Cubex

A self-hosted alternative to Clay that your AI agent can run for you.

A spreadsheet where a column can be an AI prompt, an API call or an enrichment, run with your own keys instead of credits.
Connect Claude Code, Cursor, Codex or Claude Desktop over MCP and ask for the work in plain words.

Install · How it works · Agents and API · Enrichments · Run it on a server

Works with your agent over MCP

                    

Fills columns with

      

Import a CSV, then add columns that fill themselves row by row. Do it by clicking in the browser, or tell your agent:

"Import leads.csv into a new table called Leads. Add an AI column that scores each company 1 to 10 for fit, using /company and /domain. Try it on 5 rows, show me the results, then run it on the rest and tell me how many scored 8 or more."

• AI columns. A prompt for each row that uses other columns, on any of 300+ models through your own OpenRouter key.
• HTTP API columns. Call any JSON API once per row and turn fields of the answer into columns.
• Enrichments. Work emails, phone numbers, and person and company data through treg.to, within limits you set.
• Waterfalls. One value per row from several sources in turn (your other sheets, treg tools, AI prompts, your own APIs), paying only for the steps a row reaches.
• Clean data, free. Fix names, split first and last names, drop Inc and LLC from company names, check emails, turn websites into domains and format phone numbers, with no AI and no credits.
• Formulas, free. Work a column out from each row with a JavaScript formula: pick the work email out of an API response, join names, do math. Describe it and AI writes the formula once; every row after that costs nothing.
• Lookups, free. Bring values from another sheet: a company's industry onto its people, whether a lead is already in another list, how many orders match. Match whole values or text inside them.
• Setups. Save a chain of filled columns (the prompts, models, requests, formulas and rules) once, then build the same columns on a new client's sheet in one go. Each step runs once the columns it reads are filled.
• Remove duplicates, free. Find rows that repeat an email, a domain, a phone number, a company or any columns you pick, see them before anything goes, and keep the first, last or most filled-in row of each. Or keep it that way: auto-dedupe keeps one row per key as rows arrive from a webhook, an import or an agent.
• Webhooks. Every JSON event sent to a sheet's secret URL becomes a new row.
• Agent access (MCP). Your agent can do the work for you, with only the permissions you give it. There's also a REST API.
• A real spreadsheet underneath. Several sheets per table, CSV import and export, sort, filters (by text, numbers, dates or a run's state, like the rows that failed, in and/or groups you can switch on and off), saved views (filters and hidden columns, exported on their own), search across every row, up to a million rows per sheet.
• A trash and a history. Deleted rows, sheets and tables wait 30 days in the trash, ready to restore. Each sheet's History shows what changed and who changed it: you in the app, or which agent.

Your data stays on your machine, in one SQLite file. Cubex only calls OpenRouter, the APIs you set up, and treg.to (what it sends, and when).

Install

On macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/gurmohitghuman/cubex/main/install.sh | bash

On Windows 10 or 11, in PowerShell (no administrator rights needed):

irm https://raw.githubusercontent.com/gurmohitghuman/cubex/main/install.ps1 | iex

It takes a few minutes and needs about 1.5 GB of free memory while it builds. Everything goes into ~/.cubex (on Windows, %LOCALAPPDATA%\cubex), including a private copy of Node.js, so your own Node, if you have one, is never used or changed.

Only this computer can open Cubex until you change that. To open it from other devices on your network, add --public. The installer then generates your password and prints it:

curl -fsSL https://raw.githubusercontent.com/gurmohitghuman/cubex/main/install.sh \
  | bash -s -- --public

& ([scriptblock]::Create((irm https://raw.githubusercontent.com/gurmohitghuman/cubex/main/install.ps1))) --public

• On a server reachable from the internet, keep the default and put HTTPS in front. Run Cubex on your own server walks through it step by step.
• More options: it installs the latest release, and --ref main follows the newest code instead. --port and --dir change the port and the folder, and --help lists everything.
• Coolify, Docker or from source: see Other ways to run Cubex.

