tensorcost — Python SDK

One-line wrapper for TensorCost. Adds fire-and-forget cost observability and optional inference control to OpenAI, Anthropic, AWS Bedrock, and Google Vertex AI clients without changing how your code calls the underlying provider.

Current version: 1.3.0 — period budgets + agent-run span linking (upgrade from 1.2.x if you use daily/monthly caps or Router · Agent runs). See CHANGELOG, SDK compliance, and Migrating from 0.x.

Install

pip install tensorcost

Runtime dependency: httpx only.

Quick start (observe)

The default layer is observe: telemetry only, provider call unchanged.

from openai import OpenAI
from tensorcost import wrap

client = wrap(
    OpenAI(api_key="sk-..."),
    api_key="tc_live_...",
)

resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "hello"}],
)

Anthropic works identically:

from anthropic import Anthropic
from tensorcost import wrap

client = wrap(
    Anthropic(api_key="sk-ant-..."),
    api_key="tc_live_...",
)
client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=128,
    messages=[{"role": "user", "content": "hello"}],
)

Create a long-lived SDK key in the TensorCost console (Setup → Keys). The SDK exchanges it for a short-lived JWT via POST /api/inference-proxy/sdk-token/exchange and refreshes before expiry.

Control layers (max_layer)

v1.0 uses a single ladder for how much TensorCost may intervene on a call:

off → observe → govern → steer → route

A layer is a ceiling, not a mode switch. The effective layer is always:

min(your max_layer, console published layer)

When routing is paused in the console, the published ceiling is capped at govern.

Layer	Provider path	Behavior
off	Direct	No SDK wiring; client returned unchanged
observe	Direct	Telemetry only (default when max_layer is omitted)
govern	Direct	Metadata-only POST /v1/admit before the provider call
steer	Via proxy	Proxy on the path; may steer without full live routing
route	Via proxy	Full applied-mode routing through inference-proxy

from openai import OpenAI
from tensorcost import wrap

client = wrap(
    OpenAI(api_key="sk-..."),
    api_key="tc_live_...",
    max_layer="route",
    proxy_url="https://api.my-instance.tensorcost.com",
)

wrap() raises MissingConfigError when max_layer is steer or route and no proxy_url is available. govern does not require proxy_url — admit uses base_url.

Console-published ceiling

Operators set the tenant ceiling in the console (Policy → Routing) or via PATCH /api/ai/router/layer (published_layer, default observe). The SDK fetches the effective value from GET /api/inference-proxy/v1/sdk-layer:

• 30s cache while the endpoint is reachable
• 250ms cap on a cold fetch — a hung TensorCost cannot block the model call
• Stale-while-revalidate after TTL: serve the last snapshot immediately and refresh in the background
• Decays to govern after 5 minutes unreachable (fail-safe: still run admit, never silently upgrade to route)

Publishing steer or route in the console requires a content-retention data grant. Without a grant, publish returns 409.

Govern admission

At govern and above, the SDK calls POST /api/inference-proxy/v1/admit with metadata only — no prompt or completion bodies:

• 204 + x-tc-decision: layer=govern; action=allow → proceed to the provider
• 403 + x-tc-decision: layer=govern; action=refuse; reason=… → call blocked

Admit is fail-open on transport errors: the provider call proceeds and an admit_unavailable observation is emitted when transport recovers. The admit HTTP call is capped at 500ms.

Compliance profiles (1.2.0+)

Enterprise tenants with feature.compliance_profiles assign frameworks (GDPR, HIPAA, PCI-DSS, EU AI Act, residency) in Policy → Compliance profiles. The SDK loads the compiled snapshot from GET /v1/sdk-layer and enforces it at govern and above.

• team_id on wrap(..., team_id="…") or TENSORCOST_TEAM_ID — narrows to a team overlay; must match a team-bound SDK token when the key is scoped to a team.
• In-process DLP — pattern detectors run locally before the provider; matched secrets never leave your process or hit TensorCost.
• Observations stamp compliance_frameworks and dlp_action (never secret text).
• TensorCostComplianceDeniedError / TensorCostComplianceTeamMismatchError on refuse paths.

See SDK compliance profiles.

Route layer prerequisites

max_layer="route" is necessary but not sufficient for live provider swaps. ADR-0016 gates still apply on the proxy side (LaunchDarkly flag, AI_ROUTING_LIVE_EXECUTION_PERMITTED, non-dry_run policy with rollout, routing not paused). When gated off, traffic pass-throughs to the original provider.

See ADR-0026 and inference-proxy applied mode.

Cold start (prewarm)

At Lambda/container init, the first steer/route call can otherwise pay for JWT exchange and sdk-layer fetch on the hot path. Call prewarm() once at module scope — it shares the same process-wide runtime that wrap() uses, so a later wrap() hits a warm cache.

wrap() also background-warms that shared runtime on first use; explicit prewarm() is recommended when you know traffic is about to arrive.

from openai import OpenAI
from tensorcost import prewarm, wrap

prewarm(
    api_key="tc_live_...",
    base_url="https://api.tensorcost.com",
    max_layer="route",
)

client = wrap(
    OpenAI(api_key="sk-..."),
    api_key="tc_live_...",
    max_layer="route",
    proxy_url="https://...",
)

prewarm() is best-effort: failures are swallowed and the first real call retries the fetches.

Migrating from 0.x

0.x	1.0
applied_mode=True	max_layer="route" (deprecated alias still works with a warning)
Default x-tc-max-layer: route on proxy calls	Default is now observe
Boolean on/off	Full layer ladder + console ceiling

# Before (0.6.x)
wrap(client, applied_mode=True, proxy_url="...")

# After (1.0)
wrap(client, max_layer="route", proxy_url="...")

Recommended rollout: observe → govern → route, raising the console published_layer only after each stage is validated.

Configuration

wrap() resolves: explicit kwargs → environment variables → defaults.

Variable	Purpose	Required
TENSORCOST_API_KEY	Long-lived SDK key from Setup → Keys	Yes
TENSORCOST_BASE_URL	Backend base URL (admit, sdk-layer, token exchange)	No (default https://api.tensorcost.com)
TENSORCOST_PROXY_URL	Inference-proxy base URL	Required when max_layer is steer or route
TENSORCOST_TENANT_ID	Explicit tenant id	No
TENSORCOST_ENVIRONMENT	Environment tag on every observation	No
TENSORCOST_CONNECTION_ID	Provider-connection id	No
TENSORCOST_AGENT_ID / TENSORCOST_WORKFLOW_ID	Agent-run attribution defaults	No
TENSORCOST_CUSTOMER / TENSORCOST_FEATURE	Chargeback tag defaults	No
TC_APPLICATION / TENSORCOST_APPLICATION	Routing-policy application scope (proxy)	No
TC_TAGS / TENSORCOST_TAGS	Comma-separated routing tags (proxy)	No

client = wrap(
    OpenAI(api_key="sk-..."),
    api_key="tc_live_...",
    max_layer="govern",
    application="customer-support",
    tags=["tier:enterprise"],
)

Proxy hardening

Applies when max_layer is govern, steer, or route.

Retries

from tensorcost import wrap, RetryConfig

client = wrap(
    OpenAI(api_key="sk-..."),
    max_layer="route",
    proxy_url="https://...",
    retry=RetryConfig(max_attempts=5, base_delay_ms=500.0, max_delay_ms=30_000.0),
)

5xx, network errors, and 429 are retried with jittered exponential backoff. 4xx (except 429) are never retried.

Timeouts

Proxy timeouts are split so a wedged TCP connection fails open quickly while slow model bodies still get a full budget:

client = wrap(
    OpenAI(api_key="sk-..."),
    max_layer="route",
    proxy_url="https://...",
    headers_timeout_s=2.0,  # wait for response headers (default 2.0)
    timeout_s=60.0,         # non-streaming body read (default 60.0)
    idle_timeout_s=30.0,    # streaming idle (default 30.0)
    fail_open_enabled=True,
    on_lifecycle_event=lambda e: print(e.kind, e.model),
)

Option	Default	TensorCostTimeoutError.kind	On timeout (steer/route)
headers_timeout_s	2s	"headers"	Streaming only (future SSE client)
timeout_s	60s	"total"	TTFB + body for buffered JSON; fail-open when enabled
idle_timeout_s	30s	"idle"	Reserved for future steer SSE client

Non-streaming steer/route waits up to timeout_s for headers (proxy may buffer upstream). stream=True: direct provider when fail-open enabled; MissingConfigError when disabled (1.1.1).

The headers deadline targets wedged proxy connections, not slow upstream models — body read uses timeout_s independently once headers arrive.

Typed errors: TensorCostError, TensorCostNetworkError, TensorCostTimeoutError, TensorCostProxyError, TensorCostQuotaError, TensorCostProviderError, TensorCostRunBudgetExceededError, TensorCostPeriodBudgetExceededError, TensorCostComplianceDeniedError, TensorCostComplianceTeamMismatchError.

