bside

b(oost)side — drive the Aside browser with TypeSafe Jev decisions instead of a chat LLM.

Jev is a System One model: it never generates text. Each tick it answers typed questions — which action, which element, is the goal met — with calibrated probabilities and confidence, in ~0.3–1s at ~$0.042/1M input tokens. Actions come from a schema you enumerate, so the pilot cannot hallucinate a tool call.

agent ── browse(goal) ──▶ bside ──▶ aside repl (observe/act)
                            └──▶ Jev /v1/systemone (decide)

Benchmark

Task: "Starting from google.com, search for wikipedia, open the Wikipedia site, search for Batman on Wikipedia, then in the Batman article click the link to the Gotham City article" — same browser, same loop, only the decision engine differs:

engine	wall time	steps	avg decision latency	est. cost
jev (bside)	52.8s	7	~1.0s	$0.0018
gpt-6-astra (frontier LLM)	72.9s	7	~3.6s	$0.039
gpt-5.6-luna (light LLM)	68.9s	7+retries	~2.5s	$0.011

Browser latency (page loads, snapshots, action execution) is excluded from the decision column — those numbers measure only model inference. Luna additionally attempted clicks on elements that do not exist and retried; Jev cannot emit a ref outside the enumerated candidate set. In the event log (bench/example-results.json) these show up as act_start/act_end pairs completing in ~16 ms (vs ~0.7–1.8 s for real clicks) and as extra call cycles without a matching act — the failure signature of a generated ref that isn't on the page.

Reproduce (model names shown are the ones measured; any Responses-compatible model works — results will differ):

export BSIDE_LLM_URL=https://api.openai.com/v1/responses   # or any Responses-compatible endpoint
export BSIDE_LLM_API_KEY=...
node bench/run.mjs '{"goal":"...","urls":["https://www.google.com"],"inputs":["wikipedia","batman"],"maxTicks":12}' \
  --models gpt-6-astra@1.25/10,gpt-5.6-luna@0.25/2
open bench/player.html   # side-by-side real-time replay

Install

npm i -g bside        # or: npx bside ...

Requirements: aside CLI signed in, and a TypeSafe API key (TYPESAFE_API_KEY or ~/.config/typesafe/api-key).

For the llm comparison engine, point at any OpenAI Responses-compatible endpoint: BSIDE_LLM_URL, BSIDE_LLM_API_KEY (or OPENAI_API_KEY), BSIDE_LLM_MODEL (or --model).

CLI

bside '{"goal":"Open the top story comments on Hacker News","urls":["https://news.ycombinator.com"]}'
bside --goal "Search Google for typesafe ai and open the first result" \
      --url https://www.google.com --input "typesafe ai"

Options: --attach <targetId> (drive an existing tab), --active (current tab), --ticks N, --linger <ms> (keep the tab visible after finishing).

MCP

bside --mcp

{ "mcpServers": { "bside": { "command": "bside", "args": ["--mcp"] } } }

Exposes one tool: browse(goal, urls?, inputs?, mode?, targetId?, maxTicks?).

Design notes / limits

• No text generation: type/navigate arguments come from inputs/urls candidate lists. Free text needs an upstream LLM or human.
• Text-only state: decisions read the accessibility tree, not screenshots.
• Per-tick judge, not a planner: long-horizon goals need a decomposer upstream; bside escalates on low confidence / blocked / loops.
• Each run is one aside repl call (~120s cap); session tabs close on exit. Use --attach/--active to drive persistent browser tabs.

Bonus: ChatJev

Generation is also just decisions — chat/ makes Jev "talk" without any text generation. Three modes:

• enum (--enum, default in chat/web.mjs): the faithful analog of the browser pilot — a dumb grammar mechanically enumerates ~24 complete candidate sentences (the "DOM of language"), Jev picks the best one, and a noul judge verifies it actually answers the question. On failure it resamples a fresh pool. ~2 calls per reply.
• naive (chatjev.mjs): every word is one choice over a ~500-word vocabulary, sampled from Jev's probability distribution. ~1s/word.
• structured (--struct, ?mode=struct in the web UI): plan a sentence skeleton → fill each slot inside its vocab group → verify with a noul judge → repair the broken slot.

node chat/chatjev.mjs "who are you?" --enum
# grammar enumerated 24 candidates → pick "I am an llm." (p=0.99) → verify ✓ 0.77
node chat/web.mjs   # → http://localhost:4317

MIT.