What actually shipped, and how fast
Over the past ten days, hashtag.org's changelog shows a steady cadence of releases rather than one big launch: five separate updates dated July 12 through July 21, ranging from a single-fix day (the July 17 WebGL health check) to a 26-commit release on July 20. The July 12 entry alone covered nine new features and six bug fixes tied to the RADIO_TTS_PROVIDER work and MCP architecture improvements, logged as a 20-hour day.
Releases included both new features and corrections bundled together; daily reports detailed hours spent and number of commits for each one. July 19 showed 12 features alongside 6 fixes accomplished in 23 hours. Approximately nine hours were needed to produce the 13 commits, which consisted of 6 features and 7 fixes on July 13. Eight changes occurred over roughly 12 hours on July 18.
It’s a build log shared directly, not an official review of engineering work. We are publishing this record publicly so those interested can see changes as they happen each week.
In plain terms
This section is just the receipts: what got built, on which days, and how long it took, straight from our own public changelog. We're not polishing the numbers, just reporting what we logged.
Discovery infrastructure: super-map, keyword scanning, MCP
The July 21 release focused on the super-map, the geo-spatial layer that lets agents query portals by location and intent rather than crawling pages. Shipped features included enhanced super-map functionality, expanded keyword scanning, in-app reporting configuration, and an admin control room setup, alongside one critical fix, over what the changelog describes as a 14-hour day.
Agents filter portals using hashtag keywords like #coffee, #plumber, and #lawyer, instead of only GPS location; this process is described in the Agentic Internet documentation. This scanning expansion represents a change to how information is found, not simply its appearance. An agent’s map query results are directly impacted by it.
According to the platform brief, an external agent like a Claude instance, Cursor session, or browser agent can use the MCP server to connect to the hashtag network. This allows agents to search, leave leads, buy, book, or claim their own #name presence. The July 12 release included updates to core plumbing for agent connectivity and shipped MCP enhancements along with a new RADIO_TTS_PROVIDER integration.
July 13's release added compute node humanization, SEO security fixes, and new compute marketplace features, extending the builder-facing side of the platform (six feature commits, seven fixes).
In plain terms
This is the part that lets AI agents actually find and use businesses on the platform, by location or by keyword, and the part that lets other AI tools (like Claude or Cursor) plug straight into the network to search it or book things. Think of it as the roads and doors agents use to get in.
GIGI avatars, onboarding, and the creator-facing funnel
July 20 was the heaviest release of the window: 26 commits over a 21-hour day, split into 11 features/improvements and 15 fixes, centered on the GIGI avatar experience and the marketing funnel for creators. Shipped items included a GIGI avatar cache-bust, a marketing funnel update, footer improvements, and an admin analytics fix.
GIGI functions as an AI agent, assisting visitors through chat, voice, and video, using Tavus personas aligned with platform guidelines. It handles bookings and captures leads for portal owners. Recently, a cache-busting fix addressed what seemed to be a stale asset bug rather than introducing new features. This work is occurring concurrently with ongoing development of the creator funnel.
The July 19 release built on that with onboarding and analytics work: an onboarding dropdown added to the homepage, homepage UI fixes, deeper analytics integration into the admin dashboard, and an improved humanizer component, over 12 features and 6 fixes across a 23-hour day. July 18 followed with a new analytics command center and enhanced session replay, plus fixes to OAuth issues and GIGI bugs.
In plain terms
GIGI is the AI 'employee' each business portal can run, answering questions and taking bookings on video or chat. This stretch of work was mostly about making that avatar reliable and making it easier for new creators to sign up and see their own analytics.
Newsroom, app store, and reliability fixes
A release occurred on July 16 and included two new features; a weekly agentic-web newsroom is now live at /news with an article template and cadence. This newsroom represents the publication currently being read. Also part of the update is a centralized app store, which accompanied performance and SEO enhancements logged as a 17-commit, 18-hour day.
A smaller update appeared the next day to address an actual problem for some users; those experiencing issues with WebGL were seeing a black screen where the spatial globe should be. To resolve this, the July 17 release included a check for WebGL functionality and implemented a raster fallback option, ensuring these affected users still see a functional map even if it isn't quite as detailed.
The team made eight adjustments on July 15, including tube channel video strip integration, schema improvements, and authentication fixes. Earlier, on July 14, there were just three commits, footer optimization, orphaned-page linking, and a downloads page update, completed in roughly an hour; that window saw the fewest changes.
In plain terms
This part covers the newsroom you're reading, a new app store, and a handful of fixes so the site actually works for everyone, including a fallback for people whose browsers can't render the interactive map properly.
Where this fits against the rest of the agentic web
Labs and platforms within this newsroom focus on outward-reaching agents: OpenAI’s Operator manages consumer tasks plus e-commerce, while Anthropic's Claude assists enterprise productivity by targeting document workflows. Google DeepMind’s Project Mariner supports consumer browsing and research, and Perplexity AI combines agentic search with automated browsing and shopping; generally, these are agents designed to utilize the existing web.
Within that landscape report, hashtag.org’s profile is listed under ‘Agentic Web Infrastructure,’ a place where agents arrive to read and make transactions. Releases associated with this role include the super-map and keyword scanning, which help portals become findable by location and intent; an MCP server allows any outside agent, even tools from labs, to connect and take action. GIGI provides each portal its own agent for answering inquiries, scheduling appointments, and gathering leads when a visitor arrives, whether that visitor is another agent or a person.
Keyword ranking on hashtag.space functions through an open, on-chain auction utilizing $SPACE; this means visibility depends on a stake, not a concealed algorithm. Reputation and portal identity should be verifiable directly on the blockchain, avoiding reliance on one company’s database. A noticeable trend appears in hashtag.org's traffic data from the last 30 days: bot and AI requests totaled 4.1K compared to just 3.6K human visits. This suggests that automated agents are already a core audience for this infrastructure.
hashtag.org focuses on a different area. OpenAI, Anthropic, Google DeepMind, and Perplexity are improving the agents themselves. This organization builds what those agents require: structured pages they can read. The platform uses its own model. It funds operations by returning proceeds to users through web3 rather than seeking outside investment.
- OpenAI (Operator), Anthropic (Claude computer use), Google DeepMind (Project Mariner), Perplexity AI: agents that go out and browse, compare, and transact on the existing web.
- hashtag.org: #names, geo-pinned portals, GIGI per-portal agents, an MCP server, and open on-chain keyword auctions, the infrastructure that makes a business findable, readable, and transactable when one of those outbound agents arrives.