Models

Forge has no model of its own and no server in the middle. You supply a credential, and the CLI talks to that endpoint from your machine. Nothing about a task leaves your shell except the request to the provider you configured.

Providers

Both sit behind one provider interface, which is where additional providers will be added. Only these two exist today; if you need a different vendor now, put a gateway in front of Forge rather than waiting for a native provider.

Choosing a model

There is no default model - Forge will not guess what you want to be billed for. Set one in your shell profile, or pass --model for a single run.

$ export ANTHROPIC_MODEL="claude-sonnet-4-5"$ forge "add a retry to the webhook client" # or, just for this run$ forge --model claude-opus-4-5 "refactor the scheduler"

Model ids are passed through untouched, so a new model works the day it ships without upgrading Forge. Starting a session with no model configured exits with code 2 rather than falling back to something arbitrary.

Which one to use

• A mid-tier model is the right default. Most of the loop is reading files and running commands, and it handles that at a fraction of the cost.
• Reach for the strongest model on tasks that are genuinely hard to reason about - a subtle bug, a refactor that spans packages, anything where a wrong plan wastes many iterations.
• Smaller models tend to fail in a specific way: they stop early and declare success. If a model keeps reporting work it did not verify, that is the signal to move up.

Credentials

Either an API key or a bearer token works. Forge reads the FORGE_ variables first, then the ANTHROPIC_ equivalents, so an existing setup needs nothing new.

Pointing at a gateway

Because the base URL is configurable, anything that speaks the Anthropic API works - an internal proxy, a billing gateway, a local router, or a logging layer your team already runs.

$ export ANTHROPIC_BASE_URL="https://gateway.internal/v1"$ export ANTHROPIC_AUTH_TOKEN="$GATEWAY_TOKEN"$ export ANTHROPIC_MODEL="claude-sonnet-4-5"$ forge config

Forge does not care what is on the other end. If the gateway rewrites model names, use the names it expects.

Token use and limits

• --max-tokens caps output tokens per model call. It bounds a single reply, not the run.
• --max-iterations bounds the run. This is the control that decides what a runaway task can cost you.
• Long sessions compact their own context rather than failing at the window limit, so early history is summarised as a run goes on.
• Every run prints duration, model calls, tool calls, and token usage when it finishes. Read it a few times and you will know what a task costs before you start one.