Free AI for documents: a folder-to-Word-and-Excel example

Create a checked Word and Excel report from eight fictional tasks. Download inputs, prompts and deterministic reference files, then repeat the task with a model available to you.

The More AI team · Updated October 7, 2026

Working on files, not on copy-paste

In Home, choose a project folder and ask the agent to read its documents, propose edits or create Word, Excel and PowerPoint files. Start with a working copy, review the selected permissions and inspect the resulting files in an office application.

Source files remain in the folder you selected, while an online model sees the text you send it. Signed-out work does not sync; after sign-in, project sync is on by default unless disabled. End-to-end encryption for project files is optional and starts off, so synced file content is readable by the service and web app by default.

A project folder open in Home, with file changes shown as diffs

Choose a model available to your account

Google AI Studio and Groq offer model-dependent free API access where available. Neither is universally the fastest or most generous: rate limits, region, context and workload determine what works. Check the current model and account limits, then try a small task.

  1. Install More AI and open the desktop app; an account is optional for this local provider-key route.
  2. Choose a provider and model available to your account and region, then add its API key in Settings → Providers.
  3. Copy the sample documents into a working folder and open it as a project in Home.
  4. Define a bounded task and its checks, such as a project summary with source task IDs.
  5. Review the selected file and command permissions, inspect visible diffs and open every output file before accepting the result.

What it actually does with your files

The agent can write and run a generator script to create a .docx, .xlsx or .pptx when the required tools, permissions and dependencies are available. Open the output and check figures, formulas and formatting; a generated filename or completion message does not establish correctness.

File operations and commands follow the selected tools, access mode and approvals. Use a working copy and limit the available tools to the task. Review proposed commands, visible diffs and generated files; a diff is not a backup, and permissions do not guarantee that an unwanted edit or deletion is impossible.

  • Summarise or compare a folder full of documents into one table
  • Rewrite, reformat and fix numbering across a long report
  • Build .docx, .xlsx and .pptx files from data you point it at
  • Extract figures from several files and check them against each other

Your conversations stay yours

Chats and projects live in local storage on your computer, so they survive closing the app. Signed-out work is not synchronized. After sign-in, supported conversation and project categories synchronize by default unless you disable their individual switches.

Synchronization makes supported work available in a browser or on another device; it is separate from cloud execution. End-to-end encryption for project files is optional and off by default, so synced project file content is readable by the service and web app unless you enable E2E.

More AI sync category and separate cloud-execution settings

Where free runs out

Free APIs can have per-minute, token, daily or account limits. A document task can make several model and tool calls; its cost depends on context, output, retries and the provider. Check current limits before starting and record failures instead of treating an incomplete result as finished.

Local inference has no provider per-token bill. Hardware, electricity, storage and your time still have a cost. Capability depends on the exact model, quantization, context length and available memory; test one small task before choosing it for a workflow.

A small project report you can reproduce

Download a workspace containing eight fictional tasks, project notes and prompts in English, Russian and Chinese. The files contain no customer data. The goal is a real workbook and a short Word summary, with the original task IDs preserved.

The reference XLSX and DOCX below were produced by a deterministic generator and mechanically checked. They are not outputs from a More AI test and do not demonstrate a model’s accuracy. Use them to check facts; your layout can differ.

Repeat the task in Home

Use a provider/model available to your account and region. A provider free tier can be enough for this small task, but quotas and tool support vary. Local inference requires a suitable local model and prepared dependencies; neither route guarantees a successful run.

  1. Unzip the sample into a new folder and open that folder as a project in Home. Keep the original input files.
  2. Select a model and review folder, command and network permissions. Start with confirmation for actions.
  3. Use PROMPT-en.txt: create output/project-report.xlsx and output/summary.docx; compare planned and actual hours only for done tasks; keep the missing actual value of T03 unknown.
  4. Ask the agent to explain its checks and errors. Review any dependency installation before approving it.
  5. Open the produced files in an office application and compare their values with the reference below. A success message alone is not validation.

The checks that decide whether it is correct

There are 8 source tasks: 7 done and 1 pending. Completed work has 30 planned hours and 33 actual hours, a +3-hour difference. The plan for all tasks is 32 hours. T03 is pending with 2 planned hours and no known actual value. Comparing 32 planned hours with 33 completed actual hours would mix different scopes.

The completed-work subtotals are Aurora 10 planned / 12 actual, Boreal 10 / 10 and Cedar 10 / 11. Keep the task IDs in the workbook and summary so every figure can be traced to the CSV. The reference workbook includes formulas and preserves the missing T03 value.

  • All eight IDs appear once; no source row is silently dropped.
  • T03 stays pending and its unknown actual time is not converted to zero.
  • Formulas calculate completed-work totals and a separate full-plan total.
  • The summary identifies source task IDs, the pending item and the +3-hour variance.
  • Files open successfully, and any errors or manual corrections are recorded.

An observed run, including the repair

On 7 October 2026, we tried this English-language task in More AI 0.2.0 with Z.ai (GLM), glm-5.3-flash selected. File execution was on the computer; the inference endpoint and provider bill were not independently verified. The download contains the English files and the run protocol.

The first XLSX reversed the sign of per-task variances. A follow-up repaired those formulas. We then changed only T05 actual hours from 6 to 7: completed work became 30 planned / 34 actual / +4 hours, with Boreal at 10 / 11 / +1. Separate LibreOffice recalculation checks confirmed these figures, all eight task IDs and the unknown T03 value; the Word summary matched.

This result required a repair and follow-up checks. The original XLSX needs formula recalculation, so cache-only previews can show blanks. Review print layout before sharing the files. Provider cost remains unverified.

Frequently asked questions

Can it really create Word and Excel files?

The agent can write and run a generator script to create a .docx, .xlsx or .pptx when the required tools, permissions and dependencies are available. Open the output and check figures, formulas and formatting; a generated filename or completion message does not establish correctness.

Do my documents get uploaded anywhere?

The source files stay in your folder, while an online model sees the text you send it. Signed-out work is not synchronized; after sign-in, project sync is on by default unless disabled. Optional E2E starts off, so synced project file content is readable by the service. With Ollama, prompts and inference stay local when sync, online tools, master telemetry and AI observability (including replay) are off.

Are my chats saved?

Yes — conversations and projects are kept in local storage on your computer and survive restarts. Signed-out work does not sync; after sign-in, supported categories sync by default and can each be disabled.

Which free model is best for documents?

There is no universally best free model for document work. Check your region, account quota, context and tool support, then repeat the small exercise below and compare verified outputs.

Can it delete my files?

File operations and commands follow the selected tools, access mode and approvals. Use a working copy and limit the available tools to the task. Review proposed commands, visible diffs and generated files; a diff is not a backup, and permissions do not guarantee that an unwanted edit or deletion is impossible.

How More AI works

From download to first result in three steps

All you need is a key for an AI model. If you don’t have one, OpenRouter gives you one in a couple of minutes. An account is optional.

Install the free app

For macOS, Windows and Linux. It keeps itself up to date.

Paste your key

OpenRouter, OpenAI, Anthropic, Google, DeepSeek and 30+ other providers — or a model running on your own computer. Pick a model for each task.

Ask for the result

Point it at a folder, a website or a repository and say what you need. Every change shows up for you to review.

All guides