Early Access
Request early access

Product

The people your app is built for, played by ghosts.

Ghost Company describes who uses your app, what each of them may do and what they are trying to achieve. Then it runs them together in a browser, records what happened and tells you whether a fix holds.

  • Available
  • Synthetic labs only
  • Optional, when configured
  • Disabled for customer apps

01Ghost users

A ghost is three separate decisions.

Available

Who the ghost is, what it is allowed to do and what it is trying to achieve are defined independently. Appearance never grants capability: the same ghost can be an anonymous visitor in one run and an administrator in the next.

Persona

Identity and temperament

A name, a bio, traits, pace and curiosity. Carries no permissions.

Access profile

Audience and permissions

What the ghost is granted, and what it will attempt anyway.

  • anonymous
  • registered
  • staff
  • administrator
  • service

Objectives

What it is trying to do

Goals inside the scenario, such as "submit a document and confirm it was approved".

Ghost user

Echoas Editor, submitting a document
  • Oliver MeticulousReads every value twice and notices when two screens disagree.
  • Luna ExplorerFollows every link, including the ones she was never meant to find.
  • Pixel SkepticCompares what the interface says with what the server returns.
  • Mochi ImpatientDouble-clicks, goes back mid-flow and submits forms half finished.
  • Echo OrganizedChecks that every hand-off between people actually completes.

The five built-in personas. Workspaces can register their own personas and scenarios without code changes.

02Test modes

One ghost, a crowd, or a whole company.

Available

The same engine runs all three modes. Pick the one that matches the question you are asking.

Single Ghost

One person works through one journey: sign-up, a booking, a form. The quickest way to check that a path works at all.

Ghost Crowd

Several independent people use the app at once. Useful for shared state, limited capacity and anything that should not collide.

Ghost Company

A fictional organization with units, positions and reporting lines. For approvals, escalations and hand-offs between teams.

Coordinated ghosts

In a parallel scenario each ghost acts in its own lane. An action can wait for another ghost's action to finish, and several actions can start together at a named point with a timeout. Validation rejects cycles and one-sided groups, so no ghost waits forever. Waiting is shown as coordination, never as a failure.

03Scenario Studio

Scenarios are data, and every version is kept.

Available

A scenario describes the application context, the ghosts, the workflow and a timeline of actions with checks. Studio lets you build one lane per ghost, add assertions, and see validation problems next to the card that caused them.

  1. 1Save a validated draft. Drafts cannot run.
  2. 2Publish an immutable version.
  3. 3Pick an authorized lab and preview the execution plan.
  4. 4Launch. Findings keep the exact version that produced them.

Definitions follow the versioned ghost.scenario/v1 format and can be imported and exported. Built-in templates cover booking, education, SaaS onboarding, access control, support, recruiting and e-commerce.

Scenario Studio execution lanes. Ghost A, Echo as editor: sign in, create record, share record, revoke access. Ghost B, Luna as viewer: sign in, read record, check revocation. Each action lists the actions it waits for and its assertions.
Screenshot Execution lanes in Scenario Studio. Each ghost keeps its own order; arrows mark what an action waits for.

04AI-assisted drafting

Describe the goal. Review the draft. You decide.

Optional, when configured

An administrator can describe a testing goal in plain language and receive a proposed scenario. The proposal is untrusted until a person reviews it and accepts it as a draft. Publishing and running stay separate steps. The ghosts themselves are not autonomous agents: runs follow the approved scenario and its browser assertions.

The AI can

  • Propose ghosts, access profiles, browser actions and assertions
  • Propose dependencies and start-together groups between ghosts
  • State its assumptions and the inputs it could not resolve
  • Reuse elements from tested reference recipes

The AI cannot

  • Publish, launch or save a draft, or approve its own output
  • Register or authorize targets, or change credentials and roles
  • Browse or fetch your application, run code or call tools
  • Choose a URL, host, scenario key or version
An unreviewed AI proposal in Scenario Studio, listing unresolved inputs, assumptions and validation warnings, with Reject proposal and Accept into draft buttons.
Screenshot A proposal waiting for review. This one came from the offline mock provider used for development and tests.

