pi-relay

Fan Claude Code work out to a non-Claude model through pi. Claude keeps orchestration and judgment. pi, running Kimi K3 by default or any model pi supports, does the generation. Works in any folder once installed; nothing goes into the project.

What you get

Install

On any machine:

npm i -g @earendil-works/pi-coding-agent     # pi itself
pi                                           # once, interactively: /login to your provider (e.g. fireworks)
claude plugin marketplace add jakelevirne/pi-relay
claude plugin install pi-relay@pi-relay

For a local checkout instead of GitHub: claude plugin marketplace add /path/to/pi-relay. Then in any Claude Code session:

"${CLAUDE_PLUGIN_ROOT}/bin/pi-doctor"      # or: ~/.claude/plugins/cache/pi-relay/pi-relay/*/bin/pi-doctor

Use

Ask Claude to "fan this out to kimi" or "have pi do the implementation", and the pi-fanout skill takes over. Or be explicit:

Agent({ subagent_type: 'pi-relay:pi-runner', prompt: 'MODE: text\nPROMPT:\n...' })
Workflow({ name: 'pi-relay:design-build-review', args: { spec: '...', files: [{ file: 'app.js', note: '...' }] } })

Inside your own Workflow scripts:

const kimi = (mode, prompt, opts = {}) => agent(
  `MODE: ${mode}\nPROMPT:\n${prompt}`,
  { agentType: 'pi-relay:pi-runner', model: 'haiku', effort: 'low', ...opts })

Choosing the model

Defaults are Fireworks and accounts/fireworks/models/kimi-k3. Override per machine with env vars PI_RELAY_PROVIDER, PI_RELAY_MODEL, PI_RELAY_THINKING, or per task with MODEL:, PROVIDER:, THINKING: lines in the relay task. pi --list-models kimi shows what pi knows.

Why the prompt goes through a file

Prompts contain quotes, backticks, $(...), and multi-line text. Passing them as shell arguments is where relays break. The relay agent writes the prompt with its Write tool, and pi-run pipes that file into pi's stdin. Nothing is ever shell-parsed.

One more thing pi-run handles: pi -p reads its prompt from stdin when stdin is not a terminal, and blocks forever if that stdin is an open pipe that never closes. That is exactly what agent harnesses hand it. Feeding the file on stdin sidesteps it.

Cost shape

Each pi step costs one small Haiku relay (a few thousand tokens) plus pi's own usage on your provider. Concurrency is bounded by the Workflow slot cap and the provider's rate limit.