Stunt Double MCP Server

Add a step to a workflow and connect it into the run. Only stores it: nothing executes until the workflow runs. Appends by default; `sort_order` splices it in at a position. In a run, run_checklist drives a real browser; notification emails or messages members on Slack; github_comment and actor_pr_review post on the triggering pull request; crawl_topology crawls a topology in a real browser to make a new version of its map; condition and delay only steer or pause. Config per step type: run_checklist: { checklist_id, use_deployment_url? } runs a checklist and passes its pass/fail counts to the next step. condition: { field, operator, value } continues down its True branch when it holds and ends the run otherwise. `field` is 'previous_step.status', another 'previous_step.<key>' off the last step's output, or 'trigger.<key>' off the event that started the run. `operator` is eq, neq, gt or lt. notification: { channel, recipients, template, subject?, body?, slack_channel_id? }. channel is 'email' or 'slack'; recipients is { type: 'all_members' } or { type: 'members', member_ids: [...] }; template is 'completed', 'failed' or 'custom' (custom uses subject and body). delay: { duration_seconds } waits before the next step. Waits are checkpointed, so a long one is free. github_comment: { body_template } comments on the pull request that triggered the run (GitHub App or Cursor Origin, depending on the trigger). The template may use {{passed}}, {{failed}} and {{summary}}. actor_pr_review: { actor_id, review_focus? } has an actor review the triggering pull request. crawl_topology: { topology_id } crawls a topology again and passes the new version's status, page count and what changed since the last version (added, changed, removed) to the next step.

Create a workflow (multi-step automation) that runs checks without anyone asking. Triggers: manual, schedule (cron), webhook, or vercel_event / github_event / cursor_event / linear_event / slack_event / figma_event so checklists re-run on deploys, pull requests, or design updates. Add steps with add_workflow_step (each one is wired to the step before it). A workflow with any trigger but manual starts runs on its own, so a connection without the mcp:run scope creates it paused; activate it with toggle_workflow over a connection holding mcp:run, or from the dashboard.

Delete a workflow and its steps and runs, and remove its schedule. This discards run history; prefer toggle_workflow to pause unless the user asks for deletion.

Get a workflow with its steps, the edges connecting them, the order a run would reach them, and recent runs. Read this before editing steps: `execution_order` is what actually runs, and any id in `unreachable_step_ids` is a step no run will ever reach.

Get a workflow run with step-level details. Poll this after run_workflow until status is terminal (completed, failed or cancelled).

Remove a step and close the gap: whatever led into it is repointed at whatever it led to, so the rest of the automation still runs.

Trigger a workflow run now (async), whatever its trigger type: it executes the stored steps (see get_workflow), which can drive browsers, post PR comments and notify members. Returns a run ID; poll get_workflow_run. Refused over the plan limit or budget cap.

Update a workflow's name, description, or trigger. Only the fields you pass change. Syncs the Trigger.dev schedule when the cron or trigger type changes. The trigger fields edit the workflow's one trigger; a workflow with several triggers has them edited in the dashboard instead. Changing the trigger of an active workflow that runs on a schedule, webhook or event needs the mcp:run scope; pause it first otherwise. To change steps use the step tools; to pause or resume use toggle_workflow.

Change a step's type or config in place, leaving its connections alone (move it with reorder_workflow_steps). Only stores the change. Config shapes are as in add_workflow_step.

Add a question or task to an interview section (create the section first with add_interview_section). Tasks are concrete jobs to attempt in the product ("Find the annual plan price and compare it with monthly"); questions probe comprehension and sentiment. Set expected_evidence on tasks so the report can judge success.

Attach a participant to an interview. Provide either actor_id (existing actor) or persona_spec (ad-hoc persona: name, bio, traits), not both. Vary personas and devices across participants (expertise, motivation, scepticism, mobile vs desktop) for richer findings, and use behaviour to dial scepticism, technical literacy and frustration up or down. Call once per participant, then launch_interview.

Add a section to an interview’s discussion guide. Sections group related items (e.g. "First impressions", "Checkout"); add questions and tasks to it next with add_interview_item, using the returned id as section_id.

Create a user interview: AI persona participants complete tasks and answer questions against target_url, and results are synthesised into a report. target_url can be any reachable URL, including prototypes (Figma Make publish links, Claude design artifacts, v0) and preview deployments. Build the guide with add_interview_section / add_interview_item, add participants, then launch_interview.

