konstera-sdk (Python)

Build Konstera agents in Python without hand-writing MCP client boilerplate, HTTP servers, or env-var parsing. Subclass Agent, override handle(), call run().

Status: v0.1.0 (M6). Log + LLM (stub) kernel services only. Vector / state / credentials / cost / tool / agents / human / prompt come in M7.

Install

The SDK ships as a Python package. During development, install from source:

pip install -e sdk/python

Container agents install it via a multi-stage Dockerfile — the SDK wheel is built in one stage and copied into the final image. See agents/hello-world-python/Dockerfile for the reference pattern.

Minimal agent

from konstera_sdk import Agent, AgentContext, AgentRequest, AgentResponse, run

class HelloAgent(Agent):
    async def handle(self, req: AgentRequest, ctx: AgentContext) -> AgentResponse:
        name = req.input.strip() or "world"
        await ctx.log.info(f"Hello, {name}!", input=name, language="python")
        return AgentResponse(content=f"Hello, {name}!")

if __name__ == "__main__":
    run(HelloAgent())

That's it — ~10 lines for a complete agent. run() handles:

• Reading KONSTERA_KERNEL_MCP_URL, KONSTERA_AGENT_INSTANCE_ID, KONSTERA_AGENT_ID, KONSTERA_AGENT_VERSION from the environment (injected by the kernel's ContainerAgentHost)
• Setting up the MCP client back to the kernel
• Hosting /health, /describe, /invoke on port 8080
• Routing invocations to your handle()
• Graceful shutdown

API

Agent (base class)

class Agent:
    async def initialize(self, ctx: AgentContext) -> None: ...
    async def handle(self, req: AgentRequest, ctx: AgentContext) -> AgentResponse: ...
    async def shutdown(self, ctx: AgentContext) -> None: ...

Override handle() (required). Override initialize() / shutdown() for setup/teardown.

AgentContext

ctx.instance_id     # str — this agent instance's UUID
ctx.agent_id        # str — e.g. "io.konstera.hello-world-python"
ctx.agent_version   # str — e.g. "0.1.0"
ctx.log             # LogService — ctx.log.info/warn/error
ctx.llm             # LlmService — ctx.llm.complete (stub in M6)

LogService

await ctx.log.info(message, **fields)
await ctx.log.warn(message, **fields)
await ctx.log.error(message, **fields)

Structured fields (kwargs) are JSON-serialised and enriched by the kernel with tenant/instance/invocation context, then flowed to Serilog + App Insights.

LlmService (stub in M6)

text = await ctx.llm.complete(model="claude-sonnet-4-6", prompt="Hello")

Real Claude / OpenAI / Gemini adapters land in M10 (SPEC §10). Interface is stable now; agent code written today will keep working when the stub goes away.

run(agent, port=8080)

Blocks until the process is stopped. Handles all lifecycle: reads env, creates context, calls initialize(), serves HTTP, calls shutdown() on SIGTERM.

Direct-run mode

When KONSTERA_KERNEL_MCP_URL is unset, kernel callbacks become no-ops (they log a warning and return successfully). This lets you develop agents locally without the full Konstera stack running:

python -m your_agent_module
# Serves on :8080 — POST /invoke works but ctx.log/ctx.llm skip.

Licence

Apache-2.0. See the repo-level LICENSE.