Build with the local Soterias SDK. Get started with Python and TypeScript.
SoteriasDocsBuild
Core components

Measure local runs

Inspect counts, errors, cancellations and duration distributions. Token usage and costs stay unmeasured until you explicitly supply evidence for them.

When to use this

Use this to inspect failures, cancellations and timing distributions after running your application. Supply usage or cost evidence explicitly when you have it.

Install the matching API preview

These examples use the API 2 preview, not names promised by the public RC5 release. Existing RC5 entry points remain available. The Python wheel below is verified on Windows x64; it is not a cross-platform release.

shell
python -m pip install https://docs.soterias.io/downloads/api-preview/soterias_local-0.1.0rc5%2Bapi2-cp311-abi3-win_amd64.whl
shell
npm install https://docs.soterias.io/downloads/api-preview/soterias-core-0.1.0-rc.5-api.2.tgz
npm install --save-dev typescript@5.8.3 @types/node@24

Python imports soterias. TypeScript imports @soterias/core. Use Python 3.11+ for this wheel and Node 22.23.4+ for these examples. For a .mts file, compile with npx tsc example.mts --target ES2022 --module NodeNext --moduleResolution NodeNext --skipLibCheck, then run the emitted .mjs file.

Before you run

This program creates its own local example state. After installing the matching SDK, save the chunks in order and run python example.py or compile and run example.mts. No credentials, model or service are required.

Application walkthrough

Choose Python or TypeScript. Create example.py or example.mts and append that language's chunks in order. Each language forms one complete program; indentation and open blocks carry across chunks.

Import the packages

Import the SDK and standard-library helpers. These are the real package exports used by the executable program.

"""A complete local application: Python chunks on the feature pages use this source."""
import asyncio
import json
from pathlib import Path
import sqlite3
import tempfile

from soterias import Soterias

Create an application and an ordinary database

Soterias observes this existing code; it does not secure these direct writes.

# %% Create an application and an ordinary database
# Soterias observes this existing code; it does not secure these direct writes.
root = Path(tempfile.mkdtemp(prefix="soterias-invoices-"))
database = sqlite3.connect(root / "invoices.sqlite3")
database.execute("CREATE TABLE invoices (id INTEGER PRIMARY KEY, cents INTEGER)")
sdk = Soterias(workflow="invoices", directory=root / "metrics")


def post_invoice(cents):
    row = database.execute("INSERT INTO invoices(cents) VALUES (?)", (cents,))
    database.commit()
    return {"id": row.lastrowid, "cents": cents}

Observe synchronous and asynchronous calls

Repeating the same input still runs the write twice.

# %% Observe synchronous and asynchronous calls
# Repeating the same input still runs the write twice.
first = sdk.invoke(post_invoice, 1500)
second = sdk.invoke(post_invoice, 1500)
assert first["id"] != second["id"]
tracked_post = sdk.track(post_invoice)
assert tracked_post(500)["cents"] == 500


async def review_invoices():
    await asyncio.sleep(0)
    return database.execute("SELECT SUM(cents) FROM invoices").fetchone()[0]


async def finish():
    try:
        total = await sdk.ainvoke(review_invoices)
        assert total == 3500
        tracked_review = sdk.track_async(review_invoices)
        assert await tracked_review() == total
        await sdk.flush()
        summary = sdk.summary()
        assert summary["counts"]["raw"] == 5

Inspect and export local measurements

Each export is explicit; nothing is uploaded.

        # %% Inspect and export local measurements
        # Each export is explicit; nothing is uploaded.
        for name, export in (
            ("runs.json", sdk.metrics.export_json()),
            ("runs.csv", sdk.metrics.export_csv()),
            ("runs.otlp.json", sdk.metrics.export_otlp()),
        ):
            (root / name).write_bytes(export.data)
        print(json.dumps({"status": "PASS", "total_cents": total,
                          "runs": summary["counts"]["raw"], "directory": str(root)}))
    finally:

Close the runtime and database

Wait for pending recordings before leaving the process.

