NiceKit

A multi-tenant Agent + Knowledge Base platform SDK for Python

NiceKit is a platform framework package, not a utility library. It was extracted from a production enterprise AI application. Install it, wire it up, and your project gets a multi-tenant foundation, an LLM routing layer, an agent runtime, a knowledge base and operational observability — 228 REST endpoints across 182 paths, 64 tables, 46 of them under PostgreSQL FORCE ROW LEVEL SECURITY.

pip install nicekit

from nicekit.api.v1.router import default_routers
from nicekit.runtime.app_factory import create_app
from nicekit.runtime.bootstrap import install_default_ports

install_default_ports()
app = create_app(routers=default_routers())

That is a running platform with authentication, model configuration, knowledge bases and agent chat — provided the infrastructure below is in place.

Requirements

Hard dependencies, none of them substitutable:

• Python 3.13+
• PostgreSQL with pgvector, pg_trgm, zhparser and pgcrypto
• Redis
• S3-compatible object storage
• Celery is optional — there is an inline dispatch fallback

pip install nicekit alone will not give you a running system. The repository ships a docker compose setup and a custom PostgreSQL image with the three extensions pre-built; the official postgres image will fail migrations.

What you get

Deliberately not included: any domain model. No orders, tickets, projects or customers — you define those and plug them in through extension points. NiceKit also does not abstract over the database (PostgreSQL is required by design) and ships no industry vocabularies.

Bridging your existing identity system

Authentication converges on a single dependency. Replace the PrincipalResolver behind it and the SDK plugs into whatever account system you already have — you change assembly code, not the data model.

from nicekit.api.deps import Principal, set_principal_resolver
from nicekit.api.v1.router import AUTH_ROUTER_NAMES, default_routers
from nicekit.tenancy import subject_uuid, tenant_uuid

async def resolver(request) -> Principal:
    claims = verify_host_token(request.headers["authorization"])
    return Principal(
        org_id=tenant_uuid(claims["company_id"]),   # external key -> UUID, deterministic, no mapping table
        user_id=subject_uuid(claims["user_id"]),
        role=ROLE_MAP[claims["role"]],
    )

set_principal_resolver(resolver)                    # must happen before create_app
app = create_app(routers=default_routers(exclude=AUTH_ROUTER_NAMES))

Three modes are supported: managed (SDK owns identity), bridged (your app owns it) and single-tenant (no tenant concept, or auth at the gateway).

Extension points

The SDK never imports host code. Everything domain-specific goes through registration interfaces — custom tools with seven-axis permissions, resource resolvers, context providers, entity types, notifiers, trace/usage sinks, roles, scheduled tasks and more.

from nicekit.agent.tools import ToolContext, ToolRegistry, tool_permission
from nicekit.domain.agent_permission import (
    PermissionScope, ToolCategory, ToolDelegation, ToolEffect, ToolReversibility, ToolRisk,
)

my_tools = ToolRegistry("myapp")

@my_tools.register(
    "ticket_close",
    "Close a ticket",
    {"type": "object", "properties": {"ticket_id": {"type": "string"}},
     "required": ["ticket_id"], "additionalProperties": False},
    permission=tool_permission(
        effect=ToolEffect.TRANSITION,
        risk=ToolRisk.SENSITIVE,
        categories=(ToolCategory.WORKFLOW,),
        reversibility=ToolReversibility.COMPENSATABLE,
        delegation=ToolDelegation.REVIEWABLE,     # requires AI review or human approval
        scope=PermissionScope.RESOURCE,
        material_arguments=("ticket_id",),
    ),
)
async def ticket_close(ctx: ToolContext, args: dict) -> dict:
    ...

app = create_app(routers=default_routers(), tool_registry=my_tools)

Documentation

Full documentation lives in the repository. It is written in Chinese; the repository also has an English overview.

• Documentation index
• SDK guide — 30-minute start, all extension points, common recipes
• Integration guide — bridging an existing tenancy and auth system
• Agent brief — structured, copy-pasteable facts written for AI coding assistants
• Data model — all 64 tables and the RLS mechanism

Status

0.1.2. The public API is not frozen. During 0.x, minor versions may carry breaking changes — read the changelog before upgrading.

License

Apache License 2.0 © 2026 patrickleehua