Head to head · Desktop app control · October 2026 research run

Aura Autopilot vs Glasshand

Aura Autopilot scores 58.7 (C) on agent readiness against Glasshand's 50.8 (D), and leads in 5 of 7 scored categories. Glasshand leads on schema & documentation and security & auth. Both do desktop app control.

Best desktop and mobile app automation for AI agents · All 64 desktop comparisons

Which one, for what

Aura Autopilot C

Best for Agent-driven playtests of Unreal, Unity or Godot games in the editor or in CI

Ahead on

  • Reliability, 41 against 27
  • Agent ergonomics, 69 against 62
  • Payments & pricing, 75 against 35
  • Maintenance & community, 43 against 35

Also in its favour

  • Runs on your own machine
  • Open source

Watch for

Run sharing is on by default and sends recordings, screenshots, agent traces with code excerpts and, for MCP runs, the user's Claude Code or Codex conversation to app.tryaura.dev

Glasshand D

Best for Agents that need a real iPhone or Mac, such as testing native apps in Xcode or using iOS-only apps and a real phone number.

Ahead on

  • Schema & documentation, 79 against 74
  • Security & auth, 60 against 53

Also in its favour

  • A hosted endpoint, with nothing to install

Watch for

No status page, SLA, changelog or security.txt found

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Payments & pricing Aura Autopilot +40

Aura Autopilot 75

Open-source software with paid parts around it, so the pricing lines are read for what a run needs. No payment protocol (0). TypeSafe publishes Jev's price per million input tokens and Aura's plans are public, Indie from $10 a month billed annually (20). Autopilot is free under Apache-2.0, and Aura's two-week trial needs no card (20). No Ramen account is needed with one's own keys, but a person installs the editor plugin and gets a TypeSafe key or signs in to Aura in a browser, so 10 rather than the self-hosted rule's 20 (10). 50 in all. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 50 is 75.

Glasshand 35

No x402, MPP or L402. Checkout returns a Stripe link (0). Per-device monthly prices published without a login, $60 for an iPhone and $90 for a Mac (30). No free tier or trial (0). An agent can start the device authorisation and prepare a checkout without a key, but a person signs in by email code, approves the agent and pays in a browser (5 of 30). 0 + 30 + 0 + 5 = 35.

Reliability Aura Autopilot +14

Aura Autopilot 41

Local-software reading, as a stdio MCP server with engine plugins. Official downloads per engine and OS from GitHub Releases, signed and listed in SHA256SUMS.txt, with Unreal 5.4 to 5.8, Unity 6000.0 or later, Godot 4.7 or later, Windows and macOS and Python 3.12 or later stated. The agent service isn't on a package registry; the plugin installs it, with uv and Bun, on first run (17 of 20). .github/workflows/ci.yml runs ruff and pytest on Linux, Windows and macOS with Python 3.12 and 3.13 plus a Codex end-to-end job, and CONTRIBUTING.md names download and uninstall end-to-end workflows. Whether runs pass wasn't checked, as github.com couldn't be read (15 of 25). Open issues couldn't be read, and the repository is days old (5 of 25). No changelog; releases are tagged v<version>-<commit>, the plugin protocol is versioned and the contributing rules say to add optional fields rather than change existing ones (4 of 15). Python package 0.1.0 and the Unreal plugin marked IsBetaVersion (0). 41 in all.

Glasshand 27

Hosted lines, for the HTTP API. No status page is linked from the site, agents.md or the terms (0), so no incident history could be read (5). No rate limits with numbers were found. agents.md says to back off on 429 (0). Checkout, cancel, resume and each device input take an Idempotency-Key, and agents.md says to retry a timeout or 503 with the same key and body and never to replay input blindly after a disconnect. No Retry-After is documented (12 of 15). The terms say no particular uptime is guaranteed (0). Registration and paid rentals are open worldwide with live billing, and no beta label was found (10). 0 + 5 + 0 + 12 + 0 + 10 = 27.

Maintenance & community Aura Autopilot +8

Aura Autopilot 43

The launch post is dated 31 October 2026 and Aura 1.2.3 and 1.2.4 of 30 October 2026 ship Autopilot, so the last change is within 30 days (30). GitHub releases couldn't be read, and the two Aura release-note entries aren't three (0). Issue and pull-request replies couldn't be read; the project is days old, with Discord and support email as other channels (5 of 25). Not in the official MCP registry under any namespace (searched for autopilot and ramenvr) (0). Pinned dependencies, per-platform locks in packaging/, CI on three operating systems, conventional-commit checks and code-owner review (8 of 10). 43 in all.

