kimi-calls 0.1.0

CLIs that delegate bulk reading, boilerplate generation, and chat-log extraction to Kimi K2.5 — saving tokens inside coding agents.

kimi-calls

Three CLIs that delegate the bulky parts of a coding-agent session to Kimi K2.5 via the Moonshot platform API. The reasoning stays in your main agent (Claude Code, Codex, Cursor, …); the bulk reading, boilerplate generation, and chat-log extraction get offloaded — saving tokens, latency, and money.

Tool	Replaces	Use when
ask-kimi	reading 3+ files into context	"what does this codebase do across these N files?"
kimi-write	generating boilerplate from scratch	tests, configs, docstrings, repetitive scaffolding
extract-chat	parsing session JSONL by hand	turn a Claude Code or Codex session into plain dialogue

Install

Requires Python 3.10+.

pip install kimi-calls
# or
uv tool install kimi-calls

Get a Moonshot API key

1. Sign in at https://platform.moonshot.ai (the international platform — not moonshot.cn, which is China-only).
2. Add credit under Billing (pay-as-you-go; a few dollars goes a long way).
3. Create a key under API Keys → it'll start with sk-.

Make the key available to the CLIs

The tools read MOONSHOT_API_KEY from the environment. Pick one:

Persist for every shell (recommended). Append to ~/.zshrc (macOS default) or ~/.bashrc:

echo 'export MOONSHOT_API_KEY=sk-...' >> ~/.zshrc
source ~/.zshrc

Per-project via .env. If you use direnv or a similar loader:

echo 'MOONSHOT_API_KEY=sk-...' >> .env

One-off in the current shell:

export MOONSHOT_API_KEY=sk-...

Verify it's set:

echo $MOONSHOT_API_KEY    # should print sk-...

Optional: pin a model with KIMI_MODEL (default kimi-k2.5).

Usage

ask-kimi — bulk reading

ask-kimi --paths src/api.py src/db.py src/auth.py \
         --question "Where do we handle session expiry?"

Files are concatenated before the question, so repeat calls on the same --paths list hit Moonshot's prefix cache. Keep the order stable across calls.

kimi-write — boilerplate generation

kimi-write --spec  "pytest suite covering build_corpus" \
           --context src/kimi_calls/_client.py tests/test_extract.py \
           --target  tests/test_corpus.py

Then Edit only what needs fixing. Saves the round-trips you'd otherwise spend writing scaffolding token by token.

kimi-write refuses to overwrite an existing target unless you pass --force.

extract-chat — session text for documentation

extract-chat ~/.claude/projects/.../session.jsonl -o /tmp/chat.txt
# or for Codex:
extract-chat ~/.codex/sessions/.../rollout.jsonl -o /tmp/chat.txt

When writing to -o/--output, extract-chat refuses to overwrite an existing file unless you pass --force.

Strips tool calls, system prompts, and binary blocks. Auto-detects Claude Code ({"type": "user", "message": {...}}) vs Codex ({"role": "user", "content": ...}) format. Pipe the result into ask-kimi for documentation work:

extract-chat session.jsonl -o /tmp/chat.txt
ask-kimi --paths /tmp/chat.txt docs/architecture.md \
         --question "What architecture decisions from this session belong in the doc?"

Wire it into your agent

See examples/ for ready-to-drop snippets:

• Claude Code — examples/claude-code/settings.json (permission allowlist) and CLAUDE.md (delegation rules).
• Codex — examples/codex/AGENTS.md.
• Cursor — examples/cursor/.cursor/rules/kimi-delegation.mdc (always-on project rule).

The full delegation policy lives in docs/delegation-policy.md — agent-agnostic guidance on when to delegate and when not to.

Development

uv pip install -e ".[dev]"
pytest

Tests are pure parsing — no API calls, no key required.

License

MIT.