> ## Documentation Index > Fetch the complete documentation index at: https://sandtable.dev/llms.txt > Use this file to discover all available pages before exploring further. > ## Agent Instructions > sandtable is an early-alpha, open-source CLI (TypeScript, MIT): https://github.com/sandtable-ai/sandtable. Run it as `npx sandtable-ai `, never with the bare name: npm's `sandtable` is an unrelated package. If a flag here disagrees with `npx sandtable-ai --help`, trust the help output. > Grading is deterministic: pass/fail comes from the world's end state, invariants and the tool-call log, never from an LLM judge. > To draft a domain for a user, read /guides/init and /reference/spec, then run `npx sandtable-ai validate`, `lint` and `review` before any model-backed run. Keep all example data fictional. > In Claude Code, the sandtable plugin teaches these steps as skills: `/plugin marketplace add sandtable-ai/sandtable`, then `/plugin install sandtable@sandtable`. > Every page is available as Markdown by appending .md to its URL. /llms.txt lists every page, /llms-full.txt contains the whole site, and https://sandtable.dev/mcp is a search MCP server for these docs. # sandtable vs LangWatch Scenario: declared world or test code > LangWatch Scenario writes agent tests in code with a judge agent; sandtable grades a declared world deterministically. What each does best, and when to use it. **[LangWatch Scenario](https://langwatch.ai/scenario/)** ([Apache-2.0](https://github.com/langwatch/scenario/blob/main/LICENSE)) is a library for testing agents through simulated conversations. You write each scenario in code: a description, your agent behind a [one-method adapter](https://langwatch.ai/scenario/agent-integration), a user simulator agent, and a judge agent with criteria. Scenarios run inside your test runner, such as pytest, Vitest or Jest ([test runner integration](https://langwatch.ai/scenario/basics/test-runner-integration)), and its docs list SDKs for Python, TypeScript and Go. It pairs with the [LangWatch platform](https://langwatch.ai/docs/agent-testing/overview), which shows runs as they happen and can also write and run scenarios without code. **sandtable** declares a stateful world, the customer's API as tools over it, and the rules the agent must never break. It then grades every trial deterministically on the end state and the tool-call log. ## Side by side | | sandtable | LangWatch Scenario | | - | - | - | | A test is | a task in YAML: hidden goal, persona, expected end state | code in your test suite, or a scenario in the LangWatch platform | | Stateful world | declarative tables, fresh copy per trial | bring your own: your agent's real tools, or mocks you write ([mocking guide](https://langwatch.ai/scenario/testing-guides/mocks)) | | Pass/fail | end state, invariants, required and forbidden tools; no judge | the judge agent's criteria, plus assertions you write as script steps ([scripted simulations](https://langwatch.ai/scenario/basics/scripted-simulations)) | | Rules on every tool call | checked before each call runs | assertions on tool calls that you write, run after the agent's turn ([tool calling](https://langwatch.ai/scenario/testing-guides/tool-calling)) | | Repeated trials | `-k` trials per task, reported as pass^k | the SDK runs a scenario once; the platform repeats each scenario 1 to 5 times and reports a pass rate ([run plans](https://langwatch.ai/docs/agent-testing/run-plans)) | | Drafting tests | `init` drafts the world, tools, invariants and tasks from an OpenAPI spec, an SOP and transcripts | the platform's assistant writes a scenario from text you paste, such as a ticket or a transcript ([scenarios](https://langwatch.ai/docs/agent-testing/scenarios)) | | Qualitative criteria (tone, empathy, wording) | not yet: soft checks are planned ([#16](https://github.com/sandtable-ai/sandtable/issues/16)) | yes, through the judge | | Voice agents | no, text only | yes, over the agent's real transport, such as ElevenLabs, OpenAI Realtime, Pipecat or Twilio ([voice](https://langwatch.ai/scenario/voice/getting-started)) | | Adversarial testing | no | a red-team agent that escalates attacks over several turns ([red teaming](https://langwatch.ai/scenario/advanced/red-teaming)) | ## Choose Scenario when * You want agent tests to live in your existing test runner, next to unit tests, against your agent's real code. * Your criteria need judgment: tone, helpfulness, whether an explanation was clear. * You want to steer a conversation turn by turn and assert on it in code. * You test voice agents, or want multi-turn red-teaming in the same tool. * Your agent is built on a framework it has an integration guide for, such as LangGraph, CrewAI, Pydantic AI, Mastra or the Vercel AI SDK ([agent integrations](https://langwatch.ai/scenario/agent-integration)). * You already use the LangWatch platform, or want people who don't write code to write scenarios there. ## Choose sandtable when * The rules are deterministic (eligibility, limits, ordering, disclosure), and a judge reading a transcript that "sounds fine" is not good enough. * Many tasks share one world and one rule set, and you want them declared once. * You need reliability across repeated trials, measured as pass^k. * The suite should be drafted from what a customer hands over, and reviewed by someone who does not write code. ## Use both They answer different questions. Scenario can check how the agent sounds; sandtable checks what the agent did to the customer's records and whether it broke a rule while doing it. A team can keep conversational-quality, voice and red-team tests in Scenario and gate releases on a sandtable pass^k threshold. Last reviewed October 2026. Something out of date or unfair? [Open an issue](https://github.com/sandtable-ai/sandtable/issues). This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.