Period budgets (1.3.0+)

Daily or monthly caps (tenant default, per-agent, or per-team) are configured in the console under Router · Agent runs. On applied-mode (steer / route) calls the proxy returns 403 with code: "PERIOD_BUDGET_EXCEEDED". The SDK raises TensorCostPeriodBudgetExceededError and does not fall back to the provider.

from tensorcost import TensorCostPeriodBudgetExceededError

try:
    client.chat.completions.create(...)
except TensorCostPeriodBudgetExceededError as err:
    # err.scope, err.period, err.cap_cents, err.spent_cents
    raise

Per-run caps raise TensorCostRunBudgetExceededError (RUN_BUDGET_EXCEEDED).

Govern refusals and proxy decisions appear in x-tc-decision and in observation metadata as decision_action / decision_reason when present.

After consecutive proxy 5xx failures, the circuit opens and requests fail-open to the provider. Set fail_open_enabled=False when direct provider access is unavailable.

Agent-run attribution

client = wrap(
    OpenAI(api_key="sk-..."),
    agent_id="support-bot",
)

task_client = client.with_meta(workflow_id="ticket-48291")
task_client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize this ticket"}],
)

Proxy calls forward x-tc-agent-id / x-tc-workflow-id. with_meta() works on OpenAI and Anthropic wrapped clients; Bedrock and Vertex accept wrap-time defaults only.

customer / feature are separate chargeback dimensions (max 128 chars each).

Span linking (1.3.0+): when an agent framework publishes a model-call link, observations include trace_id and parent_span_id so ai-service nests the model_call under the agent trace and run rollups subtract proxied child spend. Plain wrap() without a link still attributes by agent_id / workflow_id.

Fail-open guarantee

If TensorCost is unreachable, slow, or returns an error, your underlying provider call still completes normally. The SDK logs a warning via logging.getLogger("tensorcost").

Duplicate spend: fail-open after a failed proxy may double-bill if the proxy already reached the provider. Use fail_open_enabled=False for in-VPC sidecar-only steer.

from tensorcost import flush_observations

flush_observations(api_key=..., timeout_s=3.0)  # before Lambda return

See docs/demo/data-path-one-pager.md for the provider matrix.

Exceptions:

• Missing TENSORCOST_API_KEY → MissingConfigError at wrap() time
• Missing proxy_url when max_layer is steer or route → MissingConfigError
• Govern refusals (403 from /admit) → call blocked by design
• Steer run-budget / period-budget refusals → not fail-opened

What gets sent

Metadata only — never prompt or completion content:

• SDK version, provider, model, operation, modality
• Timestamps, token counts, correlation UUID, status
• Environment, agent/workflow ids, chargeback tags when configured
• decision_action / decision_reason when x-tc-decision is present

Proxied requests include x-tc-sdk-version, x-tc-sdk-capabilities, and x-tc-max-layer.

Authentication uses a short-lived JWT from POST /api/inference-proxy/sdk-token/exchange.

Supported providers & operations

Provider	Package	Layers	Operations
OpenAI	openai ≥ 1.0	observe → route	chat, completions, embeddings, responses (+ streaming)
Anthropic	anthropic ≥ 0.20	observe → route	messages (+ streaming)
Azure OpenAI	openai (AzureOpenAI)	observe → route	same as OpenAI (provider: openai)
AWS Bedrock	boto3 bedrock-runtime	observe, govern only	invoke, converse (+ streaming placeholders)
Google Vertex	google-genai (vertexai=True)	observe, govern only	generate_content (+ streaming)

Bedrock and Vertex: signing happens inside the AWS / Google client before the proxy can intercept — max_layer="steer" or "route" raises MissingConfigError.

import boto3
from tensorcost import wrap

bedrock = wrap(
    boto3.client("bedrock-runtime", region_name="us-east-1"),
    api_key="tc_live_...",
    max_layer="govern",  # steer/route not supported
)

Node vs Python parity (1.1)

Feature	Node (@tensorcost/sdk)	Python (tensorcost)
Control layers	✓	✓
Govern /admit	✓	✓
SDK layer fetch (250ms cold cap + SWR)	✓	✓
prewarm() / shared runtime cache	✓	✓
Proxy headers timeout (fail-open)	headersTimeoutMs	headers_timeout_s
Framework adapters	wrapLangChain, etc.	—
maxGrants / grant intersection	✓	not yet
deployment / promptTemplateId	✓	not yet

Development

pip install -e '.[dev]'
pytest

See PUBLISHING.md for the PyPI release runbook.

Further reading

• ADR-0026 — SDK control-layer ceilings
• Inference proxy SDK feature doc
• Applied mode / live routing
• CHANGELOG

License

Apache-2.0. See LICENSE for the full text.