--- title: "Use AI for more of your work. *Keep your standards.*" description: "Connect your repo and EvalGate finds the AI in your code: model calls, prompts, agents, and tools. Test what they do, catch regressions before release, and keep the proof." canonical: "https://www.evalgate.com/" last-updated: "2026-10-21" --- # Use AI for more of your work. *Keep your standards.* Works with the AI you already use Connect your repo and EvalGate finds the AI in your code: model calls, prompts, agents, and tools. Test what they do, catch regressions before release, and keep the proof. Platform: Beta. Agent Skills: Experimental. A passing check only supports the behavior actually checked. [Feature status](https://www.evalgate.com/docs/platform/feature-status) [Run a local gate →](https://www.evalgate.com/docs/quickstart-typescript) [See how it works](https://www.evalgate.com/#story) Run one credential-free local gate, break a reviewed case, and see the build block before connecting a hosted workspace. Example / Refund assistant ## The AI refunded $240 before it should have. > “Absolutely. I found order #1842 and issued the $240 refund to your Visa.” What happened - Verified It found the right customer and order. - Missed It never checked whether the refund was allowed. - Missed It didn't get the required approval. - Early It issued the money anyway. With a blocking rule in place, EvalGate stops this result and shows you why. Illustrative example — not a live evaluation. [Run the public refund demo](https://github.com/evalgate/refund-agent-demo) See the technical evidence ### Customer + order verified PASS ### Refund policy evaluated MISSING ### Required approval MISSING ### Refund executed TOO EARLY ### Gate decision BLOCK ### Evidence state INCOMPLETE This illustrative trace verifies the customer and order, then records a $240 refund. It contains no refund-policy evaluation or required approval before that action. The answer alone would not prove what happened. The sample release is blocked: the refund happened before the required checks were evidenced. Missing action evidence would be not measured, never a pass. No real payment or release was changed. Other rules can warn, flag uncertainty, or ask for review instead of stopping. You choose the boundary for each one. Works with the AI you already use - Cursor - Claude - ChatGPT - Codex - GitHub Copilot - AWS Bedrock - Databricks - OpenRouter [Building AI? SDK · MCP · API · CI/CD →](https://www.evalgate.com/developers) Why spot checks miss it ## A good answer can still hide the wrong action. - ### Reading the answer Chat logs show what the AI said. They don't show whether it checked the policy or got approval before it acted. - ### Testing prompts on their own Prompt tests score text. They don't see the tool calls, approvals, and order of steps in your real application. - ### Hearing it from customers Dashboards and incident reviews tell you after release. By then the refund has already gone out. How it works ## Correct your AI once. EvalGate checks for it next time. 1. 01 ### Work normally Use Cursor, Claude, Copilot, ChatGPT, or your own agent. Whatever helps you get the job done. 2. 02 ### Correct what matters “Don't confuse pipeline with revenue.” “Don't change unrelated notebook cells.” “Never refund without approval.” 3. 03 ### Keep the correction Accepted corrections become reusable checks that apply the next time that kind of work comes up. Each one can warn you, flag uncertainty, ask for review, or stop the result. You choose. That's an eval. You don't have to learn evals before they're useful. A saved check is not retraining a model, and it can't promise a mistake never comes back. Reviewing results, approving a new standard, and publishing changes each need someone's sign-off. [Explore the public playground ↗](https://www.evalgate.com/playground) What EvalGate does ## From the first check to the release decision. - 01 ### Find Connect a repository and EvalGate finds the AI in your code: model calls, prompts, agents, and tools. - 02 ### Test Check what your AI says and what it does, including the actions and approvals that have to happen first. - 03 ### Gate Run the same checks in GitHub Actions or GitLab CI, so a change that breaks them is caught before release. - 04 ### Prove Keep the evidence with every decision: what was required, what happened, and what was not checked. Where you already work ## Add EvalGate where you already work. - Cursor Guided - Claude Guided - ChatGPT Guided - Codex Guided - GitHub Copilot Guided - Claude Code Guided - Your application Connect your model provider [Start free and connect →](https://www.evalgate.com/auth/sign-up) Pick an agent to check your AI work: **Claude Code** to dig into a problem, **Codex** to review a change, or **Cursor** for a quick check in your editor. All three use the same EvalGate Skills. [Find EvalGate MCP on Smithery ↗](https://smithery.ai/servers/evalgate/evalgate) [Agent quick starts and verification →](https://www.evalgate.com/developers) Guided means you open a link or copy one thing into the tool, then approve it there. The EvalGate connection for AI apps is in preview. Building with EvalGate? Developer tools Public documentation, OpenAPI, SDKs, MCP interfaces, and installable skills give agents a clear path in. Private workspace data still requires authorization. # Add EvalGate’s agent skills ``` npx skills add evalgate/skills ``` # Start a local evaluation project ``` npx @evalgate/sdk init --local ``` [OpenAPI ↗](https://www.evalgate.com/openapi.json) [Find EvalGate MCP on Smithery ↗](https://smithery.ai/servers/evalgate/evalgate) [Public documentation MCP ↗](https://www.evalgate.com/api/mcp/docs) [Skill