$BASE/page/state" | wc -c

Terminology

CDPChrome DevTools Protocol. It's the remote-control signal spec that DevTools uses to talk to Chrome.

A11y snapshotA text summary of the page from the accessibility tree. Like a table of contents with only buttons and links instead of full HTML.

ref (@e0)A short ID tag on each snapshot element. The agent says 'click @e3' to act on it.

HARA file format recording a browser's network requests and responses. Like a black-box recording of traffic.

QuickJSA lightweight JavaScript engine. Used to run page JS without Chrome.

Kuri – Zig based agent-browser alternative

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

Kuri – Zig based agent-browser alternative

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources