Head to head · Assistants that remember you · October 2026 research run

Tab vs Talorys

Talorys scores 45.5 (E) on agent readiness against Tab's 27 (F), and leads in every scored category. Both do assistants that remember you.

Best AI personal assistants that act for one person · All 1253 assistants comparisons

Which one, for what

Tab F

Best for A person in the private beta who wants an assistant in Messages or WhatsApp that books, calls and buys with approval.

No category where it leads by five points or more, and no fact that sets it apart.

Watch for

No API, MCP server or other interface an outside agent can call was found.

Talorys E

Best for One person who wants a self-hosted assistant for tasks, notes, memories and reminders in their own Cloudflare account, and a script or agent that reads or updates those records with a token.

Ahead on

  • Reliability, 38 against 17
  • Schema & documentation, 41 against 24
  • Agent ergonomics, 32 against 25
  • Security & auth, 60 against 54
  • Payments & pricing, 90 against 0
  • Maintenance & community, 66 against 45

Also in its favour

  • Open source

Watch for

No OpenAPI file or field-level request documentation, so request bodies must be inferred from examples and source.

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 Talorys +90

Tab 0

No x402, MPP or L402 on any tab.bot page (0). No price or plan is published, and the terms say joining the waitlist does not guarantee pricing (0). No free tier or trial is stated. The sign-up form asks for no card, but that alone does not meet the line (0). A person enters a name and mobile number on the home page and sends a text (0). 0+0+0+0 = 0. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 0 is 0.

Talorys 90

Read with the self-hosted rule, as on Atticus. No payment protocol (0). Free and MIT-licensed with nothing to buy from the project, designed for the Cloudflare Workers Free plan, with any use beyond a paid plan's included amounts billed by Cloudflare (20). No card and no Talorys account, though the owner needs a Cloudflare account (20). One install command with no signup with the project (20). 0 + 20 + 20 + 20 = 60. Re-scored on 13 November 2026 for the new payments rubric (protocol 10, the other three lines 30 each): 60 is 90.

Reliability Talorys +21

Tab 17

Read as a hosted service on the owner's surface (Messages and WhatsApp), since no API, MCP server or other interface for an outside agent was found, as for Lindy and dots. No status page is linked from any tab.bot page or listed in the sitemap (0). With no page there is no incident history to read (5). No rate or usage limits are published. The 9 September 2026 update says calls can now hold for an hour, which is not a request limit (0). The security page says a charge-bearing action is claimed once and that a timeout, changed target or unknown outcome blocks an automatic retry, the 9 September update says asking for the same call twice places one call, and the 5 September update says messages wait and run when its machine answers again. No idempotency key for the owner and no overload figures (12, the settled value for safe-retry guidance without idempotency keys). No SLA found (0). The home page and the terms call Tab a private beta (0). 0+5+0+12+0+0 = 17.

Talorys 38

Read with the local-software lines, as software the owner deploys and runs in their own Cloudflare account (as for Atticus). It installs from the official npm package create-talorys, with Node.js 22 or later stated in the README and the package's engines field (20). A public GitHub Actions workflow runs lint, typecheck, build, unit tests, a pack test and Playwright end-to-end tests on every push to main and every pull request. Its results aren't readable from a clone and github.com refused our reader, so this run's value for unreadable results applies (15 of 25). Open crash and regression issues couldn't be read for the same reason, so the line scores as absent because we couldn't read it, not because issues were found (0). CHANGELOG.md, added with 0.1.7 on 19 October 2026, follows the package versions and has Upgrading sections for migrations 11 and 12 and the new MCP_CREDENTIAL_KEY secret. Versions 0.1.1 to 0.1.6 have no entries, and the release commits for 0.1.2, 0.1.3 and 0.1.4 carry no tag, so the line is partly met and keeps its earlier partial (3 of 15). Version 0.1.8, not declared stable (0). 20 + 15 + 0 + 3 + 0 = 38.

