What changed, and when.

Cursor Origin + Grok Voice Agent API — Cursor Origin (sys://anysphere/cursor-origin) — Anysphere's agent-first code hosting and PR platform, first verified platform entry. Grok Voice Agent API (sys://xai/grok-voice) — xAI's speech-to-speech voice platform with telephony and a no-code builder. Cursor added as a dev-tool integration target.

Icon index + /icons gallery — All 199 agent and vendor marks resolved from each homepage's declared icons (apple-touch-icon, favicon) and GitHub org avatars — browseable and copyable at /icons, served as JSON at /api/v1/icons, and now rendered across cards and spec sheets via the curated index.

Transports registry — the wires agents actually speak — New /transports section maps the channels and wire protocols agents communicate over: agent wire protocols (MCP, A2A, AGNTCY/SLIM, ANP, ACP, AG-UI, AITP), messaging channels (Slack, Telegram, Discord, WhatsApp, Signal, iMessage, email), remote access (SSH, mesh VPNs like Tailscale, HTTPS tunnels, webhooks), and realtime media (PSTN/SIP, WebRTC) — each with the cataloged systems that speak it, tagged only where documented. Exposed at /api/v1/transports and as a 'Speaks' row on every spec sheet.

Learn section — four interactive guides — New /learn section with two interactive guides: the Autonomy Dial (five delegation levels from suggests to holds-a-job, each grounded in live catalog exemplars) Which agent do you need? (a three-question quiz scored against the live catalog), Same task, different anatomy (one task traced through CLI/cloud/IDE/voice/ops agent anatomies with capability-tagged steps), and Anatomy of an MCP call (a 7-step walkthrough of a tool invocation crossing the JSON-RPC boundary).

Capability taxonomy, vendor pages, integration matrix — New atomic capability taxonomy — 28 tags (file editing, browser operation, MCP tools, agent payments, identity, sandbox runtimes…) applied across all 110 systems, filterable in the index and browsable at /capabilities. New /vendors section profiles all 89 vendors with portfolio spec sheets and curated head-to-heads. New /integrations section maps which external systems each agent connects to — GitHub, Slack, Salesforce, Twilio, MCP servers — with the tail of each matrix gated 80/20. /api/v1/systems now exposes capabilities and integrations per system.

Sixth expansion — agent services — New agent-services category covers the substrate agents can't self-provision: payments (x402, AP2, Skyfire, Payman, Nevermined, AgentCard), identity and delegated auth (Keycard, Arcade, Scalekit), tool platform (Composio), compute (E2B, Browserbase), and senses (Exa, Mem0). AgentMail recategorized from agent-communication — an inbox is a service, not a protocol.

x402 added — Coinbase's open protocol reviving HTTP 402 for stablecoin-settled machine payments.

Agent Payments Protocol added — Google's signed-mandate spec for agent payments over existing rails.

Skyfire added — Agent wallets plus KYA identity — a payments network for the agent economy.

Payman added — Agent spend with policy limits and human-in-the-loop approval.

Nevermined added — Agentic-commerce infrastructure — payment plans and metering between agents.

AgentCard added — CLI-issued virtual cards, agent email, and x402 wallet in one package.

Keycard added — Identity and policy control plane for autonomous agents.

Arcade added — Delegated OAuth and prebuilt tools so agents act on a user's behalf.

Scalekit added — MCP/auth stack covering user SSO and agent credentials.

Composio added — 250+ SaaS tools exposed to any agent framework with managed auth.

E2B added — Secure sandboxes where agents run untrusted code.

Browserbase added — Hosted headless browsers with stealth, sessions, and replay.

Exa added — Search index built for agents — meaning-based queries, clean contents.

Mem0 added — Memory layer for agents — extract, store, retrieve across sessions.

AgentMail recategorized — Moved from agent-communication to agent-services — an inbox is a service substrate, not an agent↔agent protocol.

Fifth expansion — agent communication — New agent-communication category covers agent-to-agent work: open protocols (A2A, AGNTCY, ANP, NANDA), platforms and frameworks (BeeAI, CrewAI, Microsoft Agent Framework, LangGraph, MetaGPT, CAMEL, Agency Swarm, OpenAI Agents SDK, AgentMail); OpenScout / Scout recategorized from bridge.

A2A Protocol added — Linux-Foundation open protocol for cross-vendor agent discovery and task exchange.

AGNTCY added — Cisco-originated 'Internet of Agents' stack — directory, SLIM messaging, identity.

