Sort conversations into Projects
Give it a plan file. It moves each conversation, checks that it landed, and slows down when ChatGPT says too many requests.
Ask ChatGPT, Claude, Gemini, DeepSeek, Grok and more in your own logged-in Chrome, from the terminal or from your agent. One provider or all of them at once. Your agent gets the reply, not the page.
It depends mostly on two things: how rows arrive, and what your queries look like. With 40M rows you have two realistic choices, BRIN and B-tree, and many setups end up using both.
BRIN fits when data is inserted roughly in time order (append-only logs, sensor readings, events). BRIN stores only the min and max timestamp for each block range. The index ends up kilobytes to a few MB, compared with roughly 1GB or more for a B-tree on 40M rows, and it costs almost nothing on inserts. It’s great for wide range scans like “all rows from last week” or “aggregate March”. It’s poor for very selective lookups, because it can only skip blocks, not find rows.
sql CREATE INDEX ON readings USING brin (ts) WITH (pages_per_range = 32);
For a 40M-row time series table, the best index depends entirely on how the data is physically ordered on disk. BRIN is the ideal choice if rows are appended in time order (e.g., an INSERT-only log table), while B-tree is the safer general-purpose choice if physical order is not guaranteed or your queries need precise, low-latency point lookups.
🥇 BRIN (Block Range Index): The Time-Series Specialist
If your 40M rows are stored roughly in chronological order (common for append-only time series), BRIN is dramatically more efficient than a standard index.
BRIN on the timestamp column.
For a 40M-row time series table with mostly chronological (append-only) inserts, a BRIN index is usually the best fit:
SQL CREATE INDEX CONCURRENTLY idx_ts_brin ON your_table USING brin (ts) WITH (pages_per_range = 64, autosummarize = on); Why BRIN fits Extremely small (often KB–low MB vs hundreds of MB or more for a B-tree) Low write and maintenance cost Strong for time-range queries (ts BETWEEN ... or ts > ...)
The bridge opens Chrome on its own profile. You sign in to each chat app one time, and every repo and every agent reuses it.
Opens a Chrome window on a separate profile. Sign in to the chat apps you use. The sessions stay there for every repo.
bridge chrome startName one provider or a list to fan out. Agents connect over MCP and call the same ask tool.
bridge serve# one provider you@mac ~ %PS C:\Users\you>you@ubuntu:~$ bridge ask --provider claude --json "Review this migration" # fan out to three you@mac ~ %PS C:\Users\you>you@ubuntu:~$ bridge ask --provider claude,deepseek,grok --json "same question, three models" # give Claude Code the bridge as MCP tools you@mac ~ %PS C:\Users\you>you@ubuntu:~$ claude mcp add ai-browser-bridge -- bridge serve
Each provider returns its reply text as JSON, ready for a script or an agent to read.
--json[
{ "provider": "claude", "ok": true, "text": "It depends mostly on…" },
{ "provider": "deepseek", "ok": true, "text": "For a 40M-row time…" },
{ "provider": "grok", "ok": true, "text": "BRIN on the timestamp…" }
]Short loops of real commands. Each one plays while it is on screen.
Give it a plan file. It moves each conversation, checks that it landed, and slows down when ChatGPT says too many requests.
Search titles and message content from the terminal. Matches come back as JSON.
List providers after --provider and get every reply back, keyed by provider.
Pick read-only, ask or auto. Every patch gets a checkpoint you can restore.
Type @ to pick a repo file. It is expanded into the prompt before the chat sees it.
Send a prompt to Google Flow, then download the clips into .bridge/ in your repo.
A chat page is thousands of lines of markup, scripts and old messages. The bridge reads the reply inside Chrome and hands back only the text, so your agent keeps its context for your code.
Your agent can call the chat apps. ChatGPT, Claude and Grok can read your repo through a small set of checked tools, never a shell.
Drive web chats: one prompt fanned across providers, or a `tasks` array of independent Conversations run in parallel (new or resumed). Returns an ordered, paginated result — one row per task with its reply and reopenable Conversation id/url.
Search provider conversation history by title/id and return matching conversation URLs.
Generate a Veo clip from a Start keyframe image + a shot prompt (image-to-video), then download the mp4.
List the rendered clips in the current Flow project (id + mp4 URL).
List the Flow projects in the sidebar (id + title + URL).
Download clip mp4s to the target repo's .bridge/downloads/flow directory (all clips, or the given clipIds).
Move a clip to Flow's recoverable Trash (requires confirm:true).
Rename a clip.
Add a clip to a scene (Flow's 'Add to scene' / extend).
Add a clip back to the prompt as input ('Add to prompt').
Rename the current Flow project.
Permanently delete the current Flow project (requires confirm:true; not a Trash move).
List the reference images (ingredients) attached to the current prompt.
Detach one ingredient from the current prompt.
Detach every ingredient from the current prompt.
Inspect the current ChatGPT render: streaming?, generated-image progress (loaded/pending), misfire and rate/cap-limit signals, and the latest assistant text. Pass allTabs:true to sweep every ChatGPT tab in the browser.
Where Claude Design is: home or project page, project, model + effort, whether a turn is running, open Design tabs, the actions available here, and the Share destinations left to you.
List Claude Design projects (or design systems) with id, name, url, sharing.
Choices for new work: home templates, models with effort levels (incl. More models), and design systems.
Pick the model (incl. More models) and effort in a project's composer, or on home for new projects. Like the UI, it becomes your Claude Design default.
List the files in a Design project.
Read a project's Conversations and the latest messages — what you asked and what Claude replied.
Open (or focus) the project's tab in the bridge Chrome.
Create a Design project. With a prompt: home composer with optional template, model, effort, and attachments, then wait for Claude's first reply. Without: a blank project.
Send a message in a project's Conversation (optional model, effort, design systems, attachments) and return Claude's reply. status is replied, running, or waiting-for-answer. For long turns pass wait:false, then poll design_read_conversation.
Start a new Conversation in a project and return its id.
Rename a Conversation in a project.
Rename a project, star/unstar it, or set its design systems.
Duplicate a project and return the copy's id.
Permanently delete a project (requires confirm:true).
Add repo files to a project; an existing path is replaced.
Delete files from a project (requires confirm:true).
Download project files, or the whole project as a zip, into the repo (.bridge/downloads/design by default).
Set who can open the project (private or workspace, with view/comment/edit) and return its link.
42 tools in all. Design tools that delete need confirm: true.
Search the repository using ripgrep. Locate symbols, imports, routes, tests, configs, and references.
Read a repo file with line numbers. Use after grep_code before proposing edits.
Apply a unified diff patch to the repository. Use only after reading the relevant files.
Run an allowed project test command (npm test, pytest, go test, etc.).
Show the current git diff and diff stat for the working tree.
List captured attachments in a ChatGPT conversation, including their assistant/user role.
Download one captured attachment from the active ChatGPT conversation.
Download all or selected captured attachments from the active ChatGPT conversation.
Calls reach your Mac through a Cloudflare quick tunnel. apply_patch snapshots each file first, so you can restore it.
No. The bridge drives the chat apps in your logged-in Chrome, so every request runs on your own account and plan.
macOS, Windows and Linux, with Node 24.11 or newer and Google Chrome.
Each provider keeps its selectors in one browser layer, so a fix stays in one place.
No. There is no shell. Repo tools are a fixed set, checked and kept inside the repo, under the permission mode you pick.
On your computer. The Chrome profile lives in ~/.ai-browser-bridge/chrome-profile and repo state in .bridge/ inside the repo.
Install the CLI, sign in once, and ask from your terminal or your agent.