Get started

1. Open http://localhost:3002. Whoever opens Cubex first chooses the password, unless INITIAL_PASSWORD is set (--public sets it for you).
2. For AI columns, create a key at openrouter.ai/keys (OpenRouter bills you directly for what your runs use). Add it and pick a default model in Settings → AI. Cubex never picks a model for you.
3. For enrichments, add a treg team API key in Settings → Enrichment. The enrichments guide walks through it.

Want to look around first? Click Try with sample data on the empty dashboard (or in New Table). Cubex makes a table of 25 made-up leads, at .example addresses so none of them is a real person or company, with a few messy values on purpose. A clean column and a formula column are already filled, for free, and a card above the grid shows the duplicate hiding in the list. The AI column is set up but not run: pick your model and try it on 5 rows. Delete the table when you're done; it goes to the trash like any other.

There's one account and no email, so there's no reset link. If you forget the password, run cubex reset-password on the computer Cubex runs on. Changing or resetting the password signs out every browser.

Cubex keeps running in the background and starts again after a restart: on macOS and Windows when you log in, on Linux when the computer starts. If the installer can't set that up, it tells you the one command that does. The cubex command does the rest:

Command	What it does
cubex status	Whether Cubex is running, its address and version
cubex open	Open Cubex in your browser
cubex stop, cubex start, cubex restart	Stop, start or restart it
cubex logs	Follow the server log
cubex update	Update to the latest release; your data and settings are kept
cubex rollback	Go back to the version and the data from before the last update
cubex config	Show the settings file, config.env in the Cubex folder (every setting)
cubex reset-password	Choose a new password
cubex uninstall	Remove Cubex but keep your data (--delete-data removes that too)

How it works

Import

Create a table, then import a CSV or start typing.

AI column

Click AI Column, name it, and write what you want. Type / to use another column's value. Try it on 5 rows, check the results, then run it on the rest.

• Starters. Not sure what to write? Start from a ready prompt: qualify (YES or NO), classify, write an opener, a lead score from 1 to 10, or a company profile. Each is written with your sheet's own columns; fill in the parts in square brackets.
• Web search. Turn on web search, or "fetch URLs from referenced columns", when the model needs current information. Pick the search engine and limit the searches per row, since search fees are usually most of the cost. Each row's "(Data)" cell shows what it searched for and what it cost.
• Control. Cells fill in live while a run works. Pause, resume or stop it. The column's menu says how many rows each run takes and shows its cost before it starts: Run N empty or out-of-date rows (never run, or the row's input cells changed since its result was made), Retry N failed rows and Force run all N rows. A column with failed rows says N failed, retry in its header: one click runs them again. Hover a cell and click ▶ to run that row alone; hovering ▶ shows what it costs. Stop all in the header stops every run and setup on the sheet at once. Changed the prompt? Run First Rows… in the column's menu runs it again on the sheet's top rows (10 by default), so you can check them before running the rest. A run that went wrong? Undo last run… in the column's menu puts back what its cells held before it, for 7 days: cells changed since stay as they are, and nothing runs again or is billed. HTTP API, enrichment and waterfall columns have the same controls.
• Only run if. Run only the rows that pass a rule, such as {{Industry}} is "SaaS" AND {{Employees}} > 50. Say in plain words which rows should run and your default AI model writes the rule (one small call), or write it yourself. Cubex checks the rule without AI, so the rows it leaves out cost nothing and show "⏭️ Condition not met". HTTP API and enrichment columns have it too, so one column can decide which rows the next one runs on.
• Several columns at once. One run can fill several typed columns, such as a score and the reason for it: Fill several columns from one answer in the dialog (the company profile starter sets six up), or output_columns through the API or MCP.

HTTP API column

Click the arrow next to AI Column, then HTTP API, and describe the request: method, URL, headers and body. {{column}} (or /column) inserts the row's value. Preview a few rows, then click the response fields you want as columns.

• Or click Paste a cURL command and paste one from the API's docs, or from your browser (right-click a request in DevTools, then Copy as cURL). Cubex fills in the method, URL, query, headers and body. It looks for keys in the command (sign-in headers and -u, headers and fields named like keys, and well-known key formats) and never keeps one it finds in the column: it asks to save each as a saved key, or to use one you already saved, and puts {{name}} in its place.

• Save API keys in Settings → Enrichment and use them by name, for example Authorization: Bearer {{apollo_key}}. They're stored encrypted and never shown again.