        # %% Close the runtime and database
        # Wait for pending recordings before leaving the process.
        await sdk.close_async()
        database.close()


asyncio.run(finish())

MCP composition

This complete stdio server calls this feature's real SDK operation. Its bounded tool inputs cannot replace host-selected trust, checkpoint or storage paths. Save it as mcp.mjs and configure your MCP client to launch node mcp.mjs. Public data inputs use the same environment settings as the application program. The code is a custom application tool, not a newly claimed built-in Soterias endpoint.

Install the matching adapter with npm install https://docs.soterias.io/downloads/api-preview/soterias-mcp-server-0.1.0-rc.5-api.2.tgz. Its SDK peer dependency shares the same native class identities as your installed preview.

mcp.mjs

Create mcp.mjs and append these chunks in order. Together they form one complete program; indentation carries across chunks.

Step 1

Import the SDK and standard-library helpers. These are the real package exports used by the executable program.

JavaScript
import { Soterias, observe, canonicalize, Metrics, audit, HistoricalPolicyInputs } from '@soterias/core';
import fs from 'node:fs';
import os from 'node:os';
import path from 'node:path';
import { McpServer } from '@modelcontextprotocol/server';
import { serveStdio } from '@modelcontextprotocol/server/stdio';
import * as z from 'zod/v4';

// Trusted startup: paths, trust and credentials are fixed by the host.
const root = fs.mkdtempSync(path.join(os.tmpdir(), 'soterias-mcp-'));
const sdk = new Soterias({workflow: 'invoice-tools', directory: path.join(root, 'metrics')});
const sum = input => input.cents.reduce((a, b) => a + b, 0);

// Agent input is bounded data. It cannot replace host configuration or trust.
const handle = serveStdio(() => {
  const server = new McpServer({name: 'soterias-metrics', version: '1.0.0'});
  server.registerTool('invoice_metrics', {
    description: 'Use metrics through the host-selected native SDK.',
    inputSchema: z.object({cents: z.array(z.number().int().min(0).max(1000000)).max(100)}).strict(),
  }, async input => {
    try {
      const result = await (async () => {
        sdk.invoke(sum, input);
return sdk.summary();

Step 2

Finish outstanding work and release owned resources. Cleanup errors must not silently replace an existing application error.

JavaScript
      })();
      return {content: [{type: 'text', text: JSON.stringify(result, (_key, value) => typeof value === 'bigint' ? value.toString() : value)}]};
    } catch {
      return {isError: true, content: [{type: 'text', text: 'feature_request_refused'}]};
    }
  });
  return server;
}, {maxSubscriptions: 0});
process.stdin.once('end', () => {
  void (async () => { await handle.close(); await sdk.closeAsync(); })();
});

Test the tool over a real MCP connection

Save this client next to mcp.mjs, then run node client.mjs. It starts the host, calls the actual feature tool and closes the connection. It does more than list the tool catalog.

client.mjs

Create client.mjs and append these chunks in order. Together they form one complete program; indentation carries across chunks.

Step 1

Import the SDK and standard-library helpers. These are the real package exports used by the executable program.

JavaScript
import { Client } from '@modelcontextprotocol/client';
import { StdioClientTransport } from '@modelcontextprotocol/client/stdio';
import { fileURLToPath } from 'node:url';

const client = new Client({name: 'feature-example', version: '1.0.0'});
try {
  await client.connect(new StdioClientTransport({
    command: process.execPath,
    args: [fileURLToPath(new URL('./mcp.mjs', import.meta.url))],
    env: Object.fromEntries(Object.entries(process.env).filter(([, value]) => value !== undefined)),
  }));
  const result = await client.callTool({name: 'invoice_metrics', arguments: {cents: [500, 1000]}});
  if (result.isError) throw new Error('Feature request refused');
  console.log(JSON.stringify(result));
} finally { await client.close(); }

Check the result

The application program prints a JSON result and preserves native refusals. For protected operations inspect status and target_observed: transport success does not imply a completed effect. The pure examples contain assertions for their exact expected values. Successful startup is not proof of production enrollment, a physical human review or current authority.

Put it into a workflow

Use the AML walkthrough for protected document reads, reviewed writes and signed outcome checks. Use multi-agent workflows for personal and enterprise orchestration. Keep each agent's proposals separate from operator authority.

Search guides, examples and API reference.