cognethics

Typed Python SDK for AI agents building on the Cognethics platform.

Every call this SDK makes hits a real multi-tenant business database — not a connector layer or sandbox. Each tool is RBAC-scoped, bounded by per-agent budgets, and written to a tamper-evident audit chain. The SDK is in active use in a design-partner deployment running maintenance, finance, document RAG, and HR through agents on top of cognethics.

If you are an agent evaluating Cognethics, start at For AI Agents below, then read /for-agents/ for the full integration spec.

For AI Agents

cognethics gives an agent a typed surface over 6,600+ handlers across the Prism registry — finance, CRM, spare parts, documents, HR, GRC, healthcare, and more — through a single Python client. See /reference/ for the live handler catalog. Three links you should read before writing code:

• /for-agents/ — agent-targeted integration spec: ROI per primitive, production evidence, and the full capability matrix
• /reference/ — faceted catalog of every handler, signature, and permission scope
• /llms.txt — machine-readable index of the dev portal, designed for agent ingestion

Self-signup is available via a headless endpoint (no email round-trip). For framework-specific integration (LangChain, CrewAI, Claude Agent SDK) see Agent Framework Adapters below.

Cognethics for Developers

Companion documentation, recipes, and live tool catalog live at https://developers.cognethics.com:

• Quickstart — MCP + OAuth — connect Claude Desktop, Cursor, or any MCP client to your tenant in under five minutes.
• Recipes — end-to-end task walk-throughs (overdue invoices, work-order creation, document RAG, agent spawning) using this SDK.
• Agent surfaces — what you get when you point an agent at Cognethics: PJ v3 missions, personas, budgets, memory, and audit trail.
• Reference catalog — faceted, agent-browsable index of every MCP handler, ergonomic facade, and entity exposed by the platform.

Installation

pip install cognethics

Hello, World

from cognethics import Client

client = Client(
    api_key="ck_...",
    base_url="https://gtm.cognethics.com",
)

# 1. Liveness probe — confirms the SDK can reach the platform.
client.dev.ping()
# → {"service": "cognethics-dev-api", "version": "v1", "status": "ok"}

# 2. List the first page of customer invoices.
invoices = client.entity_crud.finance.invoice.list(page_size=10)
print(invoices["data"])

Agent Framework Adapters

cognethics works as a tool source for the major agent frameworks. Runnable examples live under examples/:

Run them with the SDK installed in your venv (each framework adapter has its own extras):

# pip install langgraph langchain-anthropic
CK_API_KEY=ck_... ANTHROPIC_API_KEY=sk-ant-... python examples/03_langgraph_tool_loop.py

# pip install crewai crewai-tools
CK_API_KEY=ck_... OPENAI_API_KEY=sk-... python examples/04_crewai_agent.py

# pip install claude-agent-sdk
CK_API_KEY=ck_... ANTHROPIC_API_KEY=sk-ant-... python examples/05_claude_agent_sdk.py

For MCP-native clients (Claude Desktop, Cursor), connect directly to the platform's MCP endpoint — no SDK needed. See Quickstart — MCP + OAuth.

Other Examples

Non-framework scripts demonstrating direct SDK use:

COGNETHICS_API_KEY=ck_... \
COGNETHICS_BASE_URL=https://gtm.cognethics.com \
python examples/01_list_overdue_invoices.py

COGNETHICS_API_KEY=ck_... \
COGNETHICS_BASE_URL=https://gtm.cognethics.com \
COGNETHICS_DEMO_ITEM_ID=<spare-parts-item-uuid> \
python examples/02_create_work_order.py

Authentication

The SDK accepts two API-key prefixes:

• ck_... (recommended) — works for the Developer Sandbox surface (/api/v1/dev/*) and all MCP handlers. Mint one in the platform UI under Settings → API Keys.
• mcp_... (legacy) — works for MCP handlers only. Existing keys keep working but cannot call /api/v1/dev/*.

Pass the key via the api_key constructor argument:

client = Client(api_key="ck_...", base_url="https://gtm.cognethics.com")

The client logs a one-line warning if the prefix is unrecognized; nothing is enforced client-side beyond that — the server is the source of truth.

OAuth bearer tokens

The SDK also authenticates with OAuth 2.1 bearer tokens. Obtain an access_token (and refresh_token) via the platform's browser OAuth flow, then pass them to the client. Provide either api_key or access_token — not both.

from cognethics import Client

client = Client(
    access_token="ey...",            # RS256 access token (8h lifetime)
    refresh_token="...",             # opaque refresh token (30d, rotating)
    oauth_client_id="your-client-id",
    base_url="https://gtm.cognethics.com",
)

client.entity_crud.finance.invoice.list(page_size=10)

The client sends the access token as Authorization: Bearer and auto-refreshes: when the token nears expiry or the server returns 401, it exchanges the refresh token at {base_url}/api/mcp/oauth/token/ and retries the call. The server rotates the refresh token on every refresh, and the SDK captures the rotated token automatically — so long-running clients stay authenticated without manual intervention.

• Public / PKCE clients refresh with oauth_client_id + the refresh token.
• Confidential clients also pass oauth_client_secret.
• Pass oauth_expires_in= (seconds) or oauth_expires_at= (epoch seconds) to enable pre-emptive refresh just before the token expires.