• Responses must be JSON. HTML pages are rejected, so this isn't a scraper.

• Requests can't reach private or internal addresses (localhost, your LAN, cloud metadata), so a value that came in through a webhook or an AI answer can't point Cubex at your own network.

• A run makes 5 requests at a time by default, up to 100 (configurable).

• A request that fails because the API is busy (429), down (5xx) or doesn't answer is tried again, up to 3 times by default (0 to 5), each time a little later, following the wait the API asks for, up to a minute. POST requests are retried only on 429 and 503 unless you turn on Also retry POST requests. A request the API rejects, like a 404, isn't retried.

• Leave out empty values skips a query parameter or body field whose {{column}} is empty for that row, instead of sending it empty. Cell values always go in the way their place needs: escaped inside a JSON body, and in a URL the characters that would change the query (&, #, +) are encoded. Saved keys go in exactly as saved.

• Add a column with the response code keeps each row's HTTP status, like 200 or 404, failed rows included.

• Turn on auto-run from the column's menu and it calls the API for new rows by itself, such as rows a webhook adds, once every column the request uses is filled. It runs at most 500 rows a day unless you change that, since each row sends a real request.

Enrichment

Click the arrow next to AI Column, then Enrichment, to look things up for each row through treg.to with your own treg key. treg can ask one provider after another until one finds the answer, and bills your own treg balance.

• Find verified work emails, phone numbers, and person and company data. Lists, such as the people at a company, become rows in another sheet.
• Pick a job, match its inputs to your columns, run it on 5 rows, then on the rest.
• You see the estimated cost before each run. Every run stops at the per-row and per-run limits you set, and a daily limit ($10 by default) covers all enrichments.
• Turn on auto-run and the column fills new rows by itself, for example rows a webhook adds.

The enrichments guide has the details.

Waterfall

Click the arrow next to AI Column, then Waterfall, to fill one value per row, such as a work email or a phone number, by trying steps in order until one finds it. A step can be a treg tool, an AI prompt, a call to your own API, a free formula or a free lookup in another sheet, so the cheapest source goes first and the next one runs only on the rows it missed.

• Start from a template (a work email from a name and a domain, or a mobile or direct phone), or pick what it finds (work email, phone, company domain, LinkedIn URL, URL or any text), then match its inputs to your columns once. Every step reads those same inputs.
• Add the steps and put them in order. A step runs only on rows that have what it needs, say a name and a domain, or a LinkedIn URL; other rows skip it for free. Switch a step off to skip it.
• An answer counts only when it reads as what you asked for: "N/A" or a broken email isn't one, so the next step runs.
• A lookup step takes the answer from a list you already have, such as a CRM export in another sheet: the first row whose domain (or any column) matches. Put it first, and only the rows it can't answer reach the paid steps.
• A step can check what it found before it counts: verify a work email with treg, or test it with a formula. A value that fails is rejected and the next step runs.
• Try on 3 rows runs it for real on a few rows and writes nothing to the sheet: you see what each step did for each row. Test this step does the same for one step. For 7 days the real run reuses what that bought, at no cost.
• Before you run, you see what a row costs on average and at most, and how far down the steps rows get. Run it on 5 rows first, then on the rest, within a run limit you set.
• A column says which step answered each row, and a click on a result shows every step that row tried and what it cost.

Agents use waterfall_column; the MCP guide has the details.

AI visibility

On the dashboard or in New Table, Check AI visibility makes a table that asks AI engines the questions your buyers ask, and shows whether the answers name you. It runs on your treg key.