Agent Network Protocol added — Decentralized agent messaging spec with DID identity.

NANDA added — MIT Media Lab's Internet of AI Agents registry and index research.

BeeAI added — IBM's open-source multi-agent platform with ACP/A2A interop.

CrewAI added — Role-based multi-agent crews — the most adopted framework in the category.

Microsoft Agent Framework added — AutoGen + Semantic Kernel successor for multi-agent graphs.

LangGraph added — LangChain's stateful graph runtime with supervisor and swarm patterns.

MetaGPT added — Multi-agent framework simulating a software company's assembly line.

CAMEL added — Research-rooted framework for role-playing agent societies.

Agency Swarm added — OpenAI-native agencies with explicit agent communication flows.

OpenAI Agents SDK added — Lightweight agentic framework whose handoffs are first-class agent↔agent.

AgentMail added — Email infrastructure purpose-built for agents.

OpenScout / Scout recategorized — Moved from bridge to agent-communication — agent-to-agent messaging is the product's core.

Fourth expansion — agents beyond code — Five new categories widen the catalog past developer tools: digital employees (Alice, Ava, Piper, Sierra, Decagon, Fin, Harvey, Lindy, Juicebox, Mercor), ops agents (Cleric, Resolve, incident.io, Rootly, Bits AI, Dropzone, Simbian), computer-use agents (Browser Use, Skyvern, ChatGPT Agent, Claude Computer Use, Manus, Stagehand), research agents (ChatGPT/Gemini Deep Research, Perplexity, Genspark, Hebbia, Elicit, AlphaSense), and voice agents (Vapi, Bland, Retell, Synthflow, ElevenLabs Agents, Deepgram). New interfaces recorded: chat, browser, voice. Counts now 66 vendors / 83 systems / 13 categories.

Alice added — 11x's AI SDR — outbound email/LinkedIn prospecting plus Julian on inbound calls.

Ava added — Artisan's AI BDR — contact database, deliverability, and reply handling in one contract.

Piper added — Qualified's inbound AI SDR — website qualification and meeting routing, Salesforce-native.

Sierra added — Customer-service AI employee for consumer brands; outcome-priced resolutions across chat and voice.

Decagon added — Autonomous support agents with ops-team procedure controls (AOPs).

Fin added — Intercom's AI support agent; $0.99 per resolved conversation.

Harvey added — AI associate for legal work — drafting, review, diligence over legal corpora.

Lindy added — Build-your-own-employee platform — inbox, calendar, notes, and custom workflows.

Juicebox added — PeopleGPT recruiting agent — natural-language talent search plus outreach.

Mercor added — AI interviewer and talent marketplace for high-volume screening.

Cleric added — Dedicated AI SRE — investigates alerts and drafts root-cause hypotheses in Slack.

Resolve added — AI SRE from observability veterans — triage, investigation, remediation drafts.

incident.io added — Incident platform with AI summaries, timelines, and similar-incident recall.

Rootly added — Incident-response platform with AI-drafted summaries and retros.

Bits AI added — Datadog's in-platform agent — natural-language observability and incident investigation.

Dropzone AI added — AI SOC analyst — end-to-end alert investigation with evidence chains.

Simbian added — Autonomous SOC agents plus GRC automation.

Browser Use added — Breakout OSS browser-agent framework plus managed cloud.

Skyvern added — Vision-based open-source browser automation — no per-site selectors.

ChatGPT Agent added — OpenAI's hosted computer-use mode — Operator's successor inside ChatGPT.

Claude Computer Use added — Anthropic's GUI-control API capability — screenshots in, mouse/keyboard actions out.

Manus added — Hosted general agent — cloud-sandbox tasks returning finished artifacts.

Stagehand added — Browserbase's OSS act/extract/observe framework — half scripted, half agentic.

ChatGPT Deep Research added — OpenAI's research-agent mode — long browsing runs returning cited reports.

Gemini Deep Research added — Google's research-agent mode with Workspace integration.

Perplexity added — Answer engine with Deep Research and Labs modes.

Genspark added — Super Agent workspace — research that lands as pages, decks, and sheets.

Hebbia added — Matrix analyst agent for finance/legal over licensed and internal corpora.

Elicit added — Research agent for scientific literature — extraction tables over 125M+ papers.

AlphaSense added — Institutional market-intelligence platform with Deep Research agents.

Vapi added — API-first voice-agent orchestration — swap STT, LLM, TTS, telephony per call.

Bland AI added — Enterprise phone-agent platform with compliance posture.

