Token-efficient by design
Four tools (read, write, edit, bash) and a system prompt under about 1,000 tokens, roughly a tenth of what bigger agents send. In users' tests against Claude Code, Pi reached similar results at about half the cost.
Keep Pi exactly as it is. Add CodingPass as a provider with baseUrl in models.json and run Claude Sonnet 5.5, Opus 5.5 and Kimi K2.7 Code on a monthly coding plan from one allowance.
$10/mo until you cancel.
{ "providers": { "codingpass": { "baseUrl": "https://codingpass.ai/v1", "api": "openai-completions", "apiKey": "cp-live-…", "models": [ { "id": "claude-sonnet-5-5" }, { "id": "kimi-k2.7-code" } ] } }}The coding agent OpenClaw is built on: four tools, a short system prompt, extensions for everything else. All of it runs on CodingPass.
Four tools (read, write, edit, bash) and a system prompt under about 1,000 tokens, roughly a tenth of what bigger agents send. In users' tests against Claude Code, Pi reached similar results at about half the cost.
Pi skips sub-agents and plan mode. Add them with extensions, skills and prompt templates, or ask Pi to build the feature, then share it as a package over npm or git.
OpenClaw runs on Pi. Use the same TypeScript SDK, RPC or JSON mode to build Pi into your own apps, scripts and CI.
Continue a session later, or go back to an earlier message and try another approach. Every branch stays in the session.
Every plan works with Pi. Pick the Claude flagship you want as your default; all open-weight models come with every plan and share the same monthly allowance.
Best forDaily coding, scripts, small extensions
$10/mo
$60 Sonnet allowance a month
Best forLarge refactors, long agent sessions
$30/mo
$200 Opus allowance a month
Best forParallel sessions and the hardest problems
$100/mo
$500 Fable allowance a month
Estimates from CodingPass traffic: about 100K tokens per request at roughly 5% input, 1% output and 94% cache read. A single Pi task usually makes several requests. Allowances are metered at official API prices.
Pi loads custom endpoints from models.json. Add CodingPass once and its models show up under /model.
Sign in with Google, GitHub or email and pick a plan. Then go to Dashboard → Keys → New key. Keys start with cp-live-.
Paste the block into ~/.pi/agent/models.json with your key. To keep the key out of the file, write "$CODINGPASS_API_KEY" instead and export it in your shell.
~/.pi/agent/models.jsonRun pi, type /model and choose a CodingPass model; press Ctrl+S to keep it as the default. On Pro, add claude-fable-5-2 to the models list. Get your request ID.
pi/model$ pi > /model ● codingpass/claude-sonnet-5-5 codingpass/kimi-k2.7-code > Add a --json flag to the export command. ● Read src/commands/export.ts ● Edited src/commands/export.ts (+22 −4) ● Ran npm test · 31 passing --json prints one object per row; the table output is unchanged.
One monthly allowance covers every model, with 5-hour and weekly caps. The right model for the task makes it go further.
List a few CodingPass models in models.json, then press Ctrl+P to cycle between them. Open-weight models like Kimi K2.7 Code cost less per request; on Go each of them can use up to $40 a month from the shared allowance.
+ 4 more open-weight models, up to $40 each.
When an attempt goes wrong, run /tree and pick the message before it instead of correcting Pi in the same thread. The model only receives the active branch, so the dead end isn't sent again with every request. /compact shortens a long history further.
/tree back to an earlier message, branch from there/fork turn that branch into a new session/compact summarize older historyhttps://codingpass.ai/v1 serves every model on your plan over Chat Completions, Claude included, so one provider covers them all. See every model on CodingPass, or use Magpie, a local gateway, to put every agent on one provider.models.json. Models you connected with /login stay in the /model picker next to it.Sonnet 5.5 and open-weight models from $10 a month. One config file, no surprises, cancel anytime.