Pi coding agent on one key

• Pi installed: curl -fsSL https://pi.dev/install.sh | sh, or npm install -g --ignore-scripts @earendil-works/pi-coding-agent on Node.js 22.19 or newer.
• A RelaySwitch key from the console (email-code sign-in). Keys start with sk-relay-.
• A prepaid balance, from $10 by card or crypto, no subscription and no KYC. The console chat runs 12 free turns a day before the first top-up; Pi requests bill the balance.

Pi keeps its user configuration in ~/.pi/agent (PI_CODING_AGENT_DIR moves it). Custom providers go in models.json there:

• Any key under providers that is not a built-in id adds a provider. It needs baseUrl and an api; openai-completions calls /chat/completions under https://api.relayswitch.io/v1, which every model on this page answers.
• apiKey as "$NAME" reads an environment variable. A literal key works, and a value that begins with an exclamation point runs a command and uses its output, for a secret manager.
• A models entry needs only id. cost is optional, but once set it needs all four fields in USD per million tokens; this block sets them so Pi's session cost uses the catalog prices. cacheRead is the catalog's cached-input rate; cacheWrite is set to the input rate, since the catalog bills no separate write price. name is what /model shows.
• contextWindow and maxTokens default to 128,000 and 16,384 when omitted; the block fills them from each model's catalog specs. input adds image where the model takes one, so Pi can attach screenshots.

Field reference: Pi's models-configuration.md; the compatible-endpoint walkthrough is models.md (both read 2026-10-28). Add any other id from the Models page the same way; an entry whose id matches an existing model on that provider replaces it, and modelOverrides changes a built-in's metadata without a new list.

1. Save the block above as ~/.pi/agent/models.json.
2. Export the key in the shell that runs Pi, then start it:

3. Send a prompt. The request shows on the console's Usage page within seconds, with its tokens and its charge.

Add the export to your shell profile so new terminals have it. --model takes provider/id, an exact id or a fuzzy match on id or name; --provider relayswitch narrows the lookup to this block.

Inside Pi, /model lists every model whose provider has usable credentials. Ctrl+S on a model saves it as the default for new sessions and Ctrl+P cycles through the available ones. The same default as a file:

That file is ~/.pi/agent/settings.json. enabledModels in it, or --models on the command line, limits startup selection and cycling to a list of exact ids, fuzzy matches or globs (an array in the file, comma-separated on the command line): pi --models claude-sonnet-5-5,claude-opus-5-5 keeps the cycle to those two.

Any chat id on the Models page. The ids in the block are the usual coding set: a Claude flagship for planning and hard edits, GPT-5.5 and Gemini 3.1 Pro as second opinions, DeepSeek V4 Pro and GLM 5.3 where the input price decides. The table is what they cost today.

Input decides the cost of an agent session: the conversation and every tool result travel up again on each step. Cached input bills at the cached rate by itself; there is no switch to set.

A turn here is 80,000 tokens of earlier context, 2,000 fresh input tokens and 1,500 output tokens. The first column bills the earlier context at the cached rate, the second as if none of it were cached.

With cache hits the task runs $0.2964 on GLM 5.3 and $1.26 on Claude Opus 5.5. Run routine edits on the low end and keep the flagship for reviews, and the balance lasts longer.

Confirm the key from a shell before you open Pi:

• 401: the key is wrong or revoked, or RELAYSWITCH_API_KEY was not set in the shell that started pi. Export it and start Pi again, or confirm the value with the curl above.
• 402: the balance is too low. Top up in the console; the key works again at once.
• 404: the id in the models array is not in the catalog. Copy it from the Models page rather than typing it: a dot, a dash or a date suffix out of place is enough.
• 429: the key hit its rate limit. Wait and resend, or create a key with a higher Requests / min in the console.
• The model is missing from /model: Pi hides a custom model until it can resolve the provider's credentials. Check the variable, then open /model again; opening it reloads models.json, so a new id needs no restart either.

The full error table is in the docs.