For teams shipping AI agents, chat and copilots

v1.0 · 25 September 2026 · Changelog

Every AI call,
under control.

Change one base URL. Every model call your product makes is traced, priced and checked against your policies, and failures arrive as incidents with evidence.

Start free trial Read the quickstart

0.82 ms p95 gateway overhead, measured — how we measured gateway overhead
3‑day trial with Pro features, no card
No provider key needed to try: built‑in simulator
No mark‑up on model tokens: your provider key

01Your code

import osfrom openai import OpenAI client = OpenAI(    base_url="https://api.openai.com/v1",    base_url="https://gateway.worksi.ai/v1",    api_key=os.environ["OPENAI_API_KEY"],    api_key=os.environ["AIW_API_KEY"],) # every call is now traced and pricedreply = client.chat.completions.create(    model="gpt-5-mini", messages=history,)

import OpenAI from "openai" const client = new OpenAI({  baseURL: "https://api.openai.com/v1",  baseURL: "https://gateway.worksi.ai/v1",  apiKey: process.env.OPENAI_API_KEY,  apiKey: process.env.AIW_API_KEY,}) // every call is now traced and pricedconst reply = await client.chat  .completions.create({    model: "gpt-5-mini", messages,  })

02Gateway response443 ms

200 OKPOST /v1/chat/completions

X-AIW-Request-Id: req_af64…fa69
X-AIW-Trace-Id: 2bef9e35…6da1
X-AIW-Span-Id: 7cdbe47128f3ca27
X-AIW-Provider: simulator
X-AIW-Model: gpt-5-mini
X-AIW-Cost-USD: 0.00040925

03Span recorded

llm_callchat.completions

agent: triage-agent
model: gpt-5-mini
tokens: 213 → 178
latency: 443 ms
cost: $0.00041
status: ok

key
policies
rate limit

Header names are the gateway’s contract. Values come from trace 2bef9e35…6da1 in the built‑in demo organization.

Failures show up as complaints. A failing tool, a looping agent or a cost spike reaches customers, or the invoice, before it reaches an engineer.

The management loop

From the first request to measured value.

Six stages, one product. Each stage feeds the next, and the loop closes when what you measure changes what you control.

02Observe

See every run, not just every request.

Traces follow a run through agents, steps, model calls and tools, with latency, tokens and cost on every span.

1. The run: support‑ticket‑resolution, 9 spans on a 1.5 s timeline.
2. The first failure, flagged: lookup_order failed, and a line marks the moment.
3. Latency and cost on every span: 443 ms and $0.00041 for this model call.

Explore observability

Trace waterfall for support-ticket-resolution: 9 spans on a 1.5 s timeline with the failing lookup_order call flagged — Traces · support-ticket-resolution · sample data from the built‑in demo organization

03Detect04Understand

See what broke, where, and why.

Eleven detectors watch your traffic. Each incident arrives with the failing span, numbered evidence and a diagnosis that has to cite it.

INC‑5, reconstructed

run
triage‑agent starts a support‑ticket‑resolution run in production and hands the research step to resolver‑agent.
tool
lookup_order calls orders-db: 503 Service Unavailable (connection pool exhausted).
×3
Third failure within 10 minutes: the tool_failure detector’s threshold.
INC‑5
Incident opened. Critical. Later failures are counted as occurrences (44 so far), not opened again.
evidence
Failure location and evidence: failing span 3eb91d0f65c815a4, the error message and an 18% tool error rate.
diagnosis
78%, high confidence. Root cause: orders-db returns 503. It cites ev‑1, ev‑2, ev‑3 and ev‑6, and it is marked LLM‑assisted.
next step
Suggested: pause the support‑ticket‑resolution workflow until orders-db recovers. The pause is enforced at the gateway: 423 workflow_paused.

In our benchmark, a dependency answering 503 opened an incident 1.3 s after the failing call. How we measured — how we measured time to detect

Explore incidents & diagnosis

Incident INC‑5, What happened: lookup_order failed 3× in 10 minutes, failure location and the upstream 503 error — Incident · Overview tab · sample data from the built‑in demo organization

Diagnosis for INC‑5: root cause, 78% high confidence, LLM‑assisted, and the cited evidence — Incident · Diagnosis tab · sample data from the built‑in demo organization

05Control

Monitor first. Enforce when you’re sure.

Simulate a policy against your recorded traffic and see what it would have blocked. Then enforce it at the gateway, in the SDKs and in the MCP gateway.

Explore governance

{  "name": "Production model allowlist",  "type": "model_allowlist",  "scope_type": "environment",  "scope_id": "production",  "rules": {    "models": [      "gpt-5",      "gpt-5-mini",      "gpt-5-nano"    ]  },  "enforcement": "monitor"}

02App calls model "gpt-4o"

200 OK

HTTP/1.1 200 OKX-AIW-Model: gpt-4o # nothing is blocked; the span records:aiw.policy_id = <id of this policy>aiw.violation = "model_allowlist"

HTTP/1.1 403 Forbidden{  "error": {    "message": "model \"gpt-4o\" is not in the allowed models of policy \"Production model allowlist\"",    "type": "permission_error",    "code": "model_not_allowed",    "param": null,    "aiw_request_id": "req_…"  }}

Example policy. The response bodies are the gateway’s contract.