• Give your brand, the competitors to look for, your site, and the questions as a buyer would ask them ("What's the best CRM for a small team?"). Pick the engines: ChatGPT, Gemini, Google AI Mode, Perplexity and Copilot.
• Making the table asks nothing. Its guide shows what asking costs (a fraction of a cent per engine and question) and asks when you say so.
• You in ChatGPT, and the same column for each engine, says where the answer first names you among the brands you track ("1st of 4"), or Not named. Named by counts the engines that name you, and Named first is the brand most engines put first. The view Not named anywhere lists the questions to work on.
• The Cited sites sheet holds every source the answers cite, with its domain and whether it's your site. The Domain column's value counts show the sites AI cites most: the pages worth getting onto.
• Add questions as rows and ask again: only the new ones are asked. Change the brands you track at any time: the columns read the answers again, at no cost.
• Weeks later, Check again in the guide keeps the answers there in sheets named with the day they were asked ("Questions 2026-10-01", "Cited sites 2026-10-01"), adds Change since ("Up 2", "Down 1", "Same") and that day's Named by to the questions, then asks every question again after showing what it costs. Checking within three weeks mostly shows how answers vary between asks, and it says so. A check that couldn't ask every engine (the daily limit, say) is picked up by the next one: it keeps the same sheets and asks only the engines it missed.
• Each engine column is an ordinary enrichment column: its menu runs it again, and auto-run can ask new rows as they arrive. Answers vary from one ask to the next, so read the totals rather than one question.

Agents use ai_visibility; the MCP guide has the details.

Clean data

Click + Add Column, choose Clean data, and pick the column to clean and how. It's free: nothing is sent anywhere and no AI runs. You see the first rows before and after, and the cleaned values go into a new column, so the original stays as it was.

• Names. Fix words typed in ALL CAPS or lowercase ("JOHN SMITH" and "John SMITH" both become "John Smith"), turn "Smith, John" around, drop titles (Mr, Dr), suffixes (Jr, PhD) and nicknames, or keep just the first or last name. Words typed in mixed case, like DeShawn or McDonald, keep their capitals.
• Company names. Drop legal endings like Inc, LLC, Corp, Ltd and GmbH, and fix ALL CAPS: "ACME CO., LTD." becomes "Acme".
• Emails. Lowercase them and leave a cell blank when the address isn't valid. This checks the format only, not whether mail gets delivered.
• Domains. Get the root domain from a website or an email: "https://blog.acme.co.uk/about" and "pat@acme.co.uk" both become "acme.co.uk".
• Phone numbers. One format for every number: (212) 555-0123, +1 212 555 0123 or +12125550123. Cells that don't hold a real number are left blank.
• Text. Title Case, lowercase, or remove extra spaces and invisible characters.

Auto-run is on by default, so new rows and rows whose source changes are cleaned on their own. Any column's menu has Clean data… too.

Formulas

Click + Add Column and choose Formula to work a new column out from each row. A formula is one line of JavaScript that reads the row's cells as {{Column name}}:

• {{Response}}?.emails?.find(e => e.type == "work")?.value picks the work email out of an API response.
• `${{{First}}} ${{{Last}}}` joins two columns.
• Number({{Price}}) * 1.2 does math. Cells are text, so turn them into numbers first.

You don't need anything to use them: formulas are free and run inside Cubex, with no AI and no requests. You see what the formula gives on the first rows before you save it.

Not sure how to write one? Describe what you want, like "the work email from Response", and click Write it with AI. Your default model writes the formula once, for a fraction of a cent, and you check it in the preview. Every row after that costs nothing.

A cell holding JSON, like a whole API response, reaches the formula as an object or a list. That's how you pick an item by a condition, which an HTTP API column's JSONPath fields don't do: extract $ (the whole response) into a column, then add a formula column that reads it. A row the formula can't work out shows ❌ and the error.

Auto-run is on by default, so new rows, and rows where a column the formula reads changes, are worked out again on their own. Formulas run in a sandbox with no access to your files, the network or the server, and with time and memory limits.

Lookups

Click + Add Column and choose Lookup to fill a new column from another sheet, in any table. Pick the sheet to look in, the column there to match, and the column of this row that holds the value: for each row, Cubex finds the rows there that match and writes

• a column of the first match (top of that sheet first), like the industry of a company found by its domain,
• Yes or No, like whether a lead is already in your customer list,
• how many rows match, or
• the matching rows themselves (up to 100), as a JSON list a formula column can read.

A match is is exactly (the whole value) or contains (the value anywhere inside the other cell, 3 characters or more). Capital letters (accented ones too) and spaces around the value don't matter. It's free: no AI, and nothing is sent anywhere. You see the first matches before you save it.

