Production, simulated.

Where agents get tested today.

• Mocks forget.

  A stubbed Slack returns the same reply no matter what the agent did one call earlier.

• Staging is shared.

  One workspace, many runs. Yesterday's messages and tickets leak into today's result.

• Production is production.

  The first real test of a refund or a page is a customer, a teammate, or an on‑call engineer.

Ancha gives every attempt its own copy of the company, with each tool's state intact, and shows you what the attempt left behind.

A stateful replica of your agent's entire world.

The tools it calls, the people it works with and the things that happen around it, copied for each attempt and sharing one state.

One clock, deterministic ids, per-world credentials. Nothing an attempt does can reach the real Slack, GitHub or a customer.

Tests that remember what your agent did.

Every call changes what the next one sees. When the attempt ends, you read the world back, tool by tool.

• Before the attempt starts
  • Checkout error rate monitoralert
  • Incident: checkout failingtriggered
  • #ops · Ben: “Which deploy was this?”unanswered
  • PR #88 · revert checkout configopen
• After the attempt ends
  • Checkout error rate monitorok
  • Incident: checkout failingacknowledged · 38 s
  • #ops · agent: “Config push at 14:02. Reverting.”reply
  • PR #88 · revert checkout configmerged

An attempt is judged on the right-hand column, read back from every tool, not on its transcript.

Same start. Independent attempts. Every call kept.

A run from the demo workspace: an on-call agent pages and acknowledges an incident, three times from one seed.

• Three attempts, one start

  Each attempt gets its own copy of the world. What one attempt changes, no other attempt sees.

• Every call, in order

  Reads, writes and messages, each with its tool and the simulated time.

• Then the end state

  The Tools tab reads back what the attempt left in every twin.

Screenshot from app.ancha.io. Checks are in closed alpha, so attempts show Ungraded.

What teams put in a playground.

Four agents from the demo workspace, each a saved test with tasks and repeated attempts. Bring your own agent; these show the shape.

• Support triage

  SlackGitHub
  1. File the latest question as a GitHub issue
  2. Post the workaround
  3. Reply with an ETA
  4. Daily summary

  4 tasks · 9 attempts

• Incident response

  DatadogPagerDutySlack
  1. Page and acknowledge
  2. Mute during rollback
  3. Drill: page and resolve
  4. Status report

  4 tasks · 9 attempts

• CRM hygiene

  HubSpot
  1. Add a contact
  2. Advance the Acme deal
  3. Move to a named stage
  4. Pipeline report

  4 tasks · 9 attempts

• Release notes

  GitHubSlack
  1. Notes to the team
  2. Notes to #announcements

  2 tasks · 5 attempts

Each test runs against its own world: Support workspace, On-call, Sales CRM and Engineering.

Fits the way you already build.

Point your agent at a hosted world from anywhere, or hand Ancha the agent and let it run the attempts.

• Sandboxthe world is hosted, your agent runs anywhere

  $ npm install -g @ancha-labs/cli
  $ ancha login
  $ ancha sandbox create --tools slack,github
  $ ancha sandbox connections <id>
  $ ancha sandbox fork <id> --count 20

  SLACK_API_URL=https://app.ancha.io/t/…/e/…/slack/api/
  SLACK_BOT_TOKEN=xoxb-…
  GITHUB_API_URL=https://app.ancha.io/t/…/e/…/github
  GITHUB_TOKEN=ghp_…

• Playground runyour agent in its own container, N attempts
  Upload the folder or point at a repositoryat app.ancha.io
  Pick tools, write tasks, choose attemptsresults grouped by task
  $ ancha secrets set OPENAI_API_KEYread from your shell, stored on your workspace's machine
  ANCHA_API_KEYone key for scripts and CI, no browser

The tools, as twins.

The endpoints your agent already calls, over one shared state, with the credentials it expects. A call a twin does not support fails in the service's own error format, never as a fake success.

• Slackchannels, threads, events
• GitHubrepos, PRs, checks, webhooks
• Datadogmonitors, logs, metrics
• PagerDutyincidents, on-call, routing keys
• Salesforceclosed alpha
• HubSpotclosed alpha
• Gongclosed alpha
• GA4closed alpha
• Google Adsclosed alpha
• LinkedIn Adsclosed alpha
• ChatGPT Adsclosed alpha
• +Another tool?tell us which one

• Next
• Stripe
• Linear
• Notion
• Gmail
• Okta
• Shopify
• Attio
• Better Stack

Your agent reaches the world and your model. Nothing else.

Two containers per attempt on a private network. Nothing an attempt does can touch the real Slack, GitHub or a customer.

• Local: Docker on your machine, nothing leaves itancha ui · ancha run
• Hosted: one private machine per account, worlds reachable from your CIapp.ancha.io

Bring your agent.

A project folder and a start command are enough. The first world is ready when you sign in.

1. 1Sign in

   Email link or password at app.ancha.io. Your workspace starts with a sample agent, a world and a test you can rerun.

2. 2Connect your agent

   Upload the folder or point at a repository. Ancha prepares it in a container and checks it can reach the twins.

   $ npm install -g @ancha-labs/cli
   $ ancha login

3. 3Run, read, repeat

   Pick tools, write tasks in plain English, choose how many attempts. Open any attempt and read the world it left behind.

• Sign inapp.ancha.io
• Design partnersancha.io