Get an interview with its discussion guide (sections and their items, in order) and participants with their run status. Also useful for polling progress after launch_interview; call get_interview_report for the synthesised findings.

Get an interview participant including their full transcript from the interview run. Use this for verbatim evidence behind a report finding.

Get an interview's synthesised report (summary, themes, recommendations, per-question rollup), plus `usability` (UMUX or UMUX-Lite scores when asked). Poll after launch_interview until report_status is completed or failed; report is null until generated.

Update an interview’s name, target_url, research brief, status or usability questionnaire. Only the fields you pass change; pass at least one. To start a round use launch_interview rather than setting status to running.

Add knowledge to an actor. Choose the kind: expertise for a skill or standard it judges by in every run (brand, tone of voice, design system, compliance, accessibility), process for steps it follows, reference for facts it looks up when relevant (product docs, FAQs, policies), account for test account details. Never pass real passwords, API keys or one-time codes; actors sign in with email codes. Check list_actor_knowledge first to avoid duplicates.

Create an actor: a persistent AI user persona with its own browser, email inbox (for OTP and magic-link flows), and memory. The description shapes who the persona is, so describe a person (background, goals, tech comfort), not a job title. Call list_actors first and reuse a matching actor; afterwards, add_actor_knowledge gives it context or standards to enforce. Fails when the plan's actor limit is reached.

Get an actor's details: description, system prompt, capabilities, and status. list_actor_knowledge has the knowledge it carries into runs.

List what an actor knows, oldest first. Each entry has a kind: expertise and process entries are listed by title in every run and the actor reads one when the task calls for it; reference and account entries are found by search. Check here before add_actor_knowledge to avoid duplicates; pass an entry id to remove_actor_knowledge.

Remove a knowledge entry from an actor permanently, so future runs no longer use it. Find the entry's id with list_actor_knowledge.

Update an actor's name, description, system prompt, capabilities, or status. Only the fields you pass change; at least one is required. Set status to "archived" to soft-delete.

Create a browser checklist: an actor exercises the product at a URL and judges each check with evidence. Checks can verify flows ("a new user can complete signup") or assert standards ("headlines use sentence case"), for brand, compliance and design-system audits. Any reachable URL works, including previews and prototypes; actors handle email sign-in (OTP, magic links) through their own inboxes, but not other auth walls. Needs project_id or url, an actor (list_actors), instructions and at least one check. Search or list_checklists first to reuse one; run with run_checklist.

Delete a checklist and its checks and runs. This permanently discards run history; prefer update_checklist, or simply not running a checklist, unless the user asks for deletion.

Get a checklist: its target URL, instructions, run settings, every check (with the `id` update_checklist needs to edit one in place) and its 10 most recent runs. Call get_checklist_run with a run id for per-check results and evidence.

Get a checklist run: per-check results with evidence, score, verdict and the change since the previous run. Poll after run_checklist every 30 to 60 seconds until `status` is terminal; `outcome` is the verdict on the experience. `comparison.changes` (what got better or worse) is usually the answer worth reporting. Pass a result's `result_id` to flag_check_verdict when the user says a verdict was wrong.

Run a checklist (async): the actor drives a real browser through the product and judges each check, and may use the project's attached MCP servers (list_project_mcp_servers). Returns a run_id; poll get_checklist_run (runs take a few minutes). Counts against the workspace's plan and budget. Results reach the actor's knowledge only with retain_memory.

Update a checklist. Only the fields you pass change. Pass `checks` to replace the full set of checks, sending each existing check with the `id` from get_checklist so it keeps its run history. Changing the checks, instructions or url retires the recorded script, so the next run drives the agent afresh.

Add a standing rule and hold this project to it (a design-system rule, a tone-of-voice rule, a compliance obligation, or shared knowledge). Every checklist run, design review, interview and triage for the project is then held to it. The rule joins the workspace library, so check list_workspace_guidelines first and attach an existing rule with `set_project_guideline` rather than adding a second copy.

Create a project (a product to track with checklists, workflows, interviews and feedback). Call list_projects first and reuse a matching project rather than creating a duplicate.

