castia

Idiomatic, FastAPI-style Python SDK for Microsoft Foundry hosted agents.

castia lets a hosted agent speak Foundry's three wire protocols — Activity (Teams/Bot Framework), OpenAI responses, and invocations — through protocol-named decorators, dependency injection (Depends), and typed builders for messages, Adaptive Cards, entities, and invoke envelopes. You decorate a handler, return a value, and the framework does the rest — the auth chains, hosting, Activity routing, and telemetry stay out of your file.

Install

pip install castia
# or, with uv:
uv add castia

Requires Python 3.11+.

Quickstart

from castia import Agent, Depends, Model, Teams

app = Agent(name="my-agent")

def gpt4o() -> Model:
    return Model("gpt-4o")

@app.activity(Teams.direct, Teams.group, Teams.channel_mention)
async def reply(text: str, model: Model = Depends(gpt4o)) -> str:
    return await model.respond(text)

if __name__ == "__main__":
    app.run()

Decorate a handler with the surfaces it answers on, return a str, and the framework sends it as the Teams reply. The model is built once for the process and injected via Depends — the model choice stays visible in your file instead of being buried in the framework. (Model() with no argument falls back to AZURE_AI_MODEL_DEPLOYMENT_NAME; get_model / use_model("gpt-4o") are zero-config conveniences.)

Composing protocols with routers

Like FastAPI's include_router, an Agent composes Routers so each protocol can live in its own module:

from castia import Agent
from handlers import activity, responses, invocations

app = Agent(name="my-agent")
app.include(activity.router, responses.router, invocations.router)

Richer replies

Handlers can take a Message and reach for typed builders — Adaptive Cards, suggested actions, citations, mentions, sensitivity labels, live-typing streamers, and reactions:

from castia import Depends, Message, Model, Reaction, Router, Teams, get_model

router = Router()

@router.activity(Teams.direct)
async def reply(text: str, msg: Message, model: Model = Depends(get_model)) -> None:
    await msg.react(Reaction.eyes)
    answer = await model.respond(text)
    await msg.say(answer)

Consume a toolbox

A Foundry toolbox is a curated set of tools the platform exposes behind one MCP-compatible endpoint, with centralized auth, governance, and versioning. castia turns it into a single Responses-API mcp tool spec the model service resolves server-side — no local impl, no function loop. Build the spec and hand it to the model as an extra_specs entry:

from castia import Depends, Model, Teams, get_model, toolbox_mcp_tool, toolbox_token

@app.activity(Teams.direct)
async def reply(text: str, model: Model = Depends(get_model)) -> str:
    tool = toolbox_mcp_tool(token=await toolbox_token())   # reads TOOLBOX_* env
    return await model.respond_with_tools(
        text, tools=[], activity=None, extra_specs=[tool] if tool else []
    )

toolbox_mcp_tool() (called with no endpoint) reads the environment via resolve_toolbox_endpoint, whose precedence is:

1. an explicit full URL in TOOLBOX_ENDPOINT / TOOLBOX_MCP_ENDPOINT;
2. the platform-native TOOLBOX_<NAME>_MCP_ENDPOINT that the azd ai toolbox extension writes, keyed off TOOLBOX_NAME (see platform_endpoint_env);
3. composed from FOUNDRY_PROJECT_ENDPOINT + TOOLBOX_NAME (+ optional TOOLBOX_VERSION). The unversioned URL resolves the promoted default version, so a version bump needs no redeploy.

It returns None when no toolbox is configured, so "no toolbox" just attaches no tool. Auth is either a bearer token (minted from the container's managed identity by toolbox_token()) or a stored-connection project_connection_id. A Foundry IQ knowledge base is the same shape via knowledge_base_mcp_tool.

Deploying: the azd ai toolbox extension writes TOOLBOX_<NAME>_MCP_ENDPOINT into the azd environment, but does not auto-inject it into a hosted container — declare that env passthrough on your container yourself (there is no azure.yaml/manifest step in castia for it).

Gotcha: do not copy rai_config.rai_policy_name: Microsoft.Default from the azd ai toolbox create --help example — it is invalid on the project and 500s at tool enumeration (tools/list). Omit the policies block.