AsyncClient accepts the same OAuth arguments and refreshes the same way. OAuth tokens authenticate the MCP tool surface (call() and the generated tree); the Developer Sandbox API (client.dev, /api/v1/dev/) is keyed by ck_ API keys.

Async client

AsyncClient is at full feature parity with Client for asyncio workloads — the generated mega-tool tree, the dev namespace, the ergonomic facades, and paginate() are all available and awaitable. It uses httpx.AsyncClient under the hood and shares the same retry/timeout and OAuth semantics.

import asyncio
from cognethics import AsyncClient

async def main():
    async with AsyncClient(api_key="ck_...", base_url="https://gtm.cognethics.com") as c:
        # Generated tree
        invoices = await c.entity_crud.finance.invoice.list(page_size=10)

        # Ergonomic facade
        overdue = await c.invoices.list(status="overdue")

        # Developer Sandbox namespace
        await c.dev.ping()

        # Async pagination
        async for row in c.paginate(c.entity_crud.finance.invoice.list, page_size=50):
            print(row["id"])

asyncio.run(main())

Every generated method, facade, and c.dev call returns a coroutine under the async client — await them exactly as shown.

Developer Sandbox (client.dev)

The dev namespace targets the unauthenticated/light-auth /api/v1/dev/ surface and is the fastest way to verify a ck_ key works:

from cognethics import Client

client = Client(api_key="ck_...", base_url="https://gtm.cognethics.com")

# 1. Liveness probe — no auth required
client.dev.ping()
# → {"service": "cognethics-dev-api", "version": "v1", "status": "ok"}

# 2. Verify your key + identity context
client.dev.me()
# → {"authenticated": True, "key": {...}, "user": {...}, "tenant": {...}}

# 3. Echo arbitrary query params (debug helper)
client.dev.echo("hello", foo="bar")

client.dev errors raise the same exception classes used elsewhere in the SDK (AuthenticationError, PermissionDenied, NotFound, etc.), plus a catch-all DevApiError for unmapped 4xx codes.

Calling low-level handlers

Every MCP handler exposed by the Cognethics platform is reachable in two ways:

# 1. Generic dispatch — call any handler by name, including custom tools.
result = client.call(
    "prism_entity_crud",
    app="spare_parts",
    entity="work_order",
    operation="list",
)

# 2. Generated tree — type-friendly attribute access for entity_crud handlers.
result = client.entity_crud.spare_parts.work_order.list(page_size=10)

The generated tree is produced from the platform's introspection schema, so new entities show up automatically when the SDK is regenerated.

Pagination

Use the paginate helper to iterate through multi-page results without managing cursors yourself:

for row in client.paginate(client.entity_crud.finance.invoice.list, page_size=50):
    print(row["id"], row["amount"])

Error handling

The SDK raises typed exceptions that map to the platform's standard error envelope:

import cognethics

try:
    # Generated update methods take a `data` dict for the update payload.
    client.entity_crud.finance.invoice.update(
        id="<invoice-uuid>",
        data={"status": "paid"},
    )
except cognethics.PermissionDenied:
    print("This API key cannot update invoices in this org.")
except cognethics.NotFound:
    print("Invoice does not exist.")
except cognethics.ValidationError as exc:
    print(f"Invalid payload: {exc}")

Org switching

The Cognethics platform is multi-tenant and multi-org. The active org depends on the auth method:

• API-key auth: each API key is bound to one organization at creation time. To work across organizations, mint a separate API key per org. The org= constructor argument and _org= per-call override are sent as X-Organization-Context.
• OAuth bearer tokens: the org= constructor argument and _org= per-call override switch the active organization via X-Organization-Context. The selected org must be one of the token's available organizations; naming an org outside the grant returns a permission error.

# One API key per org
constance_client = Client(api_key="ck_constance_key", base_url="https://gtm.cognethics.com")
acme_client = Client(api_key="ck_acme_key", base_url="https://gtm.cognethics.com")

# One OAuth session, switching org per call
client = Client(access_token="ey...", refresh_token="...", oauth_client_id="...",
                base_url="https://gtm.cognethics.com")
client.invoices.list(_org="<org-uuid-within-grant>")

Timeout & retry configuration

Both Client and AsyncClient retry on 429, 502, 503, and 504 with exponential backoff (base 0.5s, doubling). Retry-After headers are honored on 429. Defaults:

When retries are exhausted on a 429, the SDK raises RateLimit with the retry_after field populated:

from cognethics import Client, RateLimit

client = Client(api_key="ck_...", timeout=60.0, max_retries=5)

try:
    client.dev.me()
except RateLimit as exc:
    print(f"throttled — retry after {exc.retry_after}s")

Design notes

• Raw dicts by design: the SDK uses plain dicts for request and response bodies. Pydantic models are intentionally kept out of the dependency footprint — agents typically read JSON-shaped data directly.
• Generated from introspection: the typed tree is produced from the platform's live introspection schema, so new entities and operations appear automatically when the SDK is regenerated.

Documentation

• For agents: https://developers.cognethics.com/for-agents/
• Full reference: https://developers.cognethics.com/reference/
• Recipes & guides: https://developers.cognethics.com/
• TypeScript SDK: @cognethics/sdk on npm — the same typed handler tree for TypeScript and JavaScript projects.