Changes in the other sheet don't update the column; click Look up every row again in its menu when they should. A lookup waits while a run is filling a column it reads there. A paused run doesn't hold it: the rows that run hasn't reached read as empty, and the lookup says so while the run is paused; look every row up again once it's done. Auto-run is on by default, so new rows and rows whose value changes are looked up on their own. Renaming columns of the other sheet keeps the lookup working; if that sheet, or a column the lookup reads, is deleted, the column stops and its header says why.

Lookups stay fast on big sheets: Cubex keeps an index of each column other sheets look up, built in the background the first time, so your sheet stays usable meanwhile. Contains on very large sheets can take a few minutes.

Remove duplicates

Open Data in the toolbar, choose Remove duplicates and pick the columns rows must match on. Upper/lower case and extra spaces never matter, and each column can match as:

• Text. "Acme Corp" matches "ACME corp".
• Email. "Jo@Acme.com" matches "jo@acme.com".
• Website or domain. "https://www.acme.com/about", "jo@acme.com" and "acme.com" all match. A Gmail or Outlook address never matches by domain, and a page on a site many businesses share, like facebook.com/joespizza, only matches the same page.
• Phone. "(212) 555-0123" matches "+1 212 555 0123". Numbers with different extensions stay apart.
• Company name. "Acme, Inc." matches "ACME LLC".
• Person name. "Smith, John" matches "john smith", while "John Smith Jr." and "John Smith Sr." stay apart.

Then choose which row of each group stays: the first, the last (usually the newest), or the most filled-in. Before anything is removed you see how many rows would go and a few of the groups, including the biggest. It's free, and it checks the whole sheet: filters don't apply.

A row missing a value in any column you picked always stays, and so does one where that cell holds a run's status like "⏭️ Not found": 5,000 rows without an email aren't one person. Removed rows go to the trash, so you can bring them back. While a run is going on the sheet, removing waits until the run finishes or you stop it.

Auto-dedupe

Keep one row per key as new rows come in, for free. In Remove duplicates, open As rows arrive, pick the key column and how it matches (the same ways as above: an email, a domain, a phone number…), and which row stays:

• Oldest row. The row already in the sheet stays; a new row with the same key goes to the trash. Good for a webhook that sends the same lead twice.
• Newest row. A new row replaces the older rows with its key, and their cells go to the trash, enriched ones too.

A row is checked when it arrives (a webhook event, a CSV import, an agent adding rows) and when someone types a key into it, so adding a row in the grid and typing an email that's already there sends it to the trash, with Restore in the message. Blank cells and cells over 200 characters never match, and values that AI, HTTP or other filled columns write are never checked: the key is a column rows arrive with (to match a cleaned value, pick the raw column and the match, like a website column as a domain). Rows already in the sheet stay as they are; the dialog says how many share a key, and Remove now cleans them up once.

What auto-dedupe removed is in the trash, one batch a day: the dialog's See them in the trash shows just those. A row a run is working on is never removed: its check waits for the run. Turning it on, changing it or turning it off waits for runs on the sheet to finish. On a big sheet it takes a moment to read the keys already there; rows that arrive meanwhile are checked once it's done.

Filters

Click Filter in the bar above the grid to choose which rows show. Each rule picks a column, a condition and a value: text (is, contains, starts with, is any of), numbers (greater than, less than), dates (is, is before, is on or after), empty or not, and on a run's column its state (has an error, not run yet). Join rules with and or or, and put some in a group to mix the two: Signed is on or after 2024-01-01 and (Plan is Pro or Plan is Team). Every rule has a switch: turn it off to see the rows without it, then back on. Changes apply as you make them, and the view keeps them. A column's own filter shows in its menu too, where you can change it.

A date can be written in a cell as 2024-01-31, 1/31/2024, 31.1.2024, Jan 31, 2024 or 31 Jan 2024, with or without a time. With slashes the month comes first, unless the first number is over 12; with dots the day comes first. A cell that isn't a date never matches a date rule, so use Is empty to find the blank ones.

Views

A view saves what you're looking at: its filters and which columns are hidden. Every sheet starts with a Default view. Use the bar above the grid to switch views, to add one (it starts as a copy of the one showing), or to rename, describe or delete one. Hide a column from its menu (Hide column) or from the 12/15 columns picker; hiding never deletes anything, and every view shares the same rows and column order.