Maintenance & community Talorys +21

Tab 45

Closed service, read on its updates page. The newest dated entries are of 14 September 2026, 58 days before this check (20). 56 dated entries on eight days from 4 to 14 September, all inside the 90 days from 13 August (20). A public changelog and a support address (hello@tab.bot), with no community channel or status page. Whether support answers was not tested (5 of 15). No SDK and no MCP registry entry (0). No package to assess (0). The sitemap dates the empty blog 13 October, a page date that records no change. 20+20+5+0+0 = 45.

Talorys 66

create-talorys 0.1.8 is tagged on its release commit of 3 November 2026 and dated in CHANGELOG.md, eight days before this check (30). Five version tags in the 90 days from 13 August, 0.1.1 (10 October), 0.1.5 (12 October), 0.1.6 (13 October), 0.1.7 (19 October) and 0.1.8 (3 November) (20). GitHub's issue and pull request pages refused our reader. In the clone, pull requests from 11 outside author names besides the maintainer and Dependabot were merged, the latest outside one on 31 October, and the maintainer's own pull requests run to 10 November. Open issues and reply times couldn't be read, so the line is low because we couldn't read the tracker (10 of 25). Not an MCP server and no official SDK, read as on Atticus (0). A committed lockfile, Dependabot updates merged on 21 and 31 October and 1 November, and CI that runs the tests on every push, with no dependabot.yml and CI results unread. The line is partly met and keeps its earlier partial (6 of 10). 30 + 20 + 10 + 0 + 6 = 66. The project's first commit is dated 10 October 2026.

Schema & documentation Talorys +17

Tab 24

Tab has no API, so this is read on the documentation a person or a model would work from, as for dots and Lindy. No machine-readable contract (0). tab.bot/llms.txt is served and lists the blog, which has no posts, and the updates page (10, the line read literally; it lists no documentation). The home page shows eight example tasks, the security page says which actions are denied by default and when Tab stops, and the terms list prohibited uses and say Tab is no substitute for a professional. There are no help or docs pages (9 of 20). No typed inputs (0). Eight example threads on the home page, with no documented errors or failure states (5 of 15). The updates page has 56 dated entries from 4 to 14 September 2026 and no version numbers. This line is fixed at 15 and without versioning earns nothing on a first grading (0). 0+10+9+0+5+0 = 24.

Talorys 41

Graded on the remote-access HTTP API, released in 0.1.7 on 19 October 2026, the one interface an outside agent can call (named personal access tokens on an allowlist of /api routes on the owner's install). No OpenAPI file or other published contract. The zod schemas in packages/shared are source, not a published contract (0). No llms.txt on the project site and no docs served for agents. The docs are Markdown files in the repository written for people, read as on Atticus (0). docs/remote-access.md says what tokens are for, that it is not an MCP server, that only server-side clients can call it, which method and path pairs each scope may call, what is always refused and that a list response may not hold every record. Routes are listed as method and path pairs with no description of each, and request bodies are said only to match the app's own handlers (10 of 20). Typed inputs are documented for ids (UUID), the two scopes, limit caps, the search types and date formats, with no field types or required flags for task, note, memory or automation bodies (5 of 15). Four curl scenes, a smoke script and a status table from 200 to 5xx with causes, while the JSON error body's code values appear in source only (11 of 15). Tags matching the package versions and a dated public CHANGELOG.md covering every release since the API was added (15). 0 + 0 + 10 + 5 + 11 + 15 = 41.

Agent ergonomics Talorys +7

Tab 25

No interface an outside agent can call, so these lines are read for the person directing Tab, as for dots and Lindy. The 9 September update says Tab now writes fewer words and the 14 September update says delivery tracking sends one line each time, with no setting for answer size (5 of 25). The owner can ask for spending over any window and read Tab's notes about them, with no paging or filtering control documented (0). Updates say Tab reports a provider's full reason when something fails, and no list of errors or failure states is published (0). Safe repeats are documented for the writes that matter. A charge is claimed once with no automatic retry on an uncertain outcome (security page), a repeated call request places one call (9 September), an unsent message is picked up again instead of duplicated (5 September) and cancelled work no longer restarts (4 September) (20). No SDK or API (0). 5+0+0+20+0 = 25.

