# How Agentlist works > Scope, evidence rules and the contribution process behind 52 personal-agent records. Edition 1; sources checked 2026-10-21 to 2026-10-26. ## What counts as a personal agent A personal agent maintains context about one person over time and uses tools to act on their behalf. Three properties decide inclusion: it remembers (preferences, people, projects or a durable workspace survive one task), it can act (tools, apps, a browser or a computer, under a person's direction), and it can return (schedules, events or continuing goals bring it back between conversations). Hosted products, local apps and self-hosted runtimes are all in scope. A new version of a product updates the existing record rather than adding one; aliases keep the old name searchable. A separate record exists only for a distinct product. Coding harnesses are catalogued as adjacent and hidden by default: the distinction is their product scope, not whether they have any persistence. - **Catalogue types:** `managed` Hosted agent; `runtime` Self-hosted runtime; `local` Local-first app; `watchlist` Under review; `adjacent` Coding tool (adjacent) ## Evidence rules Every source carries an ID, a title, a direct URL and the date it was checked; every claim references those IDs. Product facts come from official product pages, documentation, repositories, app listings or announcements, paraphrased rather than quoted. Where sources disagree, the product's own official page wins over third-party coverage, which goes stale as products reprice and rename — except for facts a vendor has reason to shade (funding, certifications, independent feedback), where a primary filing or the certifying body is the better source. A claim nobody has established is omitted from the record and filled in by the build as `unknown`. Descriptions are not independent performance tests. ## What documented, limited and unknown mean - `documented` — Supported by official documentation; not independently tested. - `limited` — Only partial or conditional support established. - `unknown` — Not established by reviewed sources; does not mean unsupported. Capabilities that look similar are kept apart: a public API, an MCP connector, browser automation, a password vault and a saved browser session are five different answers. Computer control is three independent questions, and a cloud computer says nothing about whether the agent can touch your desktop. ## Dates Source dates are the day the source was checked, per agent, so re-checking one product never restates the freshness of the rest. The catalogue reports the span of those dates rather than a single date. Launch events accept year, month or day precision (`2026`, `2026-02`, `2026-02-03`) and a day is never invented. The earliest non-update event is the launch date. A stated future ship date is recorded as an expected date on the availability object, never as a launch. - **Launch kinds:** `launch` Public launch; `preview` Research preview; `beta` Public beta; `announcement` Announcement; `release` Open-source release; `update` Major update ## Companies and funding The organization names the team that makes the product; the company object records its type and stage separately from product availability. Types are editorial groupings, not rankings — open-source software can still come from a funded startup. Funding is a dated, disclosed round at company level: round amount is kept apart from total raised and from valuation, and committed or pre-acquisition capital is labelled. Stage and funding are absent when not established, and each carries its own evidence. - **Company types:** `public` Public company; `lab` Major AI lab; `startup` Startup; `established` Established company; `independent` Independent project; `research` Research group ## The comparison fields 33 fields in 6 groups. ### Company & launch Who makes it and when it arrived. Company type and funding are context, not a quality score: an open-source project and a public company carry different risks, not better or worse ones. - **Company type** (`companyType`) - **Company stage** (`companyStage`) - **Disclosed funding** (`funding`) - **Launch date** (`launchDate`) ### What it can do What this agent can actually operate — a machine of its own, yours, or neither. Computer control is three separate questions, and most products answer only some of them. - **What stands out** (`uniqueFeatures`) - **Computer use** (`computerUse`) - **Controls your computer** (`onYourComputer`) - **Its own computer** (`agentComputer`) - **Works while your computer is off** (`whileYouAreOff`) - **Browser control** (`browserUse`) - **Voice & ambient listening** (`voice`) - **Background work & triggers** (`scheduling`) ### Apps & access Where you reach it and on what. A messaging integration is not a native app, so the two are tracked apart; an app listed here is a real first-party client. - **Native desktop apps** (`desktopApps`) - **iPhone & Android apps** (`mobileApps`) - **Web & messaging** (`channels`) - **Messaging services** (`messaging`) - **Native hardware** (`hardware`) - **Developer API** (`api`) ### Models & plans What you pay and what runs underneath. Where an agent is bundled into a larger subscription, the price shown is that subscription's — there is no separate fee for the agent. - **Models & providers** (`models`) - **Bring your own model** (`byom`) - **Included in another plan** (`bundledPlan`) - **Price & usage costs** (`pricing`) ### Vendor lock-in How hard it is to leave, and what survives if the company does. Almost no hosted vendor publishes a shutdown commitment, so an answer here is usually read from the architecture rather than from a promise. - **If the vendor stops** (`worksWithoutVendor`) - **Remote shut-off** (`remoteDisable`) - **What you keep** (`dataAfterExit`) ### Ownership & control What it asks before acting, what it remembers, and