Glasshand 35

Closed service with no changelog. The newest first-hand date is the terms and privacy policy effective 22 October 2026, 27 days before this check, read as the launch date, as beam-win's terms date was (30). No dated release entries to count (0). Support by email (support@ and hello@glasshand.com) and a demo booking link. No public tracker or community (5 of 25). No official SDK (0). No public package or CI (0). 30 + 0 + 5 + 0 + 0 = 35.

Agent ergonomics Aura Autopilot +7

Aura Autopilot 69

MCP reading. 14 tools (15 for 11 to 30), with long descriptions on probe and run_harness and a long INSTRUCTIONS.md the skill says to read in full, with no toolsets (13 of 25). read_file takes offset and limit up to 2,000 lines, list_files returns at most 500, search at most 200 with a glob, frames up to 12 and run_harness takes seconds of play (16 of 20). Errors come back as isError with what to do next (open the editor, update the agent service, call wait), and headless runs exit 0, 1, 2 or 3 (16 of 20). readOnlyHint is set on the seven read tools and destructiveHint false on all, including probe, which the description says can change the game directly. One call runs at a time and a long one answers "still going" for wait to collect (14 of 20). Two required parameters on start_run with defaults of 10 minutes of exploration and a 30-minute limit; no SDKs, which a playtester doesn't need (10 of 15). 69 in all.

Glasshand 62

Responses are small, but screenshots are full JPEG frames, and the WebSocket stream holds at most four unacknowledged frames (15 of 25). Rental lists return all items with no pagination or filters, though an account has few rentals (8 of 20). Some error states are named and explained, with no complete error list (12 of 20). An Idempotency-Key on every write and every input, with documented retry rules and a conflict on a reused key (20). No SDK. Calls are plain HTTPS with few parameters (7 of 15). 15 + 8 + 12 + 20 + 7 = 62.

Security & auth Glasshand +7

Aura Autopilot 53

Local-software reading, as tokeneconomizer was read: 30 for what leaves the machine and credentials, 20 approvals and sandboxing, 15 injection, 15 audit, 20 programme. Run sharing is on by default and sends recordings, screenshots, agent traces with the code the agent read and, for MCP runs, the user's own Claude Code or Codex conversation to app.tryaura.dev, after redacting secrets, paths, user and host names and email addresses. Keys sit in plain text in autopilot.env, the file tools never read *.env, and the localhost control endpoint carries a random secret (12). File tools write only the run's folder, notes.md and lib.py, and the user can stop or steer a run from the tab. autopilot connect codex sets default_tools_approval_mode to approve, and on Windows and macOS the sandbox is a plain child process without keys but with network and disk (10). sandbox.py names prompt injection from a game's code as the reason for the sandbox, and the frame describer is given only open questions (11). Each run keeps its agent log, harness versions, traces, decisions, report and recording, and the Feed tab shows activity (13). SECURITY.md asks for GitHub private vulnerability reporting; no security.txt on www.tryaura.dev (404), no bounty or certification found, advisories couldn't be read. Releases are signed and immutable, and pull requests need a code owner's approval (7). 53 in all.

Glasshand 60

OAuth device authorisation with four scopes, 15-minute access tokens, rotating refresh tokens and immediate revocation from the dashboard. No secret in a URL (30). rentals:read is a read-only scope, and payment and checkout stay in the person's browser. devices:control grants full input with no per-action confirmation, and a person can take control at any time (12 of 20). agents.md treats device screens, web pages, app content and incoming messages as untrusted task data that can't authorise purchases or credential disclosure. No technical control on what an agent does with screen content (10 of 15). Agent authorisation records keep last use, and control.actor names the current controller. No audit log of inputs found (6 of 15). No security.txt (404), disclosure policy, bounty or certification. The privacy policy asks for reports to support@glasshand.com (2 of 20). 30 + 12 + 10 + 6 + 2 = 60.

Schema & documentation Glasshand +5

Aura Autopilot 74

MCP reading. All 14 tools (start_run, wait, probe, run_harness, snapshot, describe_frames, restart_level, report, frames and five file tools) have JSON Schema inputs, read in autopilot/mcp.py, session.py, external.py and files.py (25). No llms.txt on www.tryaura.dev (404). The agent skill, INSTRUCTIONS.md and KIT.md are Markdown written for the agent, and the README and CONTRIBUTING.md are in the repository (7 of 10). Descriptions say what each tool does and when not to use it (frames is for a describer subagent only, describe_frames is a fallback), with server instructions (17 of 20). Required fields and an enum on kind, but report.outcome is a free string with its values only in the description, numeric ranges sit in descriptions rather than the schema, and probe takes free-form Python (9 of 15). Prompt examples in the README and skill, plain-language errors in the MCP server and documented exit codes for headless runs (11 of 15). Version numbers and a protocol version, but no public changelog, and GitHub release notes couldn't be read (5 of 15). 74 in all.

