{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "Blog · Opinion · 30 October 2026",
    "lede": "Quynh's post on seven fast-growing GitHub projects named tools that were a week old and put their numbers in a table, and it has had more views than anything else on the blog. I went looking for five more gaps like that, new this month, with people already searching and nobody lining them up side by side.",
    "readingMinutes": 6,
    "words": 1401
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/agent-tools-nobody-compares-yet",
    "json": "https://www.anchorterminal.com/blog/agent-tools-nobody-compares-yet.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/agent-tools-nobody-compares-yet.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/agent-tools-nobody-compares-yet.min.md"
  },
  "markdown": "By Vlad Cealicu, 30 October 2026. 1,401 words, about 6 min to read.\n\nOn Friday 30/10/26 we put a view counter in the footer of every page on Anchor Terminal, and the number that stood out was on Quynh's post about [seven fast-growing GitHub projects](/blog/github-trending-agent-workflows). It shows 5,354 views in the last 30 days, well ahead of the other posts (the counter is a day old and still counts some scrapers as people, so I'm reading it as a ceiling).\n\nReading it again next to mine, I think the reason is simple. It names seven tools that were a week old, every name is something people search for with almost nobody else answering yet, and it puts their numbers in a table. Most of my posts argue about things bigger sites already cover.\n\nSo I went looking for the same kind of gap in agent tooling, things that are new this month, that people already look for, and that nobody compares side by side. I found five. For each one, here's what it is, who's in it, and how much of it we list (not much yet).\n\n- 20: products that support Open Responses or were its launch partners, 9 listed\n- 22: places to find Agent Skills, none listed as a marketplace\n- 0: data-centre GPUs you can rent in orbit today\n- 16: full-duplex voice systems that run locally, 1 listed\n- 31: tools that evaluate agents, 10 listed\n\nWhat we found on 30 October 2026, and how much of it Anchor Terminal lists. Sources in each section.\n\n## 1. Open Responses\n\nOn 22/10/26 OpenAI gave Open Responses to the Agentic AI Foundation, a fund run by the Linux Foundation, and Anthropic gave it Agent Skills on the same day [[1]](#ref-1). Open Responses is an open spec for talking to any model provider, based on OpenAI's Responses API. The current version is dated 24 April 2026, and anyone can run its acceptance tests against an endpoint [[2]](#ref-2) [[3]](#ref-3).\n\nWhat someone building an agent wants to know is who supports it, and today that answer is spread across a few dozen vendors' docs. We counted 14 products that say they support Open Responses by name. Six more were launch partners, or have their logo on the spec's site, but their own docs only say \"OpenAI Responses API\".\n\n| Who | What it is | What their docs say |\n| --- | --- | --- |\n| [Vercel AI Gateway](/tools/vercel-ai-gateway) | Gateway | Text, streaming, tools, structured output and reasoning [[4]](#ref-4) |\n| Databricks | Model serving | Streaming, function tools and web search, stateless, no structured output while streaming [[5]](#ref-5) |\n| Portkey | Gateway | \"fully Open Responses compliant\", across 60+ providers [[6]](#ref-6) |\n| [LM Studio](/tools/lm-studio) | Local app | Compatible from version 0.3.39 [[7]](#ref-7) |\n| OGX (formerly Llama Stack) | Self-hosted server | 93.8% on its conformance badge [[8]](#ref-8) |\n| [Ollama](/tools/ollama) and [OpenRouter](/tools/openrouter) | Local, router | Stateless only, and their docs don't name the spec [[9]](#ref-9) [[10]](#ref-10) |\n\nI couldn't find Open Responses in the docs of LiteLLM, Groq, Together, Fireworks, Cerebras, Mistral, Anthropic, Gemini or Azure on 30/10/26. Groq, Fireworks and Azure have a Responses API of their own and don't say whether it follows the spec.\n\nThat's a compatibility table, the same kind we already keep for x402, and as far as I can tell nobody keeps it. We'll start one, with the acceptance tests as the evidence where we can run them.