The September agent-tools wave, filtered: seven launches, two worth your time

Seven agent infrastructure tools launched within about 48 hours, on September 3 and 4, 2026. The list comes from an AI-generated industry digest, so treat it as a lead, not a source: every product below was checked independently against its own documentation, repository, or launch page. We ran no live tests and held no keys. What follows is documentation-based filtering, and the ranking is earned by verification depth: real docs, real repos, and real pricing pages beat landing pages and unverifiable entries.

Worth your time now

Mireye: location infrastructure for agents. Mireye sells a single API and MCP server that returns sourced geographic data for any US coordinate: 366 indexed fields drawn from 85 authoritative sources, mostly US federal data (USGS, FEMA, NOAA, USDA, BLM, EPA, EIA), plus Overture Maps and Sentinel-2 imagery. The company is in Y Combinator's Summer 2026 batch.

What earns the pick is the provenance design. Every field in every response carries source name, source URL, a fetched timestamp, and a confidence level. Each field also returns one of three statuses: ok, absent, or failed. That distinction is the interesting part: "absent" means a flood zone genuinely does not exist at a coordinate; "failed" means the county never mapped floods at all. For an agent making a downstream decision, that gap is the difference between a correct no and a hallucinated one.

Pricing is published: a free tier of 5,000 credits per month with no card required, paid plans from $19 per month, and no overage billing (usage stops instead of accruing charges). There is also on-demand indexing: request a field that does not exist, and the company's long-running agent researches, builds, and validates it, usually within a day. The field counts, customer anecdotes (a Fortune 500 insurer that abandoned underwriting agents, data-center site selection 100x faster), and the on-demand indexing speed are vendor claims. The docs, the pricing page, and the YC listing are real.

Agent Looker: a trust layer for agents with web access. Built by Gogolook, the company behind the Whoscall caller-ID service, Agent Looker intercepts URLs and page content before they reach an agent: checking URL safety before access and scanning text for prompt injection, phishing, and malware after retrieval. It ships as an MCP server plus hooks and skills for Claude Code and Claude Cowork, and the plugin repository is public on GitHub under GPL-3.0 with detailed installation docs. The product site is live; the service itself is in private beta.

What earns the pick is that it is real code addressing a real failure mode. Prompt injection through browsed content is the attack agents face most often, and a pre-action scanning layer with URL checks and content checks as independent layers is a credible shape for the defense. The effectiveness claims are vendor claims, untested by us, and beta access is gated. But the repo, the docs, and the company behind it all check out.

Verified, but narrower than claimed

Tabbit AI: an AI-native browser with an agent CLI. Tabbit is a free public-beta browser for macOS and Windows with multi-model chat and an Agent Mode that performs web tasks. Its September 18 CLI quick-start docs show how external agents (Codex, Claude Code, WorkBuddy) can drive the browser through a /tabbit skill while reusing your signed-in sessions. It is a real product with real install documentation. It is narrower than the wave framing suggests because it is a browser, not agent infrastructure, and its privacy claims ("browsing activity is never used for training third-party public models without permission") are vendor claims.

MagiCrew: an open-source agent-team platform. MagiCrew launched on Product Hunt on September 3 with a self-hosted community edition, a cloud service, and an enterprise tier. The team says the full stack is on GitHub and runs locally via docker compose with bring-your-own keys through any OpenAI-compatible API. The caveat: the license is "Apache 2.0 with some additional terms," in the makers' own words, not plain Apache 2.0. Read the LICENSE before building on it. The product scope is also unusually broad (agent teams, workflow engine, messaging, collaborative office), which is a smell for builders evaluating focus.

Nex: a GTM app, not agent infrastructure. Nex launched on Product Hunt in early September as "Claude Cowork for high-volume GTM workflows": CRM cleanup, list qualification, and revenue recovery at volumes the makers say break general-purpose agents. The founders are ex-HubSpot, and HubSpot founder Dharmesh Shah backs the company, per their own Product Hunt post. It is a real product with a real team. But it belongs in a sales-tech evaluation, not an agent-infrastructure one. If you run GTM, look; if you build agents, skip.

Could not be verified

ARBR was described as a self-hostable, provider-neutral gateway with an OpenAI-compatible endpoint. No site, documentation, or repository under that name could be found. The gateway space is crowded with open alternatives, but none of them is ARBR.

Grove was described as a shared terminal where a developer and an agent work in the same session. Several unrelated GitHub repositories share the name, all workspace managers for coding agents; none matches that description.

Both are cut from the recommendations. They appear here so you do not waste time searching for them.

Correction (2026-09-26): ARBR is real. Since publication we found project-arbr/arbr-control-plane on GitHub: a fully documented open-source AI gateway - self-hosted, provider-neutral, OpenAI-compatible endpoint, routing, budget enforcement, prompt-injection protection, observability - matching the digest description verbatim. It does not belong in the unverifiable bucket; our original check missed it. Separately, bearlike/grove is a closer match than we acknowledged for the digest's "Grove" (its "terminal twins" are headed tmux sessions shared by developer and agent, arguably the digest's "shared terminal"), but it's fundamentally a multi-agent workspace orchestrator and the digest description still can't be confirmed, so we keep it out of the recommendations.

Update (2026-09-27): Grove is real after all. "Grove: one terminal for you and your AI agent" (grove.publishstudio.one, by Publish Studio) is a macOS terminal app - $5 one-time, 3-day free trial, macOS 15 and later - where the developer and the agent share live terminal processes through a menu-bar popover, a CLI (grove list/start/logs), and a built-in MCP server. That matches the digest's "shared terminal where a developer and an agent work in the same session" description, so the original "could not be verified" entry was wrong, and the bearlike/grove note in yesterday's correction is superseded: that workspace orchestrator is a different product with the same name. Grove belongs in the "verified, but narrower than claimed" tier - a macOS-only paid utility, not agent infrastructure. The "The decision" closing line calling ARBR and Grove "names in a digest" no longer applies to either: ARBR was confirmed on September 26, Grove today.

The decision

If you are building agents that touch physical places, Mireye's free tier is the cheapest way to test whether grounded location data beats prompt-based guessing. If your agents browse the web, Agent Looker is worth a beta slot for the pre-action scanning layer. Tabbit is worth a look if you want agents to act through your logged-in browser sessions rather than headless automation. MagiCrew deserves a license read before any adoption decision. Nex is a sales product, not a builder tool. ARBR and Grove are, as far as we can verify, names in a digest.

The wave's four themes (multi-model cost reduction, agent security, specialized data access, developer-agent collaboration) describe a real maturation of the agent stack. But a wave with two unverifiable entries in it is also a reminder to check the product before the narrative.

What remains unknown

• Whether ARBR and Grove exist as described, or at all.
• The accuracy of Mireye's 366 fields and whether on-demand indexing holds up at volume. Both are vendor claims.
• Whether Agent Looker's detection actually stops real prompt-injection and phishing attempts. No independent test exists.
• Performance and pricing details for MagiCrew's cloud and enterprise tiers, and the exact terms of its modified open-source license.
• Whether Nex's "high-volume GTM workflows" claim holds outside its makers' own anecdotes.
• Nothing in this briefing was independently tested. Every capability claim above is documentation-based.