Get one project (the product tracked by checklists, workflows, interviews and feedback), including its URL, audience and product description. Archived projects read as not found.

List the standing rules this project is held to: the workspace library rules it has attached. Read these before writing checks, interview guides or design feedback for the project, and assert them as observable checks rather than restating them. `attachment_id` is this project's copy of the rule; `id` is the library rule shared with every other project holding it.

List the third-party MCP servers attached to a project and the tools each allows. Configuration only: those tools are not callable here, but an actor in a run (checklist, interview, workflow) may call the allowed tools of servers whose in_force is true. Servers are registered and attached by workspace admins in the dashboard.

Attach a library rule to a project, detach it, or switch it off for this project alone. Attaching is how one workspace rule comes to hold for several projects without being retyped. `enabled: false` keeps the attachment but stands the rule down here, leaving every other project untouched. Safe to repeat: setting the state the project is already in changes nothing.

Add a standing rule to the workspace library. Call list_workspace_guidelines first so you do not add a near-duplicate. Attach it to the projects it applies to with `project_ids` here or `set_project_guideline` later, or set `apply_to_design_reviews` so every design review in the workspace is held to it, including the Slack and Linear ones that have no project. Codify a standard here once rather than repeating it in each checklist.

Get workspace details by ID or slug, including the admin-set controls that decide what runs in it may do. `settings` is a ceiling, never a grant: a network policy there applies on top of each worker's own, and a feature switched off there cannot be switched back on per project.

List the workspace's whole library of standing rules (design system, tone of voice, brand, compliance, accessibility, shared knowledge), with how many projects hold each one. Read this before writing a rule: attaching one that already exists with `set_project_guideline` beats typing a second copy of it.

List the members of a workspace with their role, name and join date. Use it to resolve the user ids that runs, feedback and assignments carry.

Remove a rule from the workspace library permanently. This detaches it from every project holding it, so prefer `update_workspace_guideline` with `enabled: false` when the rule may come back.

Update a rule in the workspace library, or switch it off. Every project holding it is held to the new wording, so prefer editing the shared rule over adding a near-duplicate. Pass only the fields to change; at least one is required.

Get one feedback widget comment with its full reply thread, its page location and version, and a signed screenshot URL when one was captured. Find the id with list_feedback.

List the comments stakeholders and users have left on a project through the feedback widget, newest first. Each one is pinned to a page position (often to a named element) and carries device info; filter by status "new" when triaging. Use summarise_feedback to turn a list of these into themes and an implementation brief, and get_feedback for one comment's replies and screenshot.

Update a feedback widget comment's status: "reviewed" once assessed, "resolved" when confirmed fixed (ideally after a checklist run proves it), "dismissed" sparingly. The dashboard reflects the change immediately.

Get a site's Stunt Double Index report: how well AI agents (ChatGPT, Claude, Gemini and others) can find, understand and act on it. Returns the score out of 100, band, rank, category and provider scores, frictions, failing checks and when it was scored. A blocked or opted-out site has a null score with unscored_reason. Pass a domain, or a project_id for its linked domain. Use list_index_sessions to see why a category or provider scored low.

List the agent sessions behind a site's Index score: one AI provider attempting one benchmark task in a real browser, with status, score, rubric evidence, a summary and the frictions it hit. Defaults to the latest run (pass run_id for another); filter by provider or category. Use it to explain a low score.

Search the Stunt Double Index for sites by domain or name, or browse the leaderboard. Results are ordered by overall score, highest first, with rank, sector and when each was last scored; sites the Index could not score come last with a null score and an unscored_reason. Omit query to read the top of the leaderboard; add sector to compare a site against its peers. Follow up with get_index_report for the full breakdown of any result.

Get details for a GitHub pull request (title, description, author, branches, change stats) via the workspace GitHub connection. Read it before writing a checklist for the change, then report back with comment_on_pr.

List pull requests for a GitHub repository via the workspace GitHub connection, most recently created first (up to 30). Use it to find the PR number to pass to get_pull_request or comment_on_pr.

Re-score a site on the Index after it ships changes for agents: fresh probes, then about 24 agent sessions over a few minutes. Owner only (whoever claimed it, a work email on that domain while unclaimed, or a platform admin), once every 10 minutes. Poll list_index_sessions with the run_id about every 60 seconds, then read get_index_report.

