batchgrid for Python

Run a prompt, or a whole pipeline, over every row of a DataFrame or spreadsheet. You get a cost estimate before anything runs, parallel calls with retries and rate-limit backoff, and runs that resume after an interruption.

import pandas as pd
import batchgrid

df = pd.read_csv("reviews.csv")
result = batchgrid.run(df, "classify the sentiment and extract keywords", max_cost=5)
result.data  # df with the new columns

Beta. This package drives the batchgrid CLI, so it needs Node.js 22 or newer. It uses a batchgrid on your PATH, or fetches the CLI with npx on first use.

Install

pip install "batchgrid[pandas,progress]"

pandas is needed for DataFrame input and progress adds a tqdm progress bar. File paths work without either.

Set the key of the provider you use, either as an environment variable (OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, OPENROUTER_API_KEY, XAI_API_KEY) or once with npx batchgrid config --set-key openai:sk-….

Plan, check, run

Planning asks a model, so the same request can produce a slightly different plan each time. Look at the plan and its cost first, then run it:

plan = batchgrid.plan(df, "classify the sentiment of each review")
print(plan)            # the steps, the columns they write, the estimated cost
plan.cost_usd          # 0.08

result = batchgrid.run(df, plan=plan)

A pipeline that must behave the same on every run should save the plan once and run that file. Running a saved plan does not ask the planner again:

plan.save("sentiment.json")

# later, in the pipeline
result = batchgrid.run("next_week.csv", plan="sentiment.json", max_cost=10)

The file is the same one batchgrid --save-plan writes and batchgrid --plan reads.

In the browser

batchgrid.ui() opens the batchgrid workspace in your browser with your data in the sheet. You can chat, check the plan and run it there, while the log prints in the notebook cell or terminal as it happens:

df2 = batchgrid.ui(df)

batchgrid is open at http://127.0.0.1:4817

❯ classify the sentiment of each review
  ✔ Reading data.csv - 1,000 rows · 1 column: review
  ✔ Asking gpt-5.6 for a plan - 1 step · 2.1k tokens
  ✔ Estimating the cost - About ~$0.08 for 1,000 rows
batchgrid: 100%|██████████| 1000/1000 [00:42<00:00, failed=0]
  ✔ 1000/1000 rows, 0 failed
  ↳ Sheet synced: 1,000 rows × 2 columns

Finished in the browser.

To come back, press Finish in the tab, close the tab, or interrupt the cell. ui() returns the sheet as the tab last showed it. For a DataFrame, that is a copy of yours with the new columns and any cells you edited; untouched columns keep their dtypes. ui() with no data starts on an empty sheet.

The workspace runs on your machine and uses the keys from batchgrid config or your environment. Your browser has to be able to reach the machine running Python, so this does not work for a notebook on a remote server.

From a terminal, the batchgrid command this package installs does the same:

batchgrid ui data.csv     # or: batchgrid --ui data.csv

Reference

batchgrid.run(data, prompt=None, *, plan=None, ...) -> Result

Argument	Meaning
data	A CSV, TSV or Excel path, a pandas DataFrame, or None when the request brings its own rows
prompt	What to do, in plain language. Pass either this or plan
plan	A Plan, its dict, or a saved JSON file
model	"gpt-5.6" or "provider:model", e.g. "anthropic:claude-sonnet-5". The default is the saved choice
rows	Run only the first N rows, a cheap way to try a plan
concurrency	The most requests running at once
output	Where to write the result. The extension picks the format
max_cost	In US dollars. The plan is priced first and nothing runs above it, or when the model has no price list
progress	Show a progress bar when tqdm is installed. Defaults to on
on_event	Called with each event the CLI reports: plan, progress, result…

Result has status (done or stopped), total, success, failed, input_tokens, output_tokens, duration_ms, output_path, error_log_path, and data. For DataFrame input, data is a copy of your DataFrame with the new columns added, with its index and dtypes kept. If a plan filters rows or drops columns, data holds the output as it was written.

Rows that still fail after their retries do not raise an error. Check result.failed, and look in result.error_log_path for the reasons.

batchgrid.plan(data, prompt, *, model=None, rows=None) -> Plan

Plans the request and prices it without running anything. Plan has title, summary, cost_usd, cost_formatted, rows, missing_secrets, the raw steps, and save(path) / Plan.load(path).

batchgrid.resume(data) -> Result

Ctrl+C (or a Jupyter interrupt) stops a run and keeps the rows that finished. Call resume with the same file or DataFrame to finish it. Only the unfinished rows are sent to the model.

try:
    result = batchgrid.run(df, plan="sentiment.json")
except KeyboardInterrupt:
    result = batchgrid.resume(df)

batchgrid.ui(data=None, *, open_browser=True, port=None, log=True) -> DataFrame | None

See In the browser. on_event receives the same events as in run(), plus ui, sheet, browser_error and closed.

Errors

Everything raises batchgrid.BatchgridError, and its code says what went wrong:

• no_api_key: no provider key is set.
• invalid_plan: the plan does not fit the data.
• no_plan: the planner asked a question instead of returning a plan. This raises NoPlanError, and its reply holds the question.
• missing_secrets: the plan needs keys that are not saved. This raises MissingSecretsError, and its names lists them.
• cost_limit: the run would go over max_cost. This raises CostLimitError, and its plan holds the priced plan.
• run_failed: the run itself failed.

Choosing the CLI

The batchgrid command pip installs hands everything to the Node CLI, so batchgrid config, batchgrid ui and the rest work the same whether batchgrid came from pip or npm. The package looks for the CLI in this order:

1. the cli=[...] argument
2. the BATCHGRID_CLI environment variable, e.g. node /path/to/cli/dist/index.js
3. the npm batchgrid on the PATH (never this package's own command)
4. npx "batchgrid@>=0.3.0 <1"

Development

pnpm --filter batchgrid build   # the tests drive the real CLI
cd python
uv venv && uv pip install -e ".[dev,progress]"
.venv/bin/python -m pytest

The tests replace only the model: a local server stands in for the OpenAI API.