Autonomy Docs
AI Assistants

Available Actions

The 15 MCP tools for Test Cases, Test Plans, Runs, Environments, and External Evidence.

Autonomy MCP exposes 15 tools generated from the same operation catalog as the HTTP API. The names below are stable and use the product glossary: a Test Case is an executable sequence of steps, while a Test Plan is an ordered selection of Test Cases.

Test Cases

ToolPurpose
autonomy_test_case_listList Test Cases in the organization, newest first. Filter by platform or a name query. Use before creating a case so existing coverage is reused.
autonomy_test_case_getFetch one Test Case with its full step list.
autonomy_test_case_createCreate a Test Case from explicit steps. Fill values are literal, agent (Agent-chosen at run time), or environment (named Environment runtime value); credential-shaped literals are rejected unless explicitly relocated to an Environment secret.
autonomy_test_case_update_stepsReplace a Test Case's whole step list. Requires confirm: true; without it the response describes what would change and nothing is written.

Test Plans

ToolPurpose
autonomy_test_plan_listList Test Plans (named selections of Test Cases) in the organization.
autonomy_test_plan_getFetch one Test Plan with its ordered member Test Cases.
autonomy_test_plan_createCreate a Test Plan and optionally its initial ordered member Test Cases.
autonomy_test_plan_set_membersReplace a Test Plan's ordered membership. Removing cases requires confirm: true; without it the response lists the cases that would be dropped and nothing is written.

Runs and evidence

ToolPurpose
autonomy_run_triggerQueue Runs. Give exactly one of testCaseId or testPlanId (a plan queues one Run per member case). Target a saved Environment by slug, or pass explicit targets for a fresh deployment/artifact. Returns immediately; follow with run.get.
autonomy_run_listList recent Runs, newest first. Filter by status, Test Case, Environment, or branch. Check here before triggering duplicate work.
autonomy_run_getFetch a Run's state. Pass the previous cursor and waitMs (≤ 25000) to block until something changes; timedOut: true means nothing changed and is not an error. next is advice, never authority.
autonomy_run_evidenceExternal Evidence for a Run: per step, what the product showed, what was expected, the verdict and diagnosis, neutral actions, and short-lived media links. Never model, prompt, token, or runner internals. Filter with step.

Environments

ToolPurpose
autonomy_environment_listList Environments (named deployment targets) with their per-platform targets. Runtime value names are shown; values never are.
autonomy_environment_getFetch one Environment by slug.
autonomy_environment_upsertCreate an Environment, or update one by slug. Updating an existing slug overwrites the given targets and requires confirm: true; without it the response describes the change and nothing is written.

Notes are available through the HTTP API only and are not MCP tools.

Confirming replacements

Autonomy does not ask for a second interactive confirmation inside MCP. Instead, operations that replace existing configuration require an explicit confirm: true:

  • autonomy_test_case_update_steps always replaces the whole step list.
  • autonomy_test_plan_set_members requires confirmation when the new membership drops cases.
  • autonomy_environment_upsert requires confirmation when its slug already exists and the supplied fields would be updated.

Without confirmation, the tool returns INVALID_ARGUMENT with confirmRequired: true and a wouldChange summary. Read that summary before retrying with confirm: true; do not add confirmation reflexively.

Waiting for a Run

autonomy_run_trigger returns immediately. Follow it with autonomy_run_get:

  1. Call autonomy_run_get with the returned runId.
  2. Save the response's numeric cursor.
  3. While next is wait, call it again with that cursor and waitMs up to 25000.
  4. A response with timedOut: true means the Run did not change during that wait. It is not an error; repeat with the same cursor or stop waiting.
  5. When next is evidence, call autonomy_run_evidence.

next is advice, not authority. Use the returned status, verdict, validity, progress, and step state when deciding what to do.

External Evidence

autonomy_run_evidence returns only the evidence an organization may take out of Autonomy: step actions and targets, observed and expected outcomes, verdict and validity, diagnosis, recovery counts, and short-lived screenshot, frame, or video links. Media links expire after 15 minutes; call the tool again for fresh links.

External Evidence never includes model or provider identity, prompts, token or cost figures, raw runner errors, internal tool names, or raw UI trees.

Trigger targets

autonomy_run_trigger accepts platforms with web, api, ios, and android. Use environmentSlug when the target is already saved. Use targets for a fresh deployment or artifact.

PlatformTarget fields
Webtargets.web.baseUrl
APItargets.api.baseUrl, optional targets.api.runtimeValues
iOStargets.ios.artifactUrl or scanned targets.ios.storageId, plus targets.ios.bundleId
Androidtargets.android.artifactUrl or scanned targets.android.storageId, plus targets.android.packageName

Give exactly one of testCaseId or testPlanId. Add branch, pr, and deployment.commitSha when the Runs should attach back to a pull request or commit.

OAuth scopes

Read tools require autonomy:read. Write access is separated by resource:

  • autonomy:test_cases:write
  • autonomy:test_plans:write
  • autonomy:runs:write
  • autonomy:environments:write

The autonomy:notes:write scope belongs to the HTTP-only Notes operations.

Agent recipes

Check existing coverage

Codetext
Use Autonomy MCP to list relevant Test Cases, Test Plans, and recent Runs.
If a relevant Run already exists for this branch or PR, summarize its status before triggering anything new.

Run a PR preview

Codetext
Use Autonomy MCP to trigger the relevant Test Case or Test Plan against this preview URL.
Pass the branch, PR number, and commit SHA. Wait for the Run to finish, then return its verdict and External Evidence.

Diagnose a failure

Codetext
Fetch the Autonomy Run and its External Evidence before proposing a fix.
Cite the failing step, observed outcome, diagnosis, and media link that support your conclusion.

On this page