# Fix list: Offscreen From Anchor Terminal's listing at https://www.anchorterminal.com/tools/offscreen, the October 2026 research run, assessed 17 October 2026. Grade B, 62.9 out of 100. This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public. For a coding agent working on Offscreen: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published. ## 1. Reliability, 55 out of 100, up to 9 more on the total Why it scored 55: Read with the local-software lines, since the owner runs the CLI, daemon and app. The project is six days old (first commit 11 October 2026). Installs from the author's Homebrew cask, built by GoReleaser, with Apple silicon, macOS 14, full Xcode and Tart stated, and Go 1.27 for a source build (20). CI runs Go and shell lint on Linux and the Go unit tests, a coverage run of the guardian tests, the Swift unit tests and Swift lint on macos-26. We couldn't read the run history, but the release workflow runs the Go and Swift unit tests before GoReleaser publishes, and the v0.0.5 cask was pushed on 16 October from the commit that is the head of main. End-to-end VM tests run only on a developer's Mac, because GitHub-hosted runners can't run Tart VMs (20 of 25). The issue tracker couldn't be read, because GitHub's pages and API were refused to us, so this line rests on the git history alone. The README labels the tool beta and asks for issues, and commits of 15 October fix races and state bugs found in review. Scored 10 of 25 rather than as clean or as absent. Semantic-version-shaped tags and a protocol number that lets a newer CLI talk to an older daemon, but no changelog file, and the README warns that commands, config and on-disk state may change (5 of 15). Version 0.0.5, not 1.0 and labelled beta (0). The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability): Hosted APIs, MCP servers, models and platforms. - 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own). - 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so. - 15, rate limits documented with numbers. - 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved. - 10, an SLA published for any paid tier. - 10, the surface agents use is generally available, not beta or preview. Local packages, SDKs, frameworks and stdio MCP servers. - 20, installs from an official package with supported runtimes stated. - 25, a public CI and test suite, passing on the default branch. - 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered). - 15, semver discipline and breaking changes called out in a changelog. - 15, version 1.0 or later, or declared stable. Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors. ## 2. Security & auth, 62 out of 100, up to 6.7 more on the total Why it scored 62: No credential model to grade, since there is no account or key, so read with the framework lines as for other local agent tools. 30 for what leaves the machine and what is exposed, 20 for least privilege and approvals, 15 for prompt-injection posture, 15 for audit and 20 for the security programme. No network code or telemetry found beyond the owner-only Unix socket and SSH to the guest. Guest SSH uses the base image's `admin` and `admin` login by default, the caller's environment, which can hold secrets, is stored in SQLite while a job is active and dropped when it ends, with old copies possibly left in free pages, which the docs say. Any process that can open the socket is trusted (22 of 30). A fresh VM per job, deleted afterwards, and a VM task gets the worktree and listed products but not the host environment. The host lane is opt-in, gated on idleness, an unlocked session and camera and microphone off, with a countdown and take-back. The skill lists actions to ask the person about first. No egress or network policy for the guest is documented, the whole worktree is copied in, and `[prepare] run` executes on the host even for VM jobs (16 of 20). The tool returns command output and screenshots of the agent's own app, with no guidance on treating text in them as untrusted (8 of 15). Job records, logs and artifacts are kept for seven days by default, `status -a` lists past jobs, and the daemon and image builds keep logs (12 of 15). No SECURITY.md, security.txt, bug bounty or advisories. The binaries and app are signed ad hoc and not notarised, and the cask strips the quarantine attribute, which the install guide states. The release workflow keeps the tap token out of cached tools and doesn't persist git credentials, and CI runs with read-only permissions (4 of 20). The checklist (https://www.anchorterminal.com/benchmark/#checklist-security): - 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option. - 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions. - 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10. - 0 to 15, audit logs or per-call visibility for the operator. - 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public. Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing. ## 3. Schema & documentation, 67 out of 100, up to 5.4 more on the total Why it scored 67: Graded on the CLI and its skill. `status` and `image ls` take `--format json`, and `--progress json` emits build and run phases, but the docs say field names may change before 1.0, no JSON Schema is published, and `run` output is the wrapped command's own. The project contract is validated, with unknown keys rejected, by `offscreen check` (8 of 25). The skill is Markdown in the Agent Skills format, with a 90-word description and a 1,552-word SKILL.md, and the docs are Markdown in the repository (10). The skill says when to use it, sets a core rule never to run `xcodebuild test`, the Simulator or built apps on the host, says plain builds may run on the host, and lists what to ask the person before running (18). Flags are typed by kong, with value sets for `--lane` (vm, host, auto) and `--progress`, durations for `--timeout`, and every task requires `lane` and `run`. Task commands are free shell text by design (11). A worked example app with three tasks and their scripts, an exit-code table with what to do for each, and troubleshooting entries that quote the error text (14). Tags and a daemon protocol number with stated compatibility rules. No changelog file, and the GitHub release notes couldn't be read (6 of 15). The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema): APIs and MCP servers. - 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool). - 10, llms.txt or Markdown docs served for agents. - 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference. - 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs. - 0 to 15, examples and documented error responses. - 15, versioning and a public changelog. Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference. ## 4. Payments & pricing, 60 out of 100, up to 5 more on the total Why it scored 60: No payment protocol (0). Read with the self-hosted rule. Free under MIT with nothing to buy from the author (20). No card or account (20). Installs with one Homebrew command and needs no signup, though a person grants Local Network on the first VM job (20). Tart, which the cask installs, is under FSL-1.1-ALv2, and its licensing page says organisations above 100 host CPU cores need a paid licence. That is a cost of the dependency, named in the pricing notes, and not scored here. The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments): The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/). - 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which. - 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login. - 20, a free tier or trial that doesn't need a card. - 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API). Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied. Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol. ## 5. Agent ergonomics, 74 out of 100, up to 4.2 more on the total Why it scored 74: A CLI adds no tool definitions to context, and only the skill's 90-word description is always loaded. Stdout carries the wrapped command's full output, which for `xcodebuild` can be long, while Offscreen's phases go to stderr and are hidden when stderr isn't a terminal (18 of 25). `status -a`, `-q`, `--format json` and templates, `logs` with `-f`, and artifacts in one directory per job. No control over the size of the command's output (13 of 20). Exit codes separate the command's result from a queue timeout (75), an Offscreen failure (70) and a usage error (1), errors name the fix (`Open it with: open -a Offscreen`, the Tart install command, the Local Network setting), and a failure before the command starts repeats the last 20 lines (18). Each job starts from a fresh clone, so a rerun is clean. `cancel` is safe to repeat, `init` refuses to overwrite, `check` validates without a VM, and `prune` asks first or takes `--dry-run`. Attached queued jobs fail after a daemon restart, and the docs say to detach jobs that must retry safely (16 of 20). Defaults are sensible (VM lane, two slots, no timeout) and a one-off `-- ` needs no task. One CLI, and the daemon's socket API is internal, with no SDK (9 of 15). The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics): - 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries). - 20, pagination, filtering and output-size controls. - 20, actionable, documented error responses, codes and messages an agent can recover from. - 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations. - 15, sensible defaults, few required parameters, and official SDKs in at least two languages. Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs. ## 6. Transparency & trust, 56 out of 100, up to 3.9 more on the total Made of editorial 72, provenance 40. Why it scored 56: MIT in LICENSE and in the release archive (30). No privacy policy, as for other local tools. DAEMON.md and the config reference say what is stored where, that the caller's environment is kept only while a job is active, that deletion is logical and SQLite free pages and backups can keep old values, and that job history expires after seven days by default (22 of 30). No deprecation policy. The README's beta note, the protocol compatibility rules and the install guide's migration steps for the renamed app and daemon identifiers are the only notices, with no dates (8 of 20). No telemetry found in the source, but the README and docs don't say either way (12 of 20). The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency): - 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms. - 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors). - 0 to 20, a deprecation policy or notices with dates. - 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted). The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two. Provenance checks not met in full (half of this category, computed from checked facts): - Domain age: , no registry record we could read (0 of 15) - Status page: not found (0 of 10) - Changelog: not found (0 of 10) - security.txt: not found (0 of 10) ## 7. Maintenance & community, 66 out of 100, up to 3 more on the total Why it scored 66: v0.0.5 on 16 October 2026 (30). Five tagged releases on 15 and 16 October 2026, each published to the cask (20). All 116 commits are by the one maintainer, and pull requests #89 to #91 of 15 October are his own. Outside issues and replies couldn't be read, because GitHub's pages and API were refused to us (8 of 25, from the commit history only). Not an MCP server, no SDK, and the skill is in no registry (0). Current dependencies (Go 1.27, kong 1.14, modernc SQLite 1.60), tool versions pinned in mise.toml and lint in CI. No Dependabot configuration found (8). The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance): - 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older. - 20, at least three releases or dated changelog entries in the last 90 days. - 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15. - 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models). - 10, package health, current dependencies and CI. Models are read for deprecation notice periods and model churn rather than release counts. ## What we couldn't check What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it. - Judgement call on category. Listed in code-sandboxes, not desktop-automation, because its job is running code in a throwaway machine and returning results. It doesn't click, type or read the screen for the agent, and any driving of the app inside the VM is the task's own AppleScript or System Events. `sandbox.persist` is left out because only a per-project cache survives between jobs, not the VM. - The scout dated the first release to v0.0.3 on 15 October. The tags and the cask history show v0.0.1 and v0.0.2 earlier the same day, so `firstReleased` is 2026-10-15 either way. The project was six days old when checked, and its Show HN post went up on 17 October. - Neither of our own products. It doesn't route tasks or handle payments (letme), and it isn't a personal AI platform (LocalGhost). - We did not install or run the tool. Every claim about behaviour comes from the source, the docs, the skill, the workflows, the GoReleaser config and the cask. - unchecked: GitHub issues, pull request discussions, CI run results, release notes and star count, because github.com pages and the GitHub API answered 403 to this session. The repository and tap were read by git clone. - Offscreen's docs say Tart moved to `openai/tools/tart`. The formula there names https://github.com/openai/tart and FSL-1.1-ALv2. Tart's ownership was not researched further, and tart.run's licensing page, which still describes the Fair Source tiers, may predate the move. - The vendor is written as Mark Phelps, the licence holder. The bundle identifiers point to Latent Signal Labs, LLC, whose page links his X account, but the page doesn't name Offscreen, so the company isn't given as the vendor. ## Weaknesses - First commit 11 October 2026, version 0.0.5 and labelled beta, with JSON field names that may change before 1.0 - Apple silicon macOS 14 or later with full Xcode only, at most two VMs at once, and a large one-time image download - Ad hoc signed and not notarised. The cask strips the quarantine attribute, and there is no SECURITY.md or security.txt - No egress or network policy for the guest is documented, and `[prepare] run` executes on the host with the caller's environment - End-to-end VM tests run only on a developer's Mac, because GitHub-hosted macOS runners can't run Tart VMs ## What costs an agent a turn today The notes we give agents before they call it. Each one is a workaround an agent shouldn't need. - Read `.offscreen.toml` before running anything and use its `project` key and task names. Don't guess them - Run `xcodebuild test`, the Simulator and built apps through `offscreen run `, never directly on the host - Treat exit 70 as an Offscreen failure, not a test failure, and run `offscreen version` to see whether the daemon is running - Ask the person before `--show`, `--keep`, the host lane, prune or `image build`, and cancel any kept job when done - Read screenshots and `.xcresult` bundles from the directory `offscreen artifacts ` prints ## What the review panel asked for - published golden image size - a changelog file - Publish a JSON Schema for status output - Add a limit on captured output - guest egress policy - notarised releases ## When it's done Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.