v2026.07.22
Highlights
Live Prices in Chat
Markets often move faster than the agent reasons — a long turn can outlast the price it started from. This release adds tools that keep it caught up to the live quote.
get_quote returns quotes and nothing else — up to 20 tickers per call, stocks or indices — and is documented to the agent as cheap, to be called whenever it needs a current price, including a re-check immediately before stating one in an answer. Results render as an inline quote card. Daily prices, company overviews, and the market overview open with a freshness line ([Live: AAPL $233.45 (+2.31%) — as of 14:32:05 ET, market open]), so a number carries its own timestamp and session. get_market_overview returns a single-day snapshot of index and sector movement.
Market Watch, in Settings → Experiments, pushes prices to the agent instead of waiting for it to ask. It keeps a per-thread watchlist live: the agent registers the tickers, and a compact price block is stamped onto each model call and refreshed about every 25 seconds during a turn, for up to 10 symbols. A Watching NVDA, TSLA chip appears above the composer, and a Status tab in the right panel shows current quotes. Watchlists are venue-aware: a Hong Kong watchlist stays live during HK hours while US markets are closed, including the HK lunch break. Stamps are ephemeral and never enter saved thread history, so reloads and reconnects cannot resurface a stale price.
Market Watch is opt-in and PTC-only. The toggle states its trade-off: re-injected price context can distract some models from the main task.
New Web Search Providers and Site Crawling
Exa and Parallel are selectable alongside Tavily, Serper, and Bocha in Settings → Preferences → Web Search Provider, each with its own search depths. Search, fetch, and crawl now resolve through a single provider manifest that declares each provider's capabilities and levels.
Site crawling is new and opt-in from the Experiments tab. WebCrawl walks a site and writes each page to your workspace as a markdown file with an accompanying index.jsonl, so the agent can grep and cross-reference the whole set as local files. WebMap lists a site's URLs without retrieving content. Both are PTC-only.
Kimi K3
Moonshot's flagship is available in the model picker: 1,000,000-token context, text and image input.
Experiments
A new Experiments tab in Settings lists early-access features you can enable per account. Market Watch and site crawling are the two available now, both off by default. Each entry shows its trade-offs, and features can be switched off at any time.
Changelog
New Features
- intraday market-watch live-price feed + user-facing feature flags (#325) (e62b1cd)
- web: web provider protocol — Exa/Parallel search, delegated fetch chain, Firecrawl site crawl (#327) (1df8a74)
- llms: add Kimi K3 flagship model (#326) (ec742cd)
- sandbox: separate public base for service URLs, harden token refresh (f3996bc)
- web: env-configurable entry path, and VITE_ACCOUNT_URL → VITE_PLATFORM_URL (c2e8c97)
Bug Fixes
- resolve sandbox ginlix-data token path from work dir, not $HOME (814f7d6)
Full Changelog: v2026.07.17...v2026.07.22
5 features, 1 fix — 6 changes total