Retell AI added — Voice-agent platform tuned for call quality — latency, barge-in, transfers.

Synthflow added — No-code voice agents — receptionists and outbound for SMBs and agencies.

ElevenLabs Agents added — Voice agents on ElevenLabs' TTS stack — 70+ languages, phone and web.

Deepgram Voice Agent added — Speech-native unified endpoint for real-time voice agents.

Third expansion — personal agents category — New category for self-hosted, chat-first assistant runtimes: OpenClaw and its fork family (NanoClaw, ZeroClaw, PicoClaw, nanobot) plus Nous Research's Hermes Agent. Counts now 34 vendors / 47 systems / 8 categories.

OpenClaw added — The self-hosted assistant gateway — ~29 chat channels, skills, memory, multi-agent routing; MIT, 501(c)(3)-stewarded.

Hermes Agent added — Nous Research's self-improving agent; learning loop, memory, seven backends incl. serverless.

NanoClaw added — TypeScript OpenClaw fork with container-per-agent isolation and Agent Swarms.

ZeroClaw added — Rust rewrite — single binary, ~5MB RAM, 30+ channels, hardware peripherals.

PicoClaw added — Sipeed's Go build for embedded — <10MB RAM, <1s cold start on $10 boards.

nanobot added — HKUDS's small, auditable Python assistant; MCP-native tools.

Second expansion — 9 systems added — New lab CLIs (Muse, Grok, Kimi, Qwen, Mistral), OSS agents (Crush, Kilo), JetBrains Junie, and the Instinct memory bridge. Counts now 28 vendors / 41 systems.

Muse Code added — Meta's terminal agent on Muse Spark 1.2; contributor tier free, API pay-as-you-go.

Grok Build added — xAI's open-source Rust TUI on Grok 4.6; free to try, bundled with SuperGrok / X Premium+.

Kimi Code added — Moonshot AI's single-binary agent; multi-provider incl. Anthropic/OpenAI-compatible.

Qwen Code added — Alibaba's OSS Gemini-CLI-derived agent; IDE plugins, desktop app, daemon, SDKs.

Mistral Vibe added — Mistral's OSS CLI on Devstral 2; IDE extensions via ACP.

Crush added — Charm's multi-model terminal agent; OSS, BYO provider keys.

Kilo Code added — Roo Code successor on a shared OpenCode engine; VS Code, CLI, and cloud agents.

Junie added — JetBrains agent, GA June 2026; IDE-native agentic debugging plus a terminal CLI.

Instinct added — Self-learning memory MCP server for coding agents; OSS, MIT.

Full inventory re-verification — All 32 entries re-checked against vendor pricing pages; counts moved to 20 vendors / 32 systems.

Roomote added — Roo Code team's single-tenant cloud agent; self-host free ≤10 users, cloud from $49/mo.

Augment Code added — Context-engine coding agent; Business $100/mo flat up to 50 seats.

Warp added — GPU terminal with bundled agents; free tier has no agent usage, Pro $20/mo.

Trae added — ByteDance's AI-native VS Code fork; regional pricing differs.

Zed added — Fast Rust editor with BYOK; Pro $10/mo incl. $5 token credit.

Gemini CLI added — Consumer/free path ended 2026-06-18; API-key and Code Assist routes remain.

Google Antigravity added — Agent-first IDE; individual free tier, capacity via Google AI Pro/Ultra.

Kiro added — AWS spec-driven IDE; free 50 credits, Pro $20/mo up to Power $200/mo.

Roo Code retired — Extension shut down 2026-05-15; team moved to Roomote.

Continue retired — Acqui-hired by Cursor ~June 2026; repo read-only, standalone product folded.

Cody flagged active-caution — Consumer tiers ended; enterprise-only, no public price.

Gemini Code Assist flagged active-caution — Free individual tier ended 2026-06-18.

Cursor pricing refreshed — Pro $20 / Pro+ $60 / Ultra $200; Teams Standard $40, Premium $120/user.

Windsurf plans refreshed — Free / Pro $20 / Max $200; Teams ~$80 + $40/seat under Cognition.

Devin plans refreshed — Free / Pro $20 / Max $200 / Teams $80 + $40 seat; old $500 Team plan retired.

GitHub Copilot tiers refreshed — Free / Pro $10 / Pro+ $39 / Max $100; Business $19, Enterprise $39/seat.

Initial seed catalog — 14 vendors, 24 systems across 7 categories generated from first-pass inventory.