Every release, newest first. Updates land in the app's Essentials panel too.
v0.2026.07.17.00.00.oss_00
New features
Install packages with approval — the agent has a dedicated install_package action for adding dependencies (pip, npm, brew, cargo, apt, and more). It picks the right install command for the manager you name, shows you the exact command for approval before anything runs, and works in both the app and the voice overlay. No more improvised, ad-hoc install one-liners.
The agent can run your dev server — commands that are meant to keep running (npm run dev, vite, cargo watch, a test watcher) now start in the background instead of hanging the turn until they're killed. The agent gets the startup output, notices the URL your app is serving on, and can check back later to see whether it compiled cleanly or crashed — and can answer a CLI's y/n prompt while it runs.
The voice overlay tells you when your app breaks — a new chip beside the composer shows whether what you're building is actually up (localhost:3000) or has crashed, and stays there after the turn ends, because your dev server does. Click it for a line on each background process. If read-aloud is on, a crash is spoken unprompted — so you find out your build broke without watching a log. It waits its turn: it won't talk over an answer or over you.
The agent checks your app instead of assuming it works — ask whether something works and the agent opens the page in a real browser you can see, then reports back what actually happened: uncaught errors, console errors, requests that 404'd, and the page title. On a vision-capable model it also takes a screenshot and looks at it. So "it's running on localhost:3000" becomes "it loads, but there's a TypeError in app.js" — the thing you can act on without ever opening the console. And if read-aloud is on, a broken page is spoken unprompted, the same way a crash is — so you hear "your page throws a TypeError" even if you weren't watching the answer. A red chip beside the composer also flags a broken page at a glance (click it for the reason) and clears itself the moment a later check passes.
Ship a site to a real domain — connect your own Cloudflare account in one click from Bang's MCP settings (per-user OAuth, so Bang never touches your card) and the agent can find an available domain, register it, point DNS at your site, and deploy it. Registering a domain spends real, non-refundable money, so it always stops and asks first: you get the exact domain and price with Approve or Cancel, in the app and in the voice overlay, and nothing is charged without that click. DNS changes and deploys just run. Ask about hosting before Cloudflare is connected and the agent offers to walk you through connecting it, instead of guessing or giving up.
Improvements
See what your agents are doing — orchestration now shows a compact "N Working" indicator right above the composer. Click it to expand a task-named list of every worker agent, open any one to watch its live progress in a read-only panel, and stop a single agent or Stop all from the same place. The delegation cards in the transcript are clickable too, opening the same live view. This replaces the old top pill bar.
Bug fixes
Fewer failed turns from provider blips — a brief upstream hiccup from the model provider (e.g. a transient 520/503 or a dropped connection) is now retried automatically with backoff instead of failing your turn. Real errors (bad key, bad request) still surface immediately.
No more "Thinking…" hangs after an answer — fixed a case where the agent generated its reply but the turn never completed (a slow background write could withhold the completion signal), leaving the voice overlay stuck on "Thinking…" even though the answer was ready. The turn now always finishes as soon as the reply is done.
v0.2026.07.16.15.00.oss_00
New features
Project memory — Bang now remembers durable facts and preferences. Your repo's BANG.md/AGENTS.md (legacy WARP.md still works) and a new auto-discovered .bang/memory.md are sent to the agent on every request, so it actually follows your project's rules. Save a per-project fact anytime with /remember <fact> (written to .bang/memory.md in the repo), approve the agent's Suggested memory chips to keep the good ones, and use saved cloud rules for preferences that should apply everywhere. The agent can now save memory itself too: ask it to remember something and it picks the scope — globally (applies to every project) or to the current project (.bang/memory.md). When you approve a suggested memory you also choose where it lands: Save to project (the default) or Save globally.
Improvements
Custom research subagents — run_subagent now supports named, specialized subagents that the agent picks between by task. Define your own as Markdown files (with a name, description, and instructions) in .bang/agents/, and the agent routes to the best fit; a built-in general researcher is always available. The agent can now delegate several subagents in a single turn and they research in parallel, and a subagent file can pin its own model: (with automatic fallback to your selected model if it's unavailable).
Peek into a research subagent — when the agent delegates with run_subagent, its work now appears as an expandable Sub-agent thread you can open to watch it live — each web search and interim note as it happens — and read its final report inline, instead of just a one-line "delegating…" message.
Hands-off orchestration — in /orchestrate (and /multitask), the lead agent can now split a task across parallel worker agents on its own and redirect a running worker mid-flight when things change — no need to drive each step yourself. Just say what you want done; it coordinates the workers and reports back a synthesis.
Steer running agents — redirect an agent while it's still working, no cancel/restart. Type /steer <redirection> mid-response and the running conversation folds it into its loop on the next step and adjusts course in place; background runs fold it in before they start (or resume as an automatic follow-up if it lands mid-run); and orchestrated child agents pick it up from their inbox. Works for any run or conversation, not just /orchestrate.
New /multitask command — a familiar-named shortcut for /orchestrate that fans a task out across parallel agents.
Bug fixes
File tools work again — fixed a wire mismatch where results from read_files, grep, file_glob, and edit_files were silently dropped, so the agent saw an empty result and got stuck retrying (e.g. "read the README" looping forever). These tools now return their output to the agent.
v0.2026.07.16.00.00.oss_00
New features
Web search — the agent can now search the public web with web_search to pull in current docs, library APIs, error messages, and anything beyond your codebase.
Research subagents — the agent can hand a focused research sub-task to an isolated subagent with run_subagent. It works in its own context (searching the web as needed) and reports back a concise summary, so the main conversation stays on track instead of filling up with intermediate steps.
Improvements
Longer conversations just keep working — as a chat approaches the model's context limit, Bang now automatically summarizes the earlier turns so you can keep going without hitting the wall or losing the thread.
v0.2026.07.14.00.00.oss_00
New features
Browser preview — an embedded browser now lives in the tools drawer, so you can watch what you're building without leaving Bang. The agent can point it at a dev server for you with open_browser_preview.
Improvements
The tools panel is easier to resize by dragging its edge.
An Update badge now appears in the sidebar when a newer build is available — click it to relaunch into the latest.
Read all changelogs in the Essentials panel now opens the full history at trybang.ai/changelog.