Two ready-made views are a click away in the view menu: Rows with errors (any AI, HTTP API or enrichment column has an error) and Fully filled rows (every one of those columns has a result). Export then asks whether you want this view (only its rows and visible columns) or the whole sheet. Agents can read through a view too: read_rows, export_csv and the REST API take its name.

Search

Press / (or Cmd/Ctrl+F) on the grid, or click Search in the bar above it, and type: the grid jumps to the first cell that contains your text (case doesn't matter) and highlights it. Enter goes to the next match and Shift+Enter to the previous one, round and round, and the box shows where you are ("3 of 1,240"; past 10,000 it says "10,000+"). Search goes through the whole sheet, not only the rows on screen, and follows the view: its filters pick the rows and its hidden columns are skipped. A run's own words (⏳, ❌, ⏭️) never match.

The Is any of filter lists a column's most common values with how many rows hold each. Narrow the list by typing, tick the ones you want, or type a value that isn't listed and press Enter.

Setups

When a chain of columns works on one sheet (say a formula pulls the domain, an HTTP API column looks the company up, and an AI column scores the fit), save it: open a filled column's menu and choose Save as setup…, or use Setups in the toolbar. The columns it reads come along, and you see the order the steps run in and which columns a sheet needs for it. Only how each column is filled is saved, never your rows, and saved API keys are kept by name only.

On another sheet, open ⋯ in the toolbar, then Use a setup…. Match each column the setup reads to a column of this sheet, rename any column the sheet already has, and see what it would cost before you start: AI steps priced on your rows, enrichments at treg's prices, clean and formula steps free. Test it on the first few rows, or run it on every row. The steps run in order, each once the columns it reads are filled, and a chip in the toolbar shows where it is, with pause and stop. After a run on every row, the auto-runs saved with the setup are turned on.

New rows on a sheet you used a setup on? Use it again there. The steps whose columns the sheet has already, set up the same way, show Carries on: each fills only the rows it hasn't done, with the column's own model and settings, and you see what that costs first. AI columns fill their empty rows, HTTP API columns their empty or failed rows, enrichments the rows not enriched yet, and clean and formula columns follow their columns as they always do. It's also how a setup that failed or was stopped picks up where it left off: the dialog starts from the choices you made last time on that sheet.

Setups live in Settings, Setups, where you can rename, export one as a file for another Cubex, import one, or delete it. A setup is a copy: changing or deleting it never changes a sheet you used it on.

About to change a prompt, a request or a formula? Open the sheet tab's menu and choose Setup versions… (or ⋯, This sheet's setup versions… in the toolbar), then Save version. A version keeps how every column of the sheet is set up: prompts, models, requests, enrichments, formulas, cleaners, lookups, "Only run if" rules, auto-run, descriptions and the column order, never the values in the cells. Compare and restore shows what changed since, setting by setting. Restoring puts the version's setup back and adds back columns deleted since, empty. It runs nothing and bills nothing, and it first saves the setup the sheet has now as a version of its own, so you can go back. Each column then runs or works out the restored way the next time you run it.

To keep the values as they are now, say before running a column again, choose Copy values to a new sheet in the same tab menu. The copy opens right after the sheet, with every row in the same order under the same column names, as plain columns that nothing fills: no runs, formulas or lookups come along, and cells holding a run's own words (⏳, ❌, ⏭️, ➕) are left empty. It costs nothing, a million rows take about 8 seconds, and the sheet you copy stays open to edits meanwhile.

Trash and history

Deleted rows, sheets and tables go to the trash (the trash can at the top of the dashboard), and so do rows a CSV import replaced, rows removed as duplicates, rows a list enrichment's rerun replaced and rows an agent deleted or moved to another sheet. They stay 30 days, then they're deleted; Settings, Account sets the days (0 deletes at once). Right after deleting rows in the grid, Restore in the message puts them back.

In the trash, search by table, sheet or the access token that deleted something, then restore one item or several. A sheet or table comes back as it was, with its rows, columns, runs, views and webhook. Rows go back to their old places when the sheet hasn't been sorted or replaced since, otherwise to the end of the sheet, and columns they need that the sheet lost come back too. Look inside a batch of rows to find the ones you want by any column and restore only those, or download them as CSV. A deleted column's values aren't kept.

