tai-daf-sdk 2.0.0a4

Python SDK for DAF (Declarative Agentic Framework) — thin MCP client

tai-daf-sdk

Python SDK for DAF — Declarative Agentic Framework. Thin, MCP-native client. Every call goes to the same MCP tool the backend exposes at /mcp — no separate REST surface to keep in sync, no per-resource wrappers to hand-write.

PyPI: https://pypi.org/project/tai-daf-sdk/

Install

Alpha channel while the surface stabilises:

pip install --pre tai-daf-sdk

Python ≥ 3.10.

30-second quickstart

from tai_daf_sdk import DAF

with DAF(
    base_url="https://daf-sdk-backend-dev.azurewebsites.net",
    api_key="daf_...",
) as client:
    # List LLM endpoints on your account
    endpoints = client.llm_endpoints.list()
    ep_id = endpoints["endpoints"][0]["id"]

    # Create an agent bound to that endpoint
    agent = client.agents.create(
        name="my-assistant",
        system_instructions="You answer politely.",
        llm_endpoint_id=ep_id,
        confirm=True,
    )
    print("Created:", agent["created"]["id"])

Async version

Same API, async prefix, always inside async with:

import asyncio
from tai_daf_sdk import AsyncDAF

async def main():
    async with AsyncDAF(base_url="...", api_key="daf_...") as client:
        agents = await client.agents.list()
        for a in agents["agents"]:
            print(a["id"], a["name"])

asyncio.run(main())

How the dispatch works

The SDK is a thin dispatch layer over MCP tools. client.<resource>.<method>(**kwargs) maps to daf_<method>_<resource>:

Python call	MCP tool
client.agents.list()	daf_list_agents
client.agents.create(...)	daf_create_agent (singular fallback)
client.agents.describe(name_or_id=...)	daf_describe_agent
client.agents.update(agent_id=..., patch=...)	daf_update_agent
client.agents.delete(agent_id=..., confirm=True)	daf_delete_agent
client.swarms.list()	daf_list_swarms
client.llm_endpoints.list()	daf_list_llm_endpoints

Singular fallback. Every collection is plural on the SDK (client.agents, client.swarms) so it reads naturally. When you call .create() the SDK first tries daf_create_agents (plural), and if that's not registered, falls back to daf_create_agent (singular). This matches the backend's real naming — list ops are plural, per-item ops are singular.

Escape hatch for irregular tool names. A handful of tools don't fit daf_<method>_<resource> — like daf_attach_guardrail_to_agent, daf_run_tool, daf_mcp_health_check, daf_export_agent, daf_import_agent, daf_list_node_types. Use client.call(tool_name, args_dict):

result = client.call(
    "daf_attach_guardrail_to_agent",
    {"agent_id": agent_id, "guardrail_id": gid, "confirm": True},
)

Async client has the same escape hatch:

result = await async_client.call("daf_attach_guardrail_to_agent", {...})

Errors

The SDK maps the backend's structured error codes to typed exceptions:

Backend code	Exception
VALIDATION_ERROR	BadRequestError
ENTITY_NOT_FOUND	NotFoundError
FORBIDDEN	PermissionError
NAME_TAKEN / DUPLICATE_URL	ConflictError
QUOTA_EXCEEDED	RateLimitError
Others	APIError (base class)

from tai_daf_sdk.exceptions import BadRequestError

try:
    client.agents.create(name="broken", model_provider="azure", confirm=True)
except BadRequestError as exc:
    print(f"Invalid config: {exc}")

ConnectionError is raised for transport failures (backend unreachable, DNS, TLS).

Configuration

Constructor accepts:

• base_url (str) — DAF backend URL. Local dev = http://localhost:8012, dev cloud = https://daf-sdk-backend-dev.azurewebsites.net.
• api_key (str) — DAF API key (starts with daf_). Get one from Settings → API keys in the webapp.
• token (str) — Bearer JWT alternative to API key. Only one of api_key / token at a time.
• timeout (float, default 60.0) — HTTP timeout in seconds.

Alternatively, set DAF_MCP_STDIO_COMMAND to run the SDK against a local stdio MCP server instead of HTTP.

Streaming

For tools whose backend implementation streams SSE events (daf_start_workflow, daf_subscribe_workflow), use client.stream(tool_name, args):

for event in client.stream("daf_subscribe_workflow", {"run_id": run_id}):
    if event.get("event") == "output_node":
        print(event["data"]["content"], end="", flush=True)

Async equivalent iterates via async for:

async for event in async_client.stream("daf_subscribe_workflow", {"run_id": run_id}):
    ...

Full guide + more examples

• Guide: docs/GUIDE.md — every capability with working code.
• Examples: examples/ — runnable scripts:
  • 01_hello_agent.py — create + describe + delete an agent
  • 02_swarm_workflow.py — compose a two-agent swarm via workflow drafts
  • 03_triggers.py — webhook + cron + event triggers
  • 04_memory.py — per-agent memory + shared memory
  • 05_guardrails.py — attach a guardrail to an agent
  • 06_streaming.py — subscribe to workflow SSE events

Each example is self-contained. Pass --url and --api-key (or set DAF_MCP_URL / DAF_API_KEY env vars).

Regenerating from your backend's catalog

When the backend adds tools, the SDK's tool registry catches up automatically:

python -m tai_daf_sdk.codegen \
    --mcp-url https://daf-sdk-backend-dev.azurewebsites.net \
    --api-key daf_... \
    --out tai_daf_sdk/

Regenerates:

• _tool_registry.py — MCP tool → resource+method map
• types.py — TypedDict per tool input
• client.pyi — IDE stubs for autocomplete
• models.py — Pydantic wrappers where declared

The CI workflow .github/workflows/mcp-catalog-sync.yml runs this on schedule and fails when the shipped registry drifts from the live backend catalog.

End-to-end test scenarios

scripts/sdk_tests/ — 45 scripts (24 customer sandbox stories + 21 read-only sweeps) verify the whole surface against a live backend. Same story IDs as the MCP-side suite in daf-sdk-backend/scripts/mcp_tests so you can cross-check SDK ↔ MCP agree on the customer contract.

python -m scripts.sdk_tests.run_all \
    --url https://daf-sdk-backend-dev.azurewebsites.net \
    --api-key daf_...

Filter with --set 1|2 or --filter us_01 for one scenario at a time.

License

MIT.