v1.0
Initial release
The first public release of Worksi AI Control: one management loop for AI in your apps, from connect and observe through detect, understand and control to measure, with self‑serve trials and plans.
Connect
- OpenAI‑compatible gateway: chat completions, the Responses API, embeddings and model listing, streaming and non‑streaming.
- Providers: OpenAI, OpenAI‑compatible endpoints such as Azure OpenAI and vLLM, and a deterministic built‑in simulator with fault injection.
- Workspace API keys bound to an environment, with scopes, per-key rate limits, expiry and rotation.
- Python and TypeScript SDKs: runs, steps, tools, outcomes, human approvals, operator control checks, tool policy pre-checks and a direct mode.
- OpenTelemetry ingestion over OTLP/HTTP (JSON and protobuf) with GenAI semantic‑convention mapping.
- Worksi MCP server with agent tools and operator tools, and an MCP gateway that governs any Streamable HTTP MCP server.
- Personal access tokens for the REST API and MCP operator tools.
- Guided onboarding from provider to first trace, with time to first value.
Observe
- Live activity stream, trace explorer with waterfall and span inspector, and run views.
- Overview, performance and cost dashboards.
- AI inventory of agents, models, workflows, tools and data sources, with owners, teams and risk tiers.
- Auto-discovery of unregistered agents, models, workflows and tools (shadow AI).
- AI Graph of dependencies with impact analysis.
- End‑user analytics: requests, tokens, cost and errors per end user and plan.
Detect & understand
- Eleven detectors: agent loop, repeated tool call, timeout, provider failure, tool failure, token anomaly, cost anomaly, stuck workflow, policy violation, error rate and budget threshold.
- Deduplicated incidents with severity, response-time targets, timeline and an operations inbox.
- Evidence packs with the failure location and citeable evidence items.
- Deterministic-first diagnosis with optional OpenAI-assisted reasoning, citation checks and confidence scores.
- Similar-incident search, versioned runbooks with approval, and recommendations.
- Alerts to email, Slack and signed webhooks.
Control & govern
- Pause and resume agents and workflows, stop runs, retry through signed webhooks, switch model and escalate, enforced at the gateway within seconds.
- Policies for model allow and deny lists, spend, tokens, rate, model access, tools, data capture and end‑user quotas, in monitor or enforce mode, with simulation and revision history.
- Human approvals from the console, the SDKs and MCP, with four‑eyes for high-risk actions.
- End‑user plans: enforce your own customers’ plans per request with the
X-AIW-End-UserandX-AIW-End-User-Planheaders. - Six roles with a permission matrix, capture modes with redaction, and a hash-chained audit log.
Measure
- Business outcomes, outcome definitions and baselines.
- ROI with value versus cost, net value, value multiple and a cost model that includes platform and human-review cost.
- Cost allocation with a chargeback CSV, month‑end forecast, budget burn-down and provider invoice reconciliation.
- CSV and JSON exports of audit, traces, incidents, cost and ROI.
Plans & accounts
- Starter, Pro, Business and Enterprise plans; bring your own model provider key with no token mark‑up.
- A free trial with Pro features for every new organization, no credit card required.
- Plan & billing page with usage meters, trial countdown and upgrade requests.
- Email verification, trial reminders and usage warnings at 80% and 100% of plan quotas.