source ↗](https://github.com/evalgate/skills) [Agent setup ↗](https://www.evalgate.com/agent/setup.md) [llms.txt ↗](https://www.evalgate.com/llms.txt) [Authentication ↗](https://www.evalgate.com/auth.md) [API guide ↗](https://www.evalgate.com/openapi.json.md) [MCP configuration ↗](https://www.evalgate.com/mcp.json) [Agent Skill index ↗](https://www.evalgate.com/.well-known/agent-skills/index.json) [Pricing ↗](https://www.evalgate.com/pricing) [Developer resources ↗](https://www.evalgate.com/developers) Platform: Beta. Agent Skills: Experimental. MCP: Experimental, and usable today with a scoped key. Follow the [MCP connection guide](https://www.evalgate.com/docs/platform/mcp-integration) for client-specific configuration and access requirements. Installing a Skill does not grant access to data, approve spending, or switch on application controls. ### Start free. No surprises later. Your first 10,000 test-case results each month are free. After that, it's $1 per 1,000. Your AI model costs stay on your own provider bill. [See plans and billing details →](https://www.evalgate.com/pricing) One test case checked is one result, pass or fail. Checking it several ways still counts once. An administrator approves paid checkout and changes to your organization result budget. Review the active limit in Billing before large jobs. Model-provider charges are separate. Where to next ## Pick the fastest way in. - ### Try it without an account Open the public playground and see how a check works. Nothing to install and no sign-up. [Open the playground →](https://www.evalgate.com/playground) - ### Start in your editor Add the EvalGate Skills to Claude Code, Codex, or Cursor and run a local check against your own code. [Read the quickstart →](https://www.evalgate.com/docs/quickstart) - ### Pricing that starts at zero Your first 10,000 test-case results each month are free. No card until you choose paid usage. [See pricing →](https://www.evalgate.com/pricing) EvalGate ## Connect a repo. See where your AI stands. [Run a local gate →](https://www.evalgate.com/docs/quickstart-typescript) [Read the docs](https://www.evalgate.com/docs/) Free to start. Add a card only when you choose paid usage. Before you start: common questions ### Do I need to be technical to use EvalGate? You do not need to write code to understand a test result or discuss the standards your AI should meet. Connecting an application and configuring tests may require an engineer or a coding agent. ### Does EvalGate automatically fix or retrain my AI? No. Finding a failure, fixing it, approving the fix, and training a model are different actions. EvalGate helps your team evaluate behavior and retain evidence; the supported workflow and its permissions determine what can happen next. ### What does a passing result mean? It means the tested example met the configured checks. It does not prove that every risk was tested or that future behavior is guaranteed. Missing required evidence is not a pass. ### What counts as one billed result? One saved test-case outcome counts as one result, whether it passes or fails. Running several scorers on that outcome does not multiply the result count. AI model usage is billed separately by your model provider. ### Can an agent buy EvalGate usage on its own? An authorized administrator sets up paid usage through hosted checkout. Do not treat that handoff as a fully autonomous payment API. Paid result admission uses an administrator-owned monthly result cap. Review separate provider estimates before authorizing model calls. ### Which integrations does EvalGate support? EvalGate supports GitHub Actions, a GitLab CI recipe, OpenTelemetry/OpenInference OTLP, LangChain, CrewAI, AutoGen, Vercel AI SDK, signed webhooks, and Slack delivery. Dedicated CircleCI, Teams, Jira, PagerDuty, and LlamaIndex integrations are not currently advertised. ### Does EvalGate support shadow releases? Deployable Assets can run a non-serving shadow candidate for requests routed through EvalGate's governed invocation endpoint. Shadow failures become evidence without replacing a valid serving response; arbitrary bypass traffic is not mirrored. © 2026 EvalGate [Documentation](https://www.evalgate.com/docs/) [Pricing](https://www.evalgate.com/pricing) [Contact](https://www.evalgate.com/contact) [hello@evalgate.com](mailto:hello@evalgate.com) [LinkedIn](https://linkedin.com/company/evalgate/) [Security](https://www.evalgate.com/security) [Privacy](https://www.evalgate.com/privacy) [Terms](https://www.evalgate.com/terms) Analytics choices ## Capabilities already available - GitHub Actions workflow generation and annotations, plus a documented GitLab CI recipe. - Signed outbound webhooks, Slack App delivery when configured, and Slack Incoming Webhook fallback. - OpenTelemetry/OpenInference-compatible OTLP trace ingestion and export. - Reviewed synthetic and red-team case generation before production incidents. - Immutable Prompt Hub versions, baseline/candidate experiments, and protected-environment rollback. - Non-serving shadow comparison for traffic routed through EvalGate Deployable Assets. ## Billing honesty for agents Billing is organization-scoped through Autumn with Stripe-backed hosted checkout. An authorized administrator completes paid-setup handoffs. Do not treat a checkout URL as a fully autonomous payment API. Paid result admission has a default 10,000-result UTC-month hard budget; an administrator must approve increases. Provider charges are separate. EvalGate does not currently support Machine Payments Protocol settlement. Read the [evidence-backed capability map](https://www.evalgate.com/capabilities.md) for exact entry points and current limits.