Validation status. Validated live (Foundry Responses path, App Insights-traced): the end-to-end pipe (env → compose URL → attach one mcp tool → tools/list + tools/call), a raw https://ai.azure.com bearer minted in-container (no project_connection_id required), both env forms, and unversioned→default-version resolution. Doc-derived / not yet live: knowledge_base_mcp_tool (Foundry IQ), connection-backed tools (Azure AI Search / remote-MCP / A2A), the Activity path with a toolbox, and approval-gated tools (require_approval other than "never").

Protocols

castia publishes handlers for the protocols in PUBLISHABLE_PROTOCOLS:

• Activity — Teams / Bot Framework message and invoke turns.
• responses — the OpenAI responses wire shape.
• invocations — Foundry invoke envelopes (tool execution, agent-to-agent).

Observability & evaluation

castia configures Foundry/Agent 365 telemetry for you when the agent starts. By default it emits GenAI spans (the chat {model} spans the Foundry Traces UI keys off) but does not record the prompt/response content onto them.

Recording content is what makes an agent's traces evaluable — trace-based evaluators read the input/output text from the GenAI spans, which is only present when content recording is enabled. Turn it on deliberately via configure_observability:

from castia.observability import configure_observability

# Records prompt/response text onto GenAI spans so traces can be evaluated.
configure_observability(enable_content_recording=True)

Resolution order for each flag is explicit argument > environment variable > default:

Passing nothing preserves the default behavior. Telemetry setup is best-effort: a failure is logged, never raised, so it can't break startup or a turn.

Security caveat: enabling content recording writes prompt and response text to Application Insights. Only enable it where storing that content is acceptable for your data-handling and privacy requirements.

Building a scored eval suite

Once your traces are evaluable, python -m castia eval wraps the azd ai agent eval extension to synthesize and run a scored eval suite — a generated JSONL dataset plus an auto-generated, weighted rubric (a custom multi-dimension evaluator):

# Offline gate — validate eval.yaml + rubric files, no Azure, free in CI:
python -m castia eval check

# Synthesize a rubric + dataset from the agent instruction (billable):
python -m castia eval generate --agent my-agent --max-samples 25

# Re-upload locally edited rubric/dataset files as a new version:
python -m castia eval update --evaluator-only

# Submit a scored run against the deployed agent (billable):
python -m castia eval run

check is a pure, offline referential-integrity gate: it resolves every evaluator/dataset local_uri and validates each rubric dimensions file. generate and run submit billable Foundry jobs, so both accept --dry-run to print the resolved azd command line without submitting anything. The azd wrappers need the build-time extra: pip install 'castia[deploy]'.

The rubric dimensions file is a bare JSON list where each entry is keyed by id (a stable slug like correct_outcome), with an optional always_applicable: true on the catch-all dimension. That cross-SDK shape is pinned in the monorepo at spec/conformance/rubric/.

Optimizer-readiness

The Foundry Agent Optimizer searches for a better system prompt (and, when you declare tools, better tool descriptions) by running candidates against your eval suite. Making a castia agent optimizer-ready is three things: install the runtime resolver, ship a baseline config, and source your model + instructions from that config instead of hardcoding them — so the optimizer can swap in a candidate with zero handler changes.

Bind your model dependency with configured_model() and thread its resolved instructions through:

from castia import Depends, Model, Router, configured_model

router = Router()
gpt = configured_model()  # resolves baseline (or the injected candidate) once

@router.responses()
async def reply(text: str, model: Model = Depends(gpt)) -> str:
    return await model.respond(text)   # instructions flow into responses.create

configured_model() calls load_agent_config(), which is best-effort: if the optimizer package isn't installed, resolution fails, or no config is found, it degrades to environment defaults (AZURE_AI_MODEL_DEPLOYMENT_NAME, no instructions) — the agent runs identically with or without the optimizer.