Add a site the Stunt Double Index does not track yet and give it a first score. Runs light HTTP probes (robots.txt, llms.txt, structured data, sitemaps) in under a minute; include_agent_sessions also starts live agent sessions, which may queue behind a shared daily budget. The site becomes public on the leaderboard. Public domains only; robots.txt opt-outs are respected; a tracked site is not rescored (use request_index_rerun). Then read get_index_report, and poll list_index_sessions with the run_id if sessions started.

List active actors (persistent AI user personas) in a workspace, oldest first. Prefer reusing an existing actor over creating a new one; pass an id from here to get_actor, interviews and checklists.

List the checklists in a workspace, with their target URL, actor, check count and a recent run. Check here before creating a new checklist: re-running an existing one keeps its history as a regression signal. Use get_checklist for the checks themselves.

Post a comment on a GitHub pull request via the workspace GitHub connection. Use it to report checklist or interview results where engineers already look: a short verdict, then per-check outcomes with evidence highlights. Needs the Stunt Double GitHub App installed on the repository; call get_pull_request first to confirm the PR is reachable.

Wire a step's output handle to another step, to build or repair a branch (adding and removing steps keeps the plain chain on their own). A condition's 'false' handle starts unconnected, so a failing condition ends the run. Connecting a connected handle replaces its edge; target_step_id: null disconnects it.

Get an actor conversation with its full message history, oldest message first. Find conversation ids with list_conversations.

List chat conversations with actors in a workspace, most recently active first, optionally filtered to one actor. Excludes the workspace assistant's conversations. Pass an id to get_conversation to read the messages.

Correct a wrong passed or failed verdict on one check: pass the result_id from get_checklist_run and a note on what was actually true. Later runs use it as calibration and re-record any code-run script; the run's own score is unchanged. Set clear to withdraw a correction.

Get the account this connection is acting as: id, email, name, timezone and where notifications reach them. Use `timezone` when creating a scheduled workflow rather than defaulting to UTC, and `id` to tell your own runs and feedback from a teammate's in lists that carry only user ids.

List the user interviews in a workspace, newest first, optionally filtered to one project. Returns summary rows; call get_interview for the discussion guide and participants, or get_interview_report for the findings.

Launch an interview round (async): every participant runs the discussion guide against the target URL in a real browser, and may call the allowed tools of MCP servers attached to the project (list_project_mcp_servers); then a report is synthesised. Requires at least one section with items and one participant. Poll get_interview / get_interview_report for progress (runs take several minutes).

List the live projects in a workspace, most recently opened first. Call this after list_workspaces and match a target URL or product name against these before creating anything new.

Regenerate the synthesized report from existing participant transcripts (async), e.g. after adding participants. Returns the trigger run ID; poll get_interview_report until report_status is completed or failed.

Set the order a workflow runs its steps in, as one linear sequence. Pass every step id: a list that leaves a step out is refused, since that step would be disconnected from the run (remove it with remove_workflow_step instead). This rewires the whole automation, so a branch built on the canvas (a condition with both a True and a False path) does not survive it; use connect_workflow_steps to rebuild a fork.

Search everything in a workspace (projects, actors, checklists, interviews, automations, issues, goals, feedback, knowledge, insights, design reviews, resources, conversations) by name or content. Use it instead of listing and filtering, and before creating something that may already exist. Results carry the id for the matching getter. Omit `query` to browse recent items.

Summarise a project's comments into themes ranked by severity, open questions, and a self-contained implementation brief for a coding agent. Read the brief before proposing changes instead of reasoning over raw comments. The brief leaves out work an open question holds up; `answers` says which are answered, so ask the user the rest. Shared with the dashboard: an unchanged comment set returns the saved digest instantly.

Activate or pause a workflow. Scheduled workflows have their Trigger.dev schedule activated/deactivated to match. Activating a workflow that runs on a schedule, webhook or event needs the mcp:run scope, since it then starts runs on its own; pausing never does. Pausing keeps the steps and run history; prefer it to delete_workflow.

List workflows (multi-step automations) in a workspace, newest first, with their triggers, whether each is active, its step count and its last run. Call get_workflow on one for its steps and the order a run takes them.

List the workspaces you belong to, with your role in each. Call this first: almost every other tool needs a workspace_id from here.