History, in the ⋯ menu of a sheet's toolbar, lists what changed on it in the last 90 days: columns added, renamed and deleted, runs started, stopped and undone, auto-run, views, its webhook, and rows deleted, restored, imported or sorted, each with when and who: you in the app, an access token by name, or Cubex itself. Agents read the same with get_sheet_history, and use list_trash and restore_from_trash.

Webhook

Open a sheet and choose Data, then Webhook, or let it create a table of its own. Send one test event and click the fields you want as columns. Every event after that becomes a row with those fields filled in. See the webhooks guide.

Run webhooks

In Settings, Run webhooks, add a URL and Cubex sends it a signed POST whenever an AI, HTTP, enrichment, waterfall or setup run finishes, so a script or an automation can carry on without checking again and again. Agents pass a webhook's id when they start a run to hear about that run only. See the run webhooks guide.

Agents (MCP)

Create an access token in Settings → Agent access. The page shows the setup command for your agent, with your address and token filled in. For Claude Code it's one line:

claude mcp add --scope user --transport http cubex http://localhost:3002/mcp \
  --header "Authorization: Bearer cubex_pat_..."

Then ask for the work in chat:

• "Call https://api.example.com/people?email={{email}} for each row and add the job title and LinkedIn URL as columns."
• "The last run left some cells blank. Show me why, then re-run only those rows."
• "Export the rows where Status is Qualified, with just Email and Company."

Give each token only the scopes it needs, so an agent can be read-only:

• read: list tables, read rows and check runs.
• write: also create and change tables, sheets, columns and rows.
• run: also start, pause and re-run runs, which spend your OpenRouter credits and treg balance.
• secrets: also start HTTP runs that use your saved API keys.

Give a token a spend cap to limit what an agent can spend on AI and enrichments in all (for the token's whole life, not per month). At the cap its new runs are refused and running ones pause until you raise it. The token list shows what each one spent.

Setup for Cursor, Codex, VS Code and Claude Desktop, the REST API and the full list of tools are in the agents and API guide.

The Cubex plugin for Claude Code, Codex and Cursor adds the connection (Claude Code and Codex read your address and token from environment variables, never from a config file) and skills for the common jobs: onboarding, cleaning and deduping a list, fixing failed rows, enriching with cost control, tracing where a value came from and checking limits. Paste this into your agent:

Set up Cubex by following https://raw.githubusercontent.com/gurmohitghuman/cubex/main/docs/agent-plugin.md

See the plugin guide.

Guides

Guide	What's in it
Run Cubex on your own server	From a new Linux server to Cubex on your domain: HTTPS, nightly backups, updates, keeping it safe
Other ways to run Cubex	Coolify, Docker, and running from source
Settings and your data	Every setting, where your data lives, backups, and what Cubex connects to
Performance	Measured times and memory on a sheet with a million rows
Enrichments	treg keys, jobs, costs and limits, auto-run, stored answers
Webhooks	Setting one up, reaching it from outside, what it accepts, keeping it safe
Run webhooks	A signed notice when a run finishes: setting one up, checking the signature, delivery and retries
Agents and API	MCP setup for each agent, every tool, the REST API
Agent plugin	The Cubex plugin for Claude Code, Codex and Cursor: setup from one prompt, and its skills
Contributing	Development setup, tests, and the contributor agreement

License

Cubex is open source under the GNU Affero General Public License v3.0 (AGPL-3.0). Copyright (C) 2026 Gurmohit Ghuman.

• You can use Cubex for free, host it yourself, use it in your business and run it for your clients.
• You can change it. If you let other people use a changed version, including over the web, you have to give them the source code of your version under the same license.
• The license covers the code, not the Cubex name or logo. If you share a changed version, give it its own name so people don't mistake it for Cubex.

Other names and logos in this README belong to their owners: Claude and Claude Code to Anthropic, Cursor to Anysphere, Codex and the OpenAI logo to OpenAI, and OpenRouter to OpenRouter, Inc. The same goes for the Model Context Protocol and treg.to logos. They're shown only to say which tools Cubex works with. Cubex isn't affiliated with or endorsed by any of them.