who else can see it. These are the questions that decide whether you would hand it your inbox. - **Open source & license** (`license`) - **Where it runs** (`hosting`) - **Passwords & credentials** (`passwordManager`) - **Approval before actions** (`approvals`) - **What it remembers** (`memory`) - **Edit & export your data** (`portability`) - **Privacy & isolation** (`security`) - **Availability & access** (`availability`) ## Structured vocabularies - **Availability states:** `available` Available; `waitlist` Waitlist; `reservation` Reservation; `announced` Announced; `unverified` Access unverified - **Availability restrictions:** `plan` Requires a paid plan; `region` Limited regions; `age` Age restricted; `rollout` Gradual rollout; `invite` Invite required - **Computer control · Controls your computer:** `full` Controls your computer; `partial` Limited access to your computer; `none` No access to your computer - **Computer control · Its own computer:** `cloud` Has its own cloud computer; `local` Runs on hardware you own; `none` No computer of its own - **Computer control · Works while your computer is off:** `yes` Keeps working; `no` Stops; `partial` Partly continues - **Watching the agent's own screen:** `takeover` Watch and take over; `view` Watch only; `none` No remote access - **Vendor lock-in · If the vendor stops:** `yes` Keeps working; `partial` Partly keeps working; `no` Stops working - **Vendor lock-in · Remote shut-off:** `none` Cannot be switched off remotely; `possible` Can be switched off remotely; `required` Requires a licence or activation check - **Vendor lock-in · What you keep:** `self-serve` Export it yourself; `on-request` Export on request; `partial` Partial export; `none` No export described - **Hardware states:** `available` Shipping hardware; `optional` Optional hardware; `announced` Announced, not shipping; `none` No hardware; `unknown` Not established - **App and messaging statuses:** `available` Available; `beta` Beta; `preview` Preview; `announced` Coming soon; `limited` Limited availability - **Native platforms:** `watchos` Apple Watch; `web` Web; `macos` macOS; `windows` Windows; `linux` Linux; `ios` iOS; `ipados` iPadOS; `android` Android; `mobile` Mobile app - **Messaging services:** `imessage` iMessage; `rcs` RCS; `sms` SMS; `whatsapp` WhatsApp; `telegram` Telegram; `signal` Signal; `slack` Slack; `discord` Discord; `teams` Microsoft Teams; `wechat` WeChat; `wecom` WeCom; `feishu` Feishu; `matrix` Matrix; `email` Email - **How open a data source is to third parties:** `open` Open to third parties; `gated` Open, with review; `portability` EU portability only; `closed` First-party only; `none` Reachable by nobody - **How an agent reaches a data source:** `first-party` First-party; `official-api` Official API; `portability` EU portability; `automation` Browser automation; `none` Cannot reach it - **Ecosystem relationships:** `built-on` Built on; `fork` Code fork; `renamed` Renamed - **Update types:** `release` Release; `launch` Launch; `research` Research; `availability` Availability change A messaging integration never establishes a native app, and a listed product is never a shipping product: announced hardware, a paid reservation and an open download are three different answers. ## Discussion and feedback Discussion entries link original dated threads with a short summary, and separate company or creator posts from independent community discussion. The sentiment on a thread describes that thread, not every user of the product. Aggregate feedback snapshots are separately attributed, with provider, collection window, source channels, rated user-week counts and the raw positive share. The displayed percentage is that raw share, not the provider's composite score, and an area breakdown names the provider build it came from because those numbers change with each build. Snapshots from different dates are never mixed. ## Updates and the ecosystem graph Dated updates and relationships are curated: collection gathers candidates into an ignored review inbox and nothing publishes automatically. The ecosystem graph carries only runtime dependencies, code forks and renames; inspiration, shared model providers, similar names and competing alternatives stay out of it, and fork ancestry needs explicit evidence. **Open research queue:** Copilot Autopilot (private preview, tenant-scoped); OpenClaw rename history; Verified Git fork ancestry; Permissions and memory export tests. ## Contributing - One JSON file per agent in `data/agents/.json`; the filename matches the stable, lowercase, hyphenated ID. - Run `npm run add:agent -- example-agent "Example Agent" https://example.com`, or copy `templates/agent.json`. The generator never overwrites an existing agent. - Fill in the summary, company context, launch dates, claims, plans, platforms, messaging and sources, each claim referencing its source IDs. - Save an official PNG, SVG or WebP mark in `src/logos/` and record its original URL and the official page. Logos belong to their owners. - Run `npm run verify` — build, data validation and the domain tests — and open a pull request. Maintainers review before merge; CI runs the same checks. Repository: https://github.com/sunnya97/agentlist. Generated `dist/` files, including this one, are not committed. ## Independence and licence Disclosure: Project creator affiliated with Ambi; uniform evidence rules and an editorial featured default order. Code and original catalogue contributions are MIT-licensed. Product marks remain the property of their owners, and third-party source text, linked posts and feedback data keep their original rights and attribution. Credit this catalogue as "Agentlist" with a link when you reuse it. Entry point for machines: https://agentlist.wiki/llms.txt. Everything in one file: https://agentlist.wiki/llms-full.txt. Structured dataset: https://agentlist.wiki/data.json.