generated: '2026-08-14'
method: searched
source: >-
  https://docs.ada.cx/reference/introduction/overview,
  https://docs.ada.cx/docs/optimization/testing,
  https://docs.ada.cx/docs/optimization/testing/interactive-testing,
  https://docs.ada.cx/docs/optimization/testing/simulations,
  https://docs.ada.cx/reference/integrations/getting-started,
  https://docs.ada.cx/reference/introduction/migrate-to-v-2
docs: https://docs.ada.cx/docs/optimization/testing
model: separate-instance
summary: >-
  Ada has NO test/live key split, no test-mode prefix, no magic test identifiers and no fixture
  values. Isolation is achieved by using a whole separate AI Agent instance — Ada's guidance for
  integration partners is "use a sandbox AI Agent specifically for testing and developing your
  integration, not a live AI Agent", and its v2 migration guide says "test your updated requests
  in a staging or sandbox environment". Everything below is what Ada actually publishes; nothing
  was invented.
test_live_separation:
  key_prefixes: none
  mode_flag: none
  mechanism: >-
    A dedicated development/sandbox AI Agent with its own instance host and its own API keys.
    Keys are scoped to an instance, so pointing at a different <handle>.ada.support is the
    boundary. There is no in-band way for a caller to tell a test key from a live one.
  integration_guidance: >-
    "Access to an Ada development AI Agent: Use a sandbox AI Agent specifically for testing and
    developing your integration, not a live AI Agent. This AI Agent should be dedicated
    exclusively to managing integrations."
  source: https://docs.ada.cx/reference/integrations/getting-started
console:
  name: API Playground
  url: https://docs.ada.cx/reference/introduction/overview
  hosted_in: the documentation site (Fern)
  auth: >-
    Click Login and paste your own API key. "Your key is stored securely in your local browser
    session and will be cleared when the tab is closed."
  usage: >-
    A ▶️ Play button sits next to every API call example. The endpoint URL is editable —
    double-click it to replace `example` with your instance subdomain.
  note: >-
    The playground runs against YOUR live instance with YOUR key. There is no shared demo tenant
    and no anonymous try-it, so an evaluator without an Ada account cannot execute a single call.
testing_tools:
- name: Interactive Testing
  kind: test chat
  description: >-
    Chat with the AI Agent in real time as an end user would. Supports opening the test chat in a
    new window/browser, impersonating different user types, and resetting or ending a test chat.
  docs: https://docs.ada.cx/docs/optimization/testing/interactive-testing
  note: >-
    "Launch controls no longer reroute test chats" (release note 2026-07-29) — launch/targeting
    rules are bypassed for test chats.
- name: Simulations
  kind: scripted test cases
  description: >-
    Create test cases (including from a real production conversation), define expected outcomes,
    run simulations against the Agent, and review pass/fail results with evaluation rationale.
    Voice simulations were upgraded to match real call behaviour end-to-end (2026-06-30).
  docs: https://docs.ada.cx/docs/optimization/testing/simulations
  quota: >-
    Test runs are metered. The MCP tool get_test_run_quota returns "daily test run quota, usage,
    and remaining runs"; limits were increased 2026-06-11.
  agent_access:
    mcp:
    - get_test_cases
    - get_test_runs
    - get_test_run_quota
    - edit_agent_config (creates cases, triggers runs)
    cli:
    - adacli <instance> testing cases from-conversation <conversation-id>
    - adacli <instance> testing runs create
    - adacli <instance> testing runs compare --baseline r1 --candidate r2
- name: Change sets
  kind: staged writes
  description: >-
    MCP writes to Agent behaviour stage on a change set rather than applying live: new resources
    are inactive by default, and the set is tested before being promoted, reverted or discarded.
    Shipped 2026-08-13. This is Ada's closest analogue to a test mode for write operations.
  docs: https://docs.ada.cx/mcp/tools/edit-agent-behavior
  agent_access:
    mcp:
    - edit_agent_behavior
    - list_agent_changesets
- name: Instance snapshots
  kind: config export/import/diff
  description: >-
    adacli snapshot export/import/diff moves a whole instance configuration between environments,
    which is how a sandbox Agent gets seeded from production.
  docs: https://github.com/AdaSupport/homebrew-tap
- name: Action / Playbook test
  kind: unit test
  description: >-
    adacli actions test and adacli playbooks test exercise a single web action or playbook
    without running a conversation.
test_data:
  magic_values: none
  test_cards: n/a
  fixtures: none published
  note: >-
    Ada publishes no magic identifiers, no sandbox tokens and no seeded fixture data. The
    interactive test chat and simulations generate their own conversation data inside the
    development Agent.
example_implementations:
- name: ada-integrations-demo
  url: https://github.com/AdaSupport/ada-integrations-demo
  language: python
  description: Working sample of a Knowledge integration server, including the OAuth callback flow.
- name: ada-conversations-api-demo
  url: https://github.com/AdaSupport/ada-conversations-api-demo
  language: python
- name: ada-handoffs-api-demo
  url: https://github.com/AdaSupport/ada-handoffs-api-demo
  language: python
- name: template-widget
  url: https://github.com/AdaSupport/template-widget
  language: javascript
  description: Minimal Ada widget app scaffold for the Widget SDK.