Keep Claude Code, Cursor or the agent you use. CompletionPrism™ sits under it, routes each step to the model that finishes it for least across Anthropic, OpenAI, Google and xAI, and prints the receipt. One line in, one command out.
One line under Claude Code and the tools you already use, or a coding agent that picks the model per step. Anthropic, OpenAI, Google and xAI. Remove it with one command.
curl -fsSL https://completionprism.illuminis.ai/install.sh | shnpm install -g @illuminis/comprismirm https://completionprism.illuminis.ai/install.ps1 | iexThen comprism login. No API key to export.
Install, open a new terminal, and claude is measured. Requests reach the provider byte for byte and your key is forwarded, never stored. comprism cost shows what a session cost and the five cost factors: Token Spend, Rework, Switching, Hidden Thinking and Context Re Read.
comprism work reads the project, edits files, runs your tests and shows every change first. Plan first with --plan, choose one of four permission modes, undo any change, and cap spend with --max-spend.
Four modes: read_only, approve_writes (the default), auto_edit and full_auto. No mode unlocks force push, history rewrite, download and run, credential access or production deploy. It fails open, stores no provider key, and stores prompt text only if your company turns it on.
A REST API with one optimize call, POST /api/v1/dev/optimize, answers on the cheapest model that can finish the work and returns a receipt. CI mode: comprism work --yes --mode auto_edit with an exit code. Read the API reference.
No. One line in your shell profile points it at CompletionPrism™. Your agent does not know it is there.
No. It is forwarded with each request and never read or stored.
Your request goes straight to the provider. It fails open.
comprism uninstall.