pi-cadence
Pi extension that spreads free-tier model requests across relays so the per-IP rate limit adds up
Package details
Install pi-cadence from npm and Pi will load the resources declared by the package manifest.
$ pi install npm:pi-cadence- Package
pi-cadence- Version
0.1.7- Published
- Nov 16, 2026
- Downloads
- 1,327/mo · 1,327/wk
- Author
- fn-cafeina
- License
- MIT
- Types
- extension
- Size
- 44.2 KB
- Dependencies
- 1 dependency · 2 peers
Pi manifest JSON
{
"extensions": [
"./index.ts"
]
}Security note
Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.
README
pi-cadence
Free-tier models spread across relays so the per-IP rate limit adds up.
The default provider allows 2 requests per minute per IP and per model. Your own connection is one lane; every relay adds another. Ten relays plus your own connection is 11 lanes, 22 rpm per model. Fewer relays means fewer lanes; it also works with none (2 rpm on your own connection).
Install
pi install npm:pi-cadence # from npm (recommended)
pi install git:github.com/fn-cafeina/pi-cadence # from git
pi install ./pi-cadence # local path
pi list shows what is installed, pi remove npm:pi-cadence removes it.
No API key, no /login. The provider shows up as cadence with nine models.
Use
/cadence state: lanes, free quota per model
/cadence list saved relays, numbered
/cadence add [relay] add a relay, or open the editor to paste many
/cadence remove [n] delete one by number
/cadence test [n] probe one relay by number, or all of them
/cadence help show the commands
/cadence add relay saves one relay; with no argument it opens a multiline
editor. Paste your whole proxy list, in either format, and it figures out each
line:
45.38.107.97:6014:user:pass
http://user:pass@45.38.107.97:6014
proxy.example.com:8080
Each relay is probed against the provider before it is saved. Dead ones are reported and dropped, so a broken relay never silently joins the pool. Probing costs no rate limit.
/cadence test warns when two relays share one egress IP, printing the shared
IP, because those two share a bucket and the real limit is lower than the lane
count suggests.
Relays are stored in ~/.pi/agent/cadence.json with mode 0600. If the file
is corrupt, the next save moves it to cadence.json.bak instead of overwriting it.
Then pick a model:
/model → cadence/Qwen3.8-27B
Models
| model | thinking | vision | context |
|---|---|---|---|
gpt-oss-20b, gpt-oss-120b |
low, medium, high | no | 131k |
Qwen3.8-27B |
off, low, medium, xhigh | yes | 262k |
Qwen3.6-27B |
off, low, medium, high, xhigh, max | yes | 262k |
Qwen3.5-9B, Qwen3.5-397B-A17B |
off, low, medium, high | yes | 262k |
Mistral-Small-3.2-24B-Instruct-2506 |
no | yes | 131k |
Qwen2.5-VL-72B-Instruct |
no | yes | 32k |
Meta-Llama-3_3-70B-Instruct |
no | no | 131k |
Thinking levels are mapped to what each model actually accepts. high on
Qwen3.8 became xhigh because the provider rejects high there; off does not
exist on gpt-oss, so those models offer no off.
How requests are spread
Every request takes a slot from a bucket keyed by (lane, model) and streams
through it. pi-cadence does not intercept the stream: every event, including
errors, is forwarded to Pi untouched.
A lane that has spent a slot cannot serve that model until the window resets, so the next request takes another lane. When every lane is full, the request waits for the nearest window reset (up to 65s) instead of failing. Retries are Pi's to make; each one simply takes the next slot, and the least-spent lanes come first.
The endpoint
pi-cadence talks to one fixed endpoint. Its id, base URL and rate limit live in
src/endpoint.ts, the model catalog in src/models.ts; if the service changes,
those are the only files to edit.
What this does not do
- 22 rpm is 22 HTTP requests, not 22 chat turns. One agent turn with ten tool calls spends eleven of them.
- Buckets are per model, so switching models starts a fresh 22 rpm but does not refund the previous model's usage.
- Compaction and subagents draw from the same buckets.
- Two Pi windows each count the same per-IP budget from zero; the 429s spend slots until the window resets.
Development
npm test # unit tests, no network
npm run typecheck # tsc --noEmit
npm run lint # biome lint
npm run format # biome format
License
MIT. See LICENSE.