Skip to content
Documentation

FAQ

Honest answers, including about what is unfinished.

The product

How is this different from a chatbot?

A chatbot answers. Forge acts, then checks. It reads your files, runs your tests, reads the exit code, and iterates on what came back - the loop is the product. You are not copying code out of a window; you are reviewing a diff in your repository.

Is it ready for real work?

It is version 0.1.0 and alpha. The loop, tools, permission policy, git integration, and Docker sandbox all work and are used. What you should not expect is stability guarantees, a published wheel, or MCP support.

Use it on a committed tree, on tasks you can verify. That is not a disclaimer - it is how to get value out of any coding agent today.

Does it work on Windows?

Under WSL2, yes. Native Windows is not supported, and the site does not claim it is. macOS and Linux are the tested platforms.

Do I need to be online?

Yes, for the model call. Everything else is local - there is no Forge server, no account, and no telemetry. Your code goes to the provider you configured and nowhere else.

Cost and models

What does it cost?

Forge is MIT licensed and free. Model usage is billed by whoever issued your credential, at their rates - Forge does not resell tokens and takes no cut.

How much does one task cost?

It depends on the model and the number of iterations, so any figure here would be invented. What is real: every run prints its token usage, and --max-iterations is the control that bounds the worst case. Run a few small tasks and you will know your own numbers.

Can I use a model that is not Claude?

Through a gateway that speaks the Anthropic API, yes - set ANTHROPIC_BASE_URL. Native providers for other vendors are on the roadmap and not built. See Models.

Which model should I use?

A mid-tier model for most work; the strongest one for tasks that are genuinely hard to reason about. Smaller models fail in a recognisable way - they declare success without running anything.

Control and safety

Can it break my repository?

It can change your files - that is what you installed it for. What it cannot do is push, sudo, or run anything on the deny list, in any mode. Destructive calls stop for you by default.

The honest answer: commit before you run it. Then git diff shows everything it did and git checkout . undoes all of it. See Git.

Will it push to my remote?

No. git push is denied outright, including under --yolo, with no setting to enable it. Forge commits; publishing is yours.

Is it secure?

Not a question anyone should answer with “yes”. Forge gives you a deny list that cannot be overridden, risk-based approvals, workspace-confined file access, and an optional container for shell commands. Those are real, verifiable mechanisms.

What they are not is a guarantee. You are running a model that can execute code on your machine. The sandbox inherits Docker's boundary and its weaknesses. Decide with that in mind - Permissions and Sandbox describe exactly what each layer does and does not do.

Can I stop it mid-run?

Ctrl-C. It exits with 130, leaving any edits already made in place - interrupting stops the agent, it does not roll back.

Using it

What makes a good task?

One with a finish line the agent can check itself. “Make pytest tests/test_auth.py pass” converges; “improve the auth code” does not. Naming the verification command is the single highest-leverage thing you can do.

Can I run it in CI?

Yes - that is what one-shot mode and the exit codes are for. Pass the task as an argument, add --json, and check the code. Pair --yolo with --sandbox docker so nothing hangs waiting for an approval that will never come.

Does it work on a large repository?

It explores rather than reading everything, so size is less of an issue than structure. Long runs compact their context, which means detail from early iterations is summarised. Pointing it at the relevant directory with -C helps more than any setting.

Can I use it with an existing agent config?

There is no import for another tool's configuration. Forge reads its own forge.toml plus your existing ANTHROPIC_* environment variables - if those are already set, there is nothing to configure.

The project

Is this a company?

No. It is an open-source project by Tushar Umbarkar, under the MIT licence. There is no hosted product, no pricing, and no billing - and this site does not pretend otherwise.

Why build it when Claude Code exists?

Because a harness is worth understanding from the inside. Forge is small enough to read in an afternoon and honest about its boundaries, which makes it useful both as a tool and as a reference for how one is built.

Can I contribute?

Yes. Issues and pull requests are welcome at github.com/codesensei-tushar/forge. The roadmap is the honest version of what is next, and Architecture is the fastest way into the codebase.

A question this page does not answer is worth asking in an issue. If anything here disagrees with the CLI, the CLI is right - file that as a bug.