Glasshand 79

A public OpenAPI 3.1 document at api.glasshand.com/openapi.json, version 0.1.0, with 20 operations, linked from agents.md as the API reference (25). agents.md is Markdown written for agents (10). agents.md says when to use each call, how to authenticate, how to retry, how to read the screen and what counts as untrusted data. 15 of 20 operations in the OpenAPI document carry a description (18 of 20). 14 enums and 70 required-field markers, with input types bounded, such as drag duration 100 to 3,000 ms and text up to 512 characters (13 of 15). Examples in agents.md, and named result errors such as control_taken, control_busy, authorization_pending and an idempotency conflict, but no error reference (8 of 15). Paths are under /v1. No changelog found (5 of 15). 25 + 10 + 18 + 13 + 8 + 5 = 79.

Score by category

CategoryWeightAura AutopilotGlasshandEdge
Reliability16%4127Aura Autopilot +14
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%7479Glasshand +5
Agent ergonomics13%6962Aura Autopilot +7
Security & auth14%5360Glasshand +7
Payments & pricing10%7535Aura Autopilot +40
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%4335Aura Autopilot +8
Transparency & trust7%5552Aura Autopilot +3
Negative events≤1500
TotalC 58.7/100D 50.8/100

Facts side by side

FactAura AutopilotGlasshand
KindMCP serverHTTP API
VendorRamenCodapt Inc.
Hosted endpointno (local only)https://api.glasshand.com/v1
TransportsstdioHTTP
AuthAPI keyOAuth
PricingYour planPaid
x402nono
LicenceApache-2.0Proprietary hosted service under the Glasshand Terms of Service
Tools exposed14none
Read-only variant documentednono
llms.txtnono
Last releasenone2026-10-22
Terms last updated2026-06-25no date given
Privacy policy last updated2026-08-27no date given
Customer content may train modelsyes, with an opt-outnot found in the text
Terms restrict automated accessnot found in the textnot found in the text
Terms restrict benchmarkingnot found in the textnot found in the text
Terms or service can change without noticenot found in the textnot found in the text
Arbitration or class-action waivernot found in the textnot found in the text

How the scores have moved

Verdicts

Aura Autopilot

Apache-2.0, with typed MCP tools carrying read-only hints, a separate process for model-written code and headless exit codes for CI. It is days old with no changelog, needs a TypeSafe Jev key or an Aura account, and by default shares recordings, agent traces and the user's own agent conversation with Ramen unless turned off.

Glasshand

Glasshand gives an agent a real iPhone or Mac through a small OpenAPI-described API with OAuth device sign-in, four scopes, short-lived tokens and idempotency keys on every write and input. Rentals are monthly with no trial, and no status page, SLA, changelog or security programme was found.

Before you call either

Aura Autopilot

  1. Check the game's project is open in the editor with the Autopilot tab connected before calling start_run; otherwise it answers that the agent service isn't running.
  2. When a call answers "still going", call wait until it returns; don't start another tool call meanwhile.
  3. Pass the user's request in their own words with kind verify, goal or explore, and read the INSTRUCTIONS.md that start_run names in parts with Autopilot's read_file.
  4. Tell the user that run data is shared by default and how to turn it off (the Settings page, AUTOPILOT_SHARE=0 or -PlaytestShare=0) before the first run.
  5. Never ask the user to paste a TypeSafe or provider key into the conversation; keys belong in autopilot.env or the Autopilot tab.

Glasshand

  1. Start with POST /auth/device/code and show the person the approval link. Never ask for their email code
  2. Send a new Idempotency-Key for each checkout, cancel or input, and reuse it with the same body when you retry
  3. Crop screenshots to geometry.viewport and send input coordinates from 0 to 1 within that rectangle
  4. If you get control_taken, stop sending input and wait for the person to finish
  5. Check each app's own terms before automating it on the device, since Glasshand leaves that to you

Questions

Which is better for AI agents, Aura Autopilot or Glasshand?

Aura Autopilot scores 58.7 (C) on agent readiness against Glasshand's 50.8 (D), and leads in 5 of 7 scored categories. Glasshand leads on schema & documentation and security & auth.

Do Aura Autopilot and Glasshand need an API key?

Aura Autopilot needs an API key. Glasshand uses an OAuth sign-in.

Can an agent call Aura Autopilot and Glasshand without installing anything?

Aura Autopilot runs on your own machine, with no hosted endpoint listed. Glasshand has a hosted endpoint at https://api.glasshand.com/v1.

Are Aura Autopilot and Glasshand open source?

Aura Autopilot is open source (Apache-2.0). No open-source release is listed for Glasshand.

Other comparisons with Aura Autopilot or Glasshand

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.