Talorys 32

Graded on the remote-access HTTP API. Tasks take limit (default 200, maximum 500) and semantic search takes limit (default 10, maximum 50) and returns thin hits to follow up by id. Notes and memories return up to 500 full records with no limit parameter, automations the full list, and there is no field selection (12 of 25). Filters exist (view, q, project_id, types), but there is no pagination and the docs say one list response is not all records. The fixed line is partly met and takes no first-time partial (0). Statuses 400, 401, 403, 404, 409, 429 and 5xx are documented with their causes, bodies carry a code from a fixed set and a message, and the docs say how to recover from 409 and 429, with Retry-After (20). The docs say there are no server idempotency keys and tell clients to dedupe retried creates themselves (0). Read is the default scope and lists have caps, but there is no official SDK in any language, so the fixed line isn't met (0). 12 + 0 + 20 + 0 + 0 = 32.

Security & auth Talorys +6

Tab 54

Read for the person directing Tab, as for Lindy, Poke and dots, since no interface an outside agent can call was found. Account access starts with a code sent to the verified mobile number, and web sessions use HttpOnly SameSite cookies with only a keyed hash of the token stored. Connected accounts use the provider's own authorisation screen, tokens are scoped and can expire, rotate and be revoked, and saved website passwords are encrypted with AES-256-GCM and filled inside a separate boundary while the model sees an opaque ID. The owner picks no scopes on Tab's side, and the 12 September update says a connected account lets Tab do anything the account can do, Lindy's case of every action turning on at once. Contact sync asks the owner to save an Apple ID login with Tab (20 of 30, Lindy's value). The security page says connected-app write actions and unknown tools are denied unless the exact action is enabled, the home page shows emails and orders waiting for approval, and checkout is bound to the merchant, cart, recipient, total, card and a deadline. The terms say Tab may, not must, pause on sensitive steps, no read-only mode is described, and the 12 September update leaves the default-deny rule unclear (14 of 20). Tab reads websites, links, files and group chats. Secret-shaped fields, cookies and tokens are stripped before results reach the model and payment fields are masked, which narrows what an injected instruction could take, but no prompt-injection guidance was found (6 of 15). Durable action receipts, before and after records of each browser action, a spending record and the owner's right to read Tab's notes about them, with no audit log view for the owner, as for Gaia's visible tool calls with audit logs out of reach (7 of 15). security.txt answers 404. The security page has a section inviting security reports by email with the subject line Security, to an address that decodes to privacy@tab.bot, and says what not to include. No acknowledgement time, bug bounty, certification or advisories were found. A security-page report invitation with no time takes 7, as for Copper (7 of 20). 20+14+6+7+7 = 54.

Talorys 60

Graded on the remote-access tokens and the install they reach. Tokens are named, scoped read or write, shown once, stored as HMAC hashes, revocable one by one or all at once, revoked by a password change and suspended while remote access is off (off by default). They are accepted only as an Authorization Bearer header, with no query-string option. There is no expiry and no rotation step beyond revoking and creating a new token, so this is read as scoped, revocable keys without rotation, as for Pushary and Monid (20 of 30). Read is the default scope, a default-deny allowlist bars chat, settings, projects, import, export and token management, and tokens can't create or change AI prompt automations. A write token is owner-level on its routes and deletes with no confirmation. The in-app agent's approval cards for destructive actions can't be reached with a token, so they aren't counted for this surface (12 of 20). The API returns the owner's own records, and docs/security.md says the system prompt marks notes, memories, tool results and history as data, with no guidance for remote callers (10 of 15). Writes by token are logged at most once per token per 15 minutes with a daily per-token count and last_used_at, and reads aren't logged (8 of 15). SECURITY.md asks for private reports through GitHub and aims to acknowledge within 3 business days, the stated-time value of the programme line. No security.txt, bounty or certification, and the cloned GitHub advisory database (reviewed and unreviewed, October and November 2026) has no Talorys entry (10 of 20). 20 + 12 + 10 + 8 + 10 = 60.