Approvals: Risky tools and recovery actions wait for a person. Four‑eyes approval on Business.
End‑user plans: Send X-AIW-End-User and X-AIW-End-User-Plan. Over the limit, the gateway answers 429 end_user_quota_exceeded.
Pause and stop: Pause an agent or stop a run from the console. The gateway answers 423 agent_paused or 409 run_stopped.

Approvals. Risky tools and recovery actions wait for a person. Four‑eyes approval on Business.
End‑user plans. Send X-AIW-End-User and X-AIW-End-User-Plan. Over the limit, the gateway answers 429 end_user_quota_exceeded.
Pause and stop. Pause an agent or stop a run from the console. The gateway answers 423 agent_paused or 409 run_stopped.

Benchmark · 24 September 2026

Measured, not promised.

Results from our own test runs, with the conditions attached. Your numbers depend on your traffic and your provider.

0.82ms

Gateway overhead, p95

Non-streaming chat completions ramped to 200 requests/s; 55,806 requests, 0 failures.^{¹note 1}

Telemetry never sits on the request path: if the telemetry pipeline is unavailable, the gateway buffers spans and your requests continue.

Gateway and ingestion
Metric	Result	Conditions
Streaming time to first byte, p95	1.09ms	Streaming chat completions ramped to 100 streams/s; 27,906 streams, 0 failures.
Span ingestion on one replica	2,000spans/s	360,000 spans in batches of 100; p95 request time 6.9 ms, 0 rejected spans.

8of 8

Scripted failure scenarios detected

0 false alarms in 6 healthy runs in the same round. Median time to detect: 1.1 s.^{²note 2}

Provider outage, provider_failure: 0.13 s
Policy violation, policy_violation: 0.54 s
Repeated tool call, repeated_tool_call: 0.72 s
Cost spike, token_anomaly: 0.97 s
Tool failure, tool_failure: 1.3 s
Agent loop, agent_loop: 1.9 s
Provider timeout, timeout: 5.0 s
Stuck workflow, stuck_workflow, 60 s threshold: 63.2 s

Time from the failing call to an open incident. 8 scripted failures, log scale.

Notes 1 and 2: Internal benchmark, 24 September 2026. Single-replica Docker Compose stack on a 12 vCPU / 16 GB Linux host (gateway limited to 1.5 CPU / 512 MiB), load generated with k6. The built‑in simulator provider answered model calls with a pinned 1 ms latency so the gateway’s own overhead could be isolated. Not covered: real provider latency and errors (simulator only), multiple replicas, cloud networking, TLS termination and managed data stores, multi-hour soak tests. How we measured — how we measured the gateway and detection benchmarks

Pay for the platform, not the tokens.

Bring your own provider key: your provider bills you for tokens, and Worksi never marks them up.

Compare all plans

Starter

Indie developers & B2C apps

$9/ month

or $90 / year, two months free

Seats: 2
Gateway requestsRequests: 100k / mo
Spans: 1M / mo
Retention: 14 days

Start free trial

Adds approvals, AI Graph & ROI

Pro

Startups & small teams

$29/ month

or $290 / year, two months free

Seats: 5
Gateway requestsRequests: 1M / mo
Spans: 10M / mo
Retention: 30 days

Start free trial

Business

Growing companies

$99/ month

or $990 / year, two months free

Seats: 25
Gateway requestsRequests: 10M / mo
Spans: 100M / mo
Retention: 90 days

Start free trial

Enterprise

Large / regulated orgs

Custom

Limits and terms set with you

Seats: Custom
Gateway requestsRequests: Custom
Spans: Custom
Retention: 365+ days

Contact sales

Starter
Indie developers & B2C apps
$9 / mo
2 seats · 100k requests / mo · 14 days
Adds approvals, AI Graph & ROI
Pro
Startups & small teams
$29 / mo
5 seats · 1M requests / mo · 30 days
Business
Growing companies
$99 / mo
25 seats · 10M requests / mo · 90 days
Enterprise
Large / regulated orgs
Custom
Custom limits · 365+ days

Start free trial Compare all plans

Prices in USD. We don’t take card payments online yet: when your trial ends, choose a plan with Request upgrade in the console or by email, and we activate it. Your data and setup stay. Pricing questions

Trace your first AI call in minutes.

Start free trial Talk to us

3‑day trial with Pro features, no card.
The built‑in simulator answers until you add a provider key.
Metadata‑only capture per environment, if you need it.

Terminal

curl -i \  https://gateway.worksi.ai/v1/chat/completions \  -H "Authorization: Bearer $AIW_API_KEY" \  -H "Content-Type: application/json" \  -d '{"model": "gpt-5-mini",       "messages": [{"role": "user",                     "content": "Hello"}]}' # the response headers include# X-AIW-Trace-Id and X-AIW-Cost-USD

Every AI call,under control.

From the first request to measured value.

See every run, not just every request.

See what broke, where, and why.

Monitor first. Enforce when you’re sure.

Measured, not promised.

Pay for the platform, not the tokens.

Starter

Pro

Business

Enterprise

Starter

Pro

Business

Enterprise

Trace your first AI call in minutes.

Every AI call,
under control.