xiaoguai (Python wrapper)

pip install xiaoguai — a thin Python launcher that bundles the Rust xiaoguai CLI binary inside a platform-specific wheel.

On Debian 12 / Ubuntu 24 and other PEP 668 "externally-managed" systems, pip install into the system Python is blocked. Use pipx instead: sudo apt install -y pipx && pipx ensurepath && pipx install xiaoguai.

After install:

xiaoguai --help
xiaoguai chat --mock --prompt "hello"
xiaoguai doctor    # self-check: database / providers / Ollama / port

To keep xiaoguai serve running in the background across reboots: xiaoguai service install (systemd on Linux — needs sudo; a per-user launchd agent on macOS — no root).

The console script forwards every argument to the bundled native binary. There is no Python agent logic in this package — it exists so pip users have an install path alongside Cargo, Homebrew, and the standalone tarball.

Supported platforms

The CI matrix produces wheels for:

Other platforms (Alpine / musl, Windows, FreeBSD) are out of scope. Build from source instead (needs Rust 1.94 or newer):

cargo install --path crates/xiaoguai-cli

Offline hosts: the wheel bundles the native binary, so it installs without network on the target. Download it on a connected machine and carry it over:

pip download xiaoguai --no-deps --only-binary=:all: \
  --platform manylinux_2_28_x86_64 --python-version 3.12 -d ./offline
# transfer ./offline, then on the air-gapped host:
pip install --no-index ./offline/xiaoguai-*.whl

For the systemd/package route and the SEC-01 credential step, see Offline / air-gapped install.

HTTP client (wave-3)

pip install 'xiaoguai[client]' — adds xiaoguai.client.XiaoguaiClient, a synchronous HTTP client for the xiaoguai-api REST server (requires httpx>=0.25).

Note: the client snippets below predate the single-user pivot (DEC-033). The live API now serves on :7600 with optional HTTP Basic auth and no tenant scoping. The bundled binary launcher above is the supported path; treat these examples as illustrative pending a client refresh.

Covered endpoints (v1.2.x)

Quick start

from xiaoguai.client import XiaoguaiClient

with XiaoguaiClient("http://localhost:7600", token="my-bearer-token") as c:
    # HotL — boundary policy admin
    policy = c.create_hotl_policy(
        tenant_id="my-tenant-uuid",
        scope="llm_call",
        window_seconds=3600,
        max_count=100,
        escalate_to="ops@example.com",
    )
    policies = c.list_hotl_policies(tenant_id="my-tenant-uuid", scope="llm_call")
    c.delete_hotl_policy(policy.id)

    # Outcomes — ROI telemetry
    c.record_outcome(
        tenant_id="my-tenant",
        agent_name="sales-bot",
        kind="revenue_usd",
        value=1500.0,
        description="Closed enterprise deal",
    )
    summary = c.outcomes_summary(tenant_id="my-tenant", range="7d")
    ts = c.outcomes_timeseries(tenant_id="my-tenant", range="30d", kind="hours_saved")

    # Skills — pack marketplace
    catalog = c.list_skill_catalog()
    pack = c.install_skill(tenant_id="my-tenant", pack_slug="rag-legal")
    installed = c.list_installed_skills(tenant_id="my-tenant")
    c.uninstall_skill(pack.id)

Error handling

from xiaoguai.client import (
    XiaoguaiNotFoundError,
    XiaoguaiValidationError,
    XiaoguaiConflictError,
)

try:
    c.install_skill(tenant_id="t1", pack_slug="rag-legal")
except XiaoguaiConflictError:
    print("already installed")
except XiaoguaiNotFoundError:
    print("unknown pack slug")

Typed models

HotlPolicy, HotlVerdict, OutcomeRecord, OutcomeSummary, OutcomeTimeseries, InstalledSkillPack, SkillPackEntry — all frozen dataclasses with from_dict class methods.

Troubleshooting

If xiaoguai after a fresh install prints "native binary not bundled", the wheel matched on architecture but its package data is empty (rare — usually an sdist install rather than a wheel). Set XIAOGUAI_PY_DEBUG=1 to see the resolution path the launcher tried.

Documentation

Full documentation, configuration, and architecture notes live in the upstream repository — see the main README.

License

Apache-2.0. Same license as the upstream Rust project.