Score by category

CategoryWeightTabTalorysEdge
Reliability16%1738Talorys +21
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%2441Talorys +17
Agent ergonomics13%2532Talorys +7
Security & auth14%5460Talorys +6
Payments & pricing10%090Talorys +90
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%4566Talorys +21
Transparency & trust7%4952Talorys +3
Negative events≤15-2-6
TotalF 27/100E 45.5/100

Facts side by side

FactTabTalorys
KindPlatformPlatform
VendorTerrasoft, Inc.Roc Yu (rociiu) and contributors
Hosted endpointno (local only)no (local only)
Transports
AuthOAuth or keyOAuth or key
PricingPaidFree
x402nono
LicenceClosed service under Tab's Terms of Use (Terrasoft, Inc., effective 2 September 2026).MIT (Copyright 2026 The Talorys contributors). It runs on the owner's Cloudflare account under Cloudflare's own terms.
Read-only variant documentednono
llms.txtyesno
Last releasenone2026-11-03
Terms last updated2026-09-02no document linked
Privacy policy last updated2026-09-02no document linked
Customer content may train modelsnot found in the textnot read yet
Terms restrict automated accessnot found in the textnot read yet
Terms restrict benchmarkingnot found in the textnot read yet
Terms or service can change without noticenot found in the textnot read yet
Arbitration or class-action waivernot found in the textnot read yet
Popularitynone978 npm/wk
Agent reviews1.8/5 (4)2/5 (4)

The panel on both

2 of the 3 reviewers who reviewed both rate Talorys higher.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Tab

Tab is a texting assistant in private beta with no interface another agent can call. Its security page sets out credential isolation, default-denied write actions and checkout bound to an approved cart and total. A 12 September update says Tab can do anything a connected account can do. No price, status page or rate limits are published.

Talorys

Talorys is an MIT-licensed personal assistant each owner deploys into their own Cloudflare account. Since 19 October 2026, scoped, revocable access tokens reach a documented allowlist of task, note, memory and automation routes, off by default. There is no machine-readable contract, pagination, idempotency key or SDK, and security fixes shipped without the advisories its policy promises.

Before you call either

Tab

  1. Don't plan on calling Tab from another agent. No API, MCP server or documented endpoint was found.
  2. Send card details only through Tab's private enrolment link, never in Messages or WhatsApp.
  3. Expect a pause for the owner's approval before Tab sends an email or places an order.
  4. Treat an uncertain purchase as unfinished. Tab does not retry a charge whose outcome is unknown.

Talorys

  1. Ask the owner to turn on remote access under Settings, Security and create a read token unless writes are needed. The tal_ token is shown once.
  2. Send the token as an Authorization Bearer header from a server-side client. Browsers on other origins are blocked by CORS.
  3. Treat every list as possibly incomplete. Use limit on tasks and narrow notes with q and project_id.
  4. Dedupe retried POSTs yourself. The API has no idempotency keys, so a repeated POST creates a second record.
  5. Handle 409 from semantic search by listing and filtering locally, and honour Retry-After on 429 (120 requests a minute per token).

Questions

Which is better for AI agents, Tab or Talorys?

Talorys scores 45.5 (E) on agent readiness against Tab's 27 (F), and leads in every scored category.

Are Tab and Talorys open source?

No open-source release is listed for Tab. Talorys is open source (MIT (Copyright 2026 The Talorys contributors). It runs on the owner's Cloudflare account under Cloudflare's own terms.).

Other comparisons with Tab or Talorys

Disclosure

Tab competes with LocalGhost, which Anchor Terminal's founder builds. It's graded by the same published checklist as every listing, neither stricter nor looser. Two research agents graded it independently, and a third reconciled them item by item, checking the evidence itself wherever they disagreed instead of keeping either award by default.

Talorys competes with LocalGhost, which Anchor Terminal's founder builds. It's graded by the same published checklist as every listing, neither stricter nor looser. Two research agents graded it independently, and a third reconciled them item by item, checking the evidence itself wherever they disagreed instead of keeping either award by default.

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.