Ship a baseline under .agent_configs/baseline/:

.agent_configs/baseline/
  metadata.yaml       # model, instruction_file, (optional) tool_file pointers
  instructions.md     # the system prompt the optimizer tunes
  tools.json          # optional: tool specs the optimizer may reword

If your agent declares tools with app.tools(...), keep the baseline tools.json in sync with the code using the build-time reconciler:

python -m castia optimize          # write/refresh .agent_configs/baseline/tools.json
python -m castia optimize --check   # CI drift gate (exits non-zero, writes nothing)

Config resolution order is first-wins: OPTIMIZATION_CONFIG (inline JSON) → resolver API (OPTIMIZATION_CANDIDATE_ID + OPTIMIZATION_RESOLVE_ENDPOINT) → local .agent_configs/ → environment defaults. An explicit config_dir argument (or OPTIMIZATION_LOCAL_DIR) affects only the local source — pass it anchored to your app root so the baseline resolves the same under python app.py and python -m castia. That contract is pinned for every SDK in spec/conformance/optimization/.

Switching to a reasoning (or RFT-tuned) model

The optimizer's model search can land on a reasoning model — an o-series or GPT-5 deployment, or one you mint yourself with reinforcement fine-tuning (RFT). Those models take a reasoning.effort control that plain chat models don't. Model exposes it as reasoning_effort (minimal|low|medium|high):

o4 = use_model("o4-mini-rft-2025", reasoning_effort="high")

An unset effort omits the field entirely, so chat models are called exactly as before; a bad level raises at construction rather than as a 400 mid-turn. An operator can also switch a deployed agent onto a reasoning model with zero code by setting MODEL_REASONING_EFFORT — an explicit argument still wins. Because configured_model() builds its Model through the same path, that env override flows through to the resolved candidate automatically.

Responses-only constraint: the optimizer accepts only single-protocol responses agents — submitting a multi-protocol agent (one that also speaks activity/invocations) is rejected with a 400 at submission. Project a responses-only sibling from the same handler code with app.responses_only(), deploy that as its own service, optimize it, then apply the winning .agent_configs candidate back to your live agent.

The runtime resolver and the reconciler need the optimizer extra: pip install 'castia[optimize]'.

Reinforcement fine-tuning (RFT)

RFT is the fourth lifecycle step — build → evaluate → optimize → switch models. It trains a reasoning model against a grader (a reward function) instead of labeled answers, minting a new fine-tuned deployment that becomes a candidate in the optimizer's model search. castia ships the build-time tooling to prepare, validate, and (behind one guarded seam) submit an RFT job — the same three-seam shape as the eval suite: pure builders, offline validators, and one billable submit seam.

# Offline gate — validate an RFT dataset (+ grader), no Azure, free in CI:
python -m castia finetune check --dataset train.jsonl --validation val.jsonl --grader grader.json

# Bridge an eval rubric into a score_model grader (offline):
python -m castia finetune grader --rubric rubric.json --model gpt-4o --out grader.json

# Submit a billable RFT job (use --dry-run to print the payload and submit nothing):
python -m castia finetune submit --model o4-mini --dataset train.jsonl \
  --validation val.jsonl --grader grader.json --reasoning-effort high --dry-run

A grader is one of string_check, text_similarity, score_model, python, multi, or endpoint (preview); templates reference two namespaces only — {{ sample.output_text }} and {{ item.<field> }}. Datasets are JSONL chat messages[] rows whose final message role must be user, with extra top-level keys as the item.* ground truth; both train and validation splits are required. rubric_to_score_model() bridges an eval rubric straight into a score_model grader — the natural tie between the evaluate and switch-models steps.

⚠️ Provisional / doc-derived. The grader JSON schema, the RFT hyperparameter names, and that fine_tuning.jobs.create accepts this payload are derived from the Foundry RFT how-to and have not been confirmed against a live RFT job. The builders and validators are fully offline-tested; treat the submitted wire shape as provisional until a real submission validates it. The language-neutral contract is pinned in spec/conformance/graders/.

Design

castia is deliberately import-cheap: import castia never pulls in the instrumented Azure/OpenAI/httpx stacks, so telemetry can be configured before those libraries load. The heavy imports are deferred into the methods that need them.

License

MIT © 2026 Seth Juarez