Only approved or redacted sensitive material

Do not assume open source means offline. Before using sensitive code or data, confirm the selected model endpoint is local and disconnect cloud provider keys.

A safer tool to use instead

OpenCode

A tool for people who write code, installed on your own computer, that proposes file changes and commands.

First thing to tryWith a colleague who writes code, make a new folder holding a five-row invented CSV, connect a local model, and ask for a script that writes a cleaned copy.

CostThe client is free and a local model costs only your own hardware, while any cloud provider you select bills you under its own account rules.

Download it

Do this with it

  • Data

    Redact first

    Get cleaning steps for a spreadsheet, without uploading it

    20 minutes · Gives you numbered steps

    Answer in [ANSWER LANGUAGE], even if the material below is in another language.
    
    I have a spreadsheet I cannot upload. I will describe it instead. The three rows below are invented by me, not copied out of the file, and I will never paste a real row. Do not ask me for the file or for real rows.
    
    Columns and what they hold:
    [LIST THE COLUMN NAMES AND WHAT EACH CONTAINS]
    
    Three invented rows, in the same shape as the real ones:
    [PASTE THREE INVENTED ROWS]
    
    What is wrong with it:
    [DESCRIBE THE PROBLEMS: mixed date formats, duplicates, mixed digits: ١٢٣ alongside 123, trailing spaces, merged cells, several answers in one cell]
    
    What I need at the end:
    [DESCRIBE THE RESULT YOU WANT]
    
    Give me, in this order:
    A. Numbered steps, in the order I should do them, each saying what to do and what the file will look like afterwards.
    B. The formula, or the script, for each step that needs one. Say which program it is for.
    C. The assumptions you made about my data, one per line, and for each one what it will silently break if the assumption is wrong and how I would spot that.
    D. What to check after each step to know it worked, including the row count before and after.
    
    Work only from my description. If my description is not enough to write a step safely, say what you need to know instead of guessing.
    How to use it
  • Data

    Redact first

    Find the difference between receipts and a ledger

    20 minutes · Gives you numbered steps

    Answer in [ANSWER LANGUAGE], even if the material below is in another language.
    
    My receipts and my spreadsheet do not agree, and I can upload neither of them. I will describe the spreadsheet instead. Real rows are never pasted here, so do not ask me for them and do not ask me to upload a file.
    
    Columns and what each one holds:
    [LIST THE COLUMN NAMES AND WHAT EACH CONTAINS]
    
    Three invented rows in the same shape as the real ones:
    [PASTE THREE FAKE ROWS]
    
    What does not match:
    [DESCRIBE THE DIFFERENCE: the two totals differ, a receipt has no row, a row has no receipt, one payment is split across two receipts, mixed digits: ١٢٣ alongside 123, two currencies, dates written in two formats]
    
    Rules:
    1. Work only from my description and the invented rows. Never invent a figure of mine, and never assume a total.
    2. If a step cannot be done without seeing my data, say so and tell me what to look at myself.
    3. Order the steps so that the cheapest check comes first.
    
    Give me, in this order:
    A. Numbered steps, in the order I should do them, each one narrowing down where the difference is.
    B. The formula or the script for each step that needs one, and which program it is for.
    C. The assumptions you made about my spreadsheet, as a list, and for each one, what it silently breaks if it is wrong.
    D. What to check after each step, including the row count and the total before and after.
    How to use it

Plain verdict

Good for technical participants who want control over the model provider. Open source does not mean offline: a cloud provider still receives prompts, code, and tool results.

Good for

Small scripts, repeatable data cleaning, code explanation, and local-model experiments in a disposable project folder.

Avoid for

Non-technical participants, production repositories, unrestricted shell access, cloud use with sensitive files, or generated commands you cannot explain.

Data lane

Local client with two model lanes. A verified local Ollama endpoint can keep model traffic on the machine. Any cloud provider applies its own account, retention, training, and jurisdiction rules.

Ten-minute safe trial

Create a new folder with a five-row synthetic CSV. Connect a local model, turn off the network, and ask for a script that writes a new cleaned file. Review the proposed files before allowing changes.

Human check

Inspect every diff and command. Compare row counts and totals, preserve the original, and check that no credential or real identity entered logs or test data.

Safer or constraint-based alternatives

Use a reviewed spreadsheet formula for one-off work. If a cloud coding assistant is approved, keep Privacy Mode on and open only a synthetic-data project.

What happens to your data

Where this data goes

  • your computer in local mode
  • the selected provider in cloud mode

Trains on what you type?

Depends on the plan or account type.

How long it is kept

OpenCode is a local client. It can call a model running locally through Ollama, or a cloud provider. Local model traffic stays on the machine. Cloud prompts, code, and tool results follow the selected provider's terms and retention.

Account required

No account is needed.

Getting it, and using it in Iraq

Access in Iraq

Works in Iraq

The open source client has no country gate. A selected cloud provider can introduce its own account, country, and billing rules.

Language quality

No one has published a measurement for this.

  • Arabic (Modern Standard)

    Natural-language performance depends on the selected model, not OpenCode itself. Support is not a measured quality result. Test a short sample with a fluent reviewer.

  • Iraqi Arabic

    No current first-party benchmark for Iraqi Arabic was found. Test the exact dialect and task with a fluent reviewer.

  • Kurdish (Sorani)

    No current first-party quality benchmark for Sorani was found. Do not infer quality from general multilingual claims.

Sources

Last checked: 2026-09-01