Disabled by default. Operators enable it with a provider key on the server, never in the browser. Target addresses, credentials, evidence and run data are never sent to the model.

05Browser execution

Real Chromium, in a place that was approved first.

Synthetic labs only

Browser scenarios run with Playwright and Chromium inside a fresh sandbox per run, against operator-approved synthetic labs. Pass or fail comes from assertions on what the browser actually saw.

One sandbox per run

Each run gets its own gVisor-isolated browser that is destroyed afterwards.

Only the approved origin

All traffic leaves through a proxy that allows only the run's authorized lab.

Reset before every attempt

Labs are reset to a known synthetic state, and readiness is checked before any action.

Testing your own application

Running ghosts against customer applications is disabled until independent security review and release approval are complete. Built-in demo runs in Mission control use a scripted runner and are labelled that way in the product.

06Findings and evidence

Facts first. Hypotheses labelled.

Available
Observed by
The ghost and access profile that saw it, and the exact scenario version
Reproduce
Numbered steps another person can follow
Expected
What the check required
Observed
What actually happened in the browser
Evidence
Event timeline and available screenshots, kept private to your workspace
Investigate
A starting point, labelled as a hypothesis until someone confirms it
Fix prompt
Reviewed text you can paste into a coding assistant, without credentials or URLs
A guided test result: a problem was found, five checks passed and one failed, followed by the finding 'The editor didn't see the approval' with expected, observed, why it matters and where to investigate.
Screenshot A verified problem from the guided document-approval test, run in Chromium on a synthetic demo.

Findings are deduplicated across runs with occurrence counts and move through open, acknowledged and resolved. Reports and single findings export to Markdown.

07Regression checks

Know when a fixed bug comes back.

Synthetic labs only

A saved check repeats one published scenario version in a lab. Choose All assertions pass, or approve a complete run as a baseline and choose No new failures since a trusted run.

Comparisons only conclude when conditions match: the same scenario version, lab configuration, initial fixture and execution options. Otherwise the result is inconclusive, and it says so. A missing finding is not treated as a fix unless the same checks ran and passed.

Assertion groups compared between two runs. Every check passes in both except 'check revocation', which fails in both runs.
Screenshot Run comparison: one check failed in both runs under comparable conditions, so it is reported as recurring, not new.

08CLI and CI

Run the same check from a terminal or a pipeline.

Synthetic labs only

Generate a token scoped to one saved check. It expires after 30 days and can be rotated or revoked. The CLI prints one versioned JSON object and an optional Markdown summary without credentials, URLs or response bodies.

A reusable GitHub Actions workflow runs the check on a dedicated runner in your isolated environment, and can require the lab to report the exact commit under test.

# Run a saved regression check
npm run --silent ghost -- check \
  --org "$GHOST_ORG_ID" \
  --check "$GHOST_CHECK_ID" \
  --revision "$GITHUB_SHA" \
  --markdown ghost-regression.md \
  > ghost-regression.json
CLI exit codes
ExitMeaning
0Completed and met the saved pass policy
1Completed and failed the saved policy
2Infrastructure, authorization, timeout or inconclusive evidence
64Invalid CLI configuration
130Interrupted locally

09Workspaces and accounts

The parts a team needs around the tests.

Available
  • Workspaces and projectsOwners, administrators and members, with invitations and per-workspace isolation.
  • Sign-inEmail verification, password reset, active sign-in review and optional two-step sign-in with an authenticator app.
  • Live runsEvents stream to the Ghost office as they happen and survive refreshes and reconnects.
  • Usage limitsUsage measured in ghost-minutes, with limits on concurrency, ghosts per run and monthly use.

Want ghosts on your app first?

Early access starts with teams whose apps fit the current lab boundary. Tell us what you are building.