Capability, pricing and access changes, and notable events. Each entry shows its date and sources. We only use the word "breaking" when the primary source's own date is within 48 hours. Unconfirmed reports are labelled with a confidence level and are never presented as fact.
2026-10-02 · Guide
What OpenAI's documentation says about the Agents API public beta: sessions, sandbox options, billing and data limits.
Confidence: Vendor documentation, attributed; public beta · Sources: developers.openai.comhelp.openai.comopenai.com
2026-10-02 · Guide
How Poke's documentation says recipes and custom MCP servers connect apps, and how to disconnect or refresh them.
Confidence: Vendor documentation, attributed · Sources: poke.compoke.com
2026-10-02 · Agent update · Lucas
Lucas's FAQ lists a free tier, Pro at $20/month, Max at $100/month and a seven-day reverse trial. Lucas now shows as having a free tier in the price filter. Numeric limits are not published.
Confidence: Vendor FAQ, attributed; not tested · Sources: meetlucas.ai
2026-10-02 · Agent update · Comma
Comma lists Value at $0/month, Pro at $80/month shown as billed yearly, and Enterprise as custom. Credits are priced at 1M per US$1 and do not roll over. Comma now shows as having a free tier in the price filter.
Confidence: Vendor pricing page, attributed; not tested · Sources: comma.surf
2026-10-02 · Agent update · Poke
Poke lists Free at $0, Pro at $19/month and Ultra at $199/month, billed monthly. No numeric limits are shown. Poke now shows as having a free tier in the price filter.
Confidence: Vendor pricing page, attributed; not tested · Sources: poke.com
2026-10-02 · New agent · Wabi
Wabi 2.0 is an iPhone messenger that handles plans and errands and builds small apps in chat, per its App Store listing. TechCrunch (29 Sep 2026) reports it is invite-only. Pricing is not published on the pages we read. Not tested.
Confidence: Vendor listing plus trade press · Sources: apps.apple.comtechcrunch.comwabi.ai
2026-10-02 · New agent · szn
szn is an iMessage-first personal agent from a former Siri and Apple Intelligence engineer. Its site describes calls, emails and follow-ups handled from one thread. Press from 27-28 Sep 2026 says it launches in October; price, countries and approval rules were not disclosed. Not tested.
Confidence: Vendor site plus trade press; launch date is press-reported · Sources: theszn.aiwww.ithinkdiff.comwww.macobserver.com
2026-10-02 · News · Anthropic
Anthropic's blog says mods are small TypeScript functions, shipped inside plugins, that can rewrite a prompt, block or retry a tool call, approve or deny a permission request, or change the interface. They work in the CLI and desktop app. Anthropic warns mods are not sandboxed and run with the same machine access as Claude Code, so install only from trusted sources. On Team and Enterprise plans a built-in sec-default mod loads first to stop user-installed mods overriding permission deny rules. Vendor description, not tested by us.
Confidence: Primary source (Anthropic blog); the page shows no date in our fetch, trade press dates it 1-2 Oct 2026. Vendor claims attributed. · Sources: claude.com
2026-10-01 · Site note
Sixteen agents at launch, one coding-agent guide and a comparison tool are live. Every record carries its sources and the date it was checked. Nothing has been hands-on tested.
Confidence: Not applicable
2026-10-01 · Guide
A documentation-led guide comparing what each vendor says about access, approval, retention and deletion.
Confidence: Vendor statements, attributed
2026-10-01 · Guide
Six jobs (coding, personal admin, errands, customer service, team workflows, building your own) mapped to the agents whose vendors document them.
Confidence: Vendor statements, attributed
2026-10-01 · Guide
Errand agent versus Task-based personal agent: approvals, spending, published pricing and what is still unknown.
Confidence: Vendor statements, attributed · Sources: wajo.aicomma.surf
2026-10-01 · Guide
Eight questions to ask before connecting accounts, with what 16 vendors say about approvals and data, and where they say nothing.
Confidence: Vendor statements, attributed
2026-10-01 · Guide
What agents are, how they differ from chatbots, how they work, what they are used for, what can go wrong, and where to start.
Confidence: Explainer; product statements attributed
2026-10-01 · Site note
Filter by free tier or open source documented, published USD price verified, usage-based, or price not verified. Unknown prices stay labelled as unknown.
Confidence: Not applicable
2026-10-01 · Guide
Which agents document a free tier or are open source, what is capped, and which prices we could not verify.
Confidence: Vendor statements, attributed
2026-10-01 · Guide
New app versus existing repository, how each lets you review changes, and published pricing, with a way to test them yourself.
Confidence: Vendor statements, attributed · Sources: docs.replit.comclaude.comdocs.devin.ai
2026-10-01 · Guide
How Microsoft and Salesforce price agent work (credits, conversations, licences) and what to ask about governance.
Confidence: Vendor pricing pages, attributed · Sources: www.microsoft.comwww.salesforce.com
2026-10-01 · Guide
Other agents in the directory grouped by what you want done, with published prices and caveats.
Confidence: Vendor documentation, attributed · Sources: help.manus.im
2026-10-01 · Guide
What LangChain's docs say LangGraph is, and when a framework or a ready-made agent is the shorter path.
Confidence: Vendor documentation, attributed · Sources: docs.langchain.com
2026-10-01 · Guide
What Intercom says counts as a billable Fin outcome and which handoffs are not charged.
Confidence: Vendor documentation, attributed · Sources: www.intercom.com
2026-10-01 · Guide
Who can run, edit and share, write-action approvals, connector constraints and API triggers, per OpenAI's help article.
Confidence: Vendor documentation, attributed · Sources: help.openai.com
2026-10-01 · Agent update · Claude Code
Anthropic's release notes for v2.1.287 (published 1 October 2026 18:00 UTC) say plugins can now modify deeper behavior (Claude Mods), add a built-in mod where a side agent flags things you or Claude might miss (for first-party sessions with telemetry on), and add URL prompts from MCP servers. It also lists a fix so a dangerous rm keeps its always-ask safeguard when the command also redirects output. If an MCP server stops connecting after the update, the notes say to add "bareElicitationCapability": true to its config entry.
Confidence: Primary source (vendor release notes) · Sources: github.com
2026-10-01 · Guide
What Meta's own safety post says about Sentinel approvals, credentials, purchases and training data, with the caveats it states.
Confidence: Vendor documentation, attributed · Sources: research.meta.ai
2026-10-01 · Guide
Devin's own billing documentation on Free, Pro, Max and Teams, full and flex seats, the $80 Teams minimum and on-demand credits.
Confidence: Vendor documentation, attributed · Sources: docs.devin.ai
2026-10-01 · Guide
How Anthropic's help article says usage is metered and the five habits it recommends.
Confidence: Vendor documentation, attributed · Sources: support.claude.com
2026-10-01 · Guide
How Replit's documentation says Agent work is charged: effort-based checkpoints, Plan Mode, plan features and spend controls.
Confidence: Vendor documentation, attributed · Sources: docs.replit.com
2026-10-01 · News · JetBrains
JetBrains says the Air plugin (JetBrains Marketplace, or native in 2026.3 EAP builds) runs agents you already use, such as Codex, GitHub Copilot, Junie and Cursor, in parallel sessions inside the IDE. It says Air ships with no agents installed and sends no data until you sign in to an agent or a JetBrains AI plan. Free Junie Lite runs are offered during the EAP. Early access, vendor description, not tested by us.
Confidence: Primary source (vendor blog, 1 Oct 2026); early access, vendor claims attributed · Sources: blog.jetbrains.com
2026-10-01 · Guide
What Microsoft's documentation lists for admins: data policies, credit caps, audit logs, security scans, Agent 365 and the Lockbox exclusions.
Confidence: Vendor documentation, attributed · Sources: learn.microsoft.com
2026-09-30 · Agent update · Claude Code
Anthropic's release notes for v2.1.286 (published 30 September 2026 19:10 UTC) add an "N of M" count to stacked permission prompts, and fix several session-resume and login issues. Relevant if you review approvals often.
Confidence: Primary source (vendor release notes) · Sources: github.com
2026-09-30 · New agent · DoorDash
DoorDash says you can now text it what you want, and it finds options, builds the cart and checks out in the text thread, using your past orders. Access is a US waitlist beta. DoorDash also announced a connector for company AI tools to order team lunches. Announced 30 September 2026; the assistant is now in the directory.
Confidence: Vendor announcement plus trade press · Sources: about.doordash.comtechcrunch.comabout.doordash.com
2026-09-30 · New agent · Mindtrip
Mindtrip says its assistant learns traveler preferences, remembers them between trips and can run agents that watch prices and options. This comes from Mindtrip's own press release dated 30 September 2026, including its beta figures, which are unaudited. No price was stated. Now in the directory.
Confidence: Vendor press release only · Sources: www.prnewswire.com
2026-09-29 · Agent update · Claude Code
Release notes for v2.1.285 (29 September 2026) add an allowedProviders managed setting to limit which API providers a machine may use, and an environment variable to disable the WebFetch tool.
Confidence: Primary source (vendor release notes) · Sources: github.com
2026-09-29 · New agent · Manus
Manus documents a rebuilt Manus 2.0 (Cascade engine, Manus Studio desktop workspace, Cloud Computer, Automations, Remote Control) and Cue, an invitation-only app where each agent has its own email, phone number, wallet and computer. Manus says its efficiency figures come from one tested configuration, not an average. InfoWorld reported the launch on 29 September 2026. Cue is now in the directory.
Confidence: Vendor documentation plus trade press · Sources: help.manus.imwww.infoworld.com
2026-09-29 · Agent update · Devin
Cognition says OpenAI's GPT-6.1 Sol is live in Devin Desktop and Devin CLI. The cost and score figures in the post come from Cognition's own FrontierCode benchmark and are not independent.
Confidence: Primary source (vendor blog); benchmark claims are the vendor's · Sources: devin.ai
2026-09-29 · Agent update · Muse
Meta's announcement of 29 September 2026 says Muse, its personal agent for the US and Canada, now has skills and connectors for running a business, including Asana, Box, Canva, Dropbox, Figma, QuickBooks, Klaviyo, Notion, Shopify, Slack, Stripe, Zoom and Facebook and Instagram business accounts. Meta says nothing publishes, sends or spends without your approval, and that Muse is free for most of what people need, with paid subscription plans for more. Those are Meta's statements; we have not tested them, and it does not give plan prices.
Confidence: Primary source (vendor announcement) · Sources: about.fb.com
2026-09-28 · Agent update · Claude Code
The v2.1.284 release notes (28 September 2026) say Claude Sonnet 5.5 is now the default Sonnet model on the Anthropic API, with 1M context and $2/$10 per million tokens, cache reads $0.20 per million. These figures are Anthropic's, as stated in the notes.
Confidence: Primary source (vendor release notes) · Sources: github.com
2026-09-25 · Agent update · Replit Agent
Replit says it has acquired Atta and that you can ask Replit Agent to visualise data, with interactive charts in chat. Replit describes it as aimed at business analysis without writing SQL.
Confidence: Primary source (vendor blog) · Sources: replit.com