\n\n## 2. Agent Skills\n\nA skill is a folder with a SKILL.md file in it, instructions and maybe a few scripts that an agent loads when the task needs them [[11]](#ref-11). The spec's own showcase lists 46 agents that load skills [[12]](#ref-12), and we found 7 more whose docs say they do, Cline, Zed, Augment, Antigravity, Kilo Code, Devin Local (which used to be Windsurf) and Warp.\n\nWhere skills come from is messier. We counted 19 third-party marketplaces and directories, plus Anthropic's own repository, OpenAI's (now deprecated) and GitHub's `gh skill` command. Their counts go from 193 hand-picked skills to 1.56 million on one leaderboard, and how they check what they list goes from nothing at all to their own security scans. Skillselion says 87.8% of the skills it tracks have never had a security audit [[13]](#ref-13).\n\nSkills listed, by directory\n\n| Directory | Skills |\n| --- | --- |\n| AgenticSkills | 193 |\n| Agensi | 16000 |\n| Skillselion | 64756 |\n| LobeHub | 334426 |\n| MCP Market | 372895 |\n| skills.sh (all time) | 1555267 |\n\nEach series has a zero-based scale.\n\nNone of those 22 is listed on Anchor as a marketplace. Grading them is the same job we do for MCP registries, and with numbers that far apart (and most of it unchecked) someone needs to.\n\n## 3. Compute in orbit\n\nSince Google's [haiku from orbit](/blog/suncatcher-haiku-from-orbit) everyone is writing about space data centres, and nearly all of it is news. The question I couldn't find an answer to is who you can rent compute from up there today.\n\nThe answer is small chips only. TakeMe2Space publishes \"$2/min*\" for time on its OrbitLab satellite, an Nvidia Orin NX (the asterisk isn't explained) [[14]](#ref-14). Kepler leases compute on about 40 Orin chips across 10 satellites, by contract, and has 18 customers [[15]](#ref-15). Data-centre GPUs come later, and the first planned offer is Crusoe Cloud on a Starcloud satellite from early 2027 [[16]](#ref-16).\n\n| Who | What's up there | Can you use it? |\n| --- | --- | --- |\n| Starcloud | 1 Nvidia H100, since November 2025 | Not yet, Crusoe Cloud from early 2027 |\n| ADA Space with Zhejiang Lab | 12 computing satellites, May 2025 | Not that they describe |\n| Google Suncatcher | 4 TPUs, October 2026 | No, research only |\n| Kepler | About 40 Nvidia Orin chips | Yes, by contract |\n| TakeMe2Space | 1 Nvidia Orin NX | Yes, \"$2/min*\" |\n\nSources for the table [[16]](#ref-16) [[17]](#ref-17) [[18]](#ref-18) [[15]](#ref-15) [[14]](#ref-14).\n\nNobody agrees on the money either. Google's paper says it gets roughly comparable to a data centre's energy bill on the ground once launches cost under $200/kg, which it expects around the mid-2030s [[19]](#ref-19). Starcloud's CEO told McKinsey nearer $500/kg [[20]](#ref-20). BCG puts space at 2.5 to 3 times the total cost today [[21]](#ref-21), and IEEE Spectrum quotes ABI Research at 10 times or more [[22]](#ref-22).\n\nWe list none of these companies. It's a small set, so it's one page and a handful of listings.\n\n## 4. Voice agents that run on your own machine\n\nBlock showed Dessa this month, a voice model that listens and talks at the same time (full-duplex, so you can interrupt it like a person) and runs on a MacBook Pro M4 Max or better. It's a research preview with no weights to download [[23]](#ref-23).\n\nIt isn't alone. We counted 16 full-duplex voice systems that run on your own hardware, and 15 of them have weights you can download. Kyutai's Moshi is the one most of the others are built on, and its Apple backend runs on an iPhone [[24]](#ref-24). Kyutai tested its Hibiki translation model on an iPhone 16 Pro [[25]](#ref-25). NVIDIA released NemotronLabs VoiceChat 11B in August, open, full-duplex and able to call tools [[26]](#ref-26).\n\nWe list one full-duplex model that runs locally ([MiniCPM-o 4.5](/tools/minicpm-o-4-5)) and four that only run in someone's cloud. Moshi isn't listed, and about six models are built on it. This one sits between Anchor and LocalGhost, a voice agent that doesn't send your voice anywhere, so I'll admit I care about it more than the other four.\n\n## 5. Grading the steps an agent takes\n\nCisco published FAITH on 23/10/26, a harness that runs an agent and records every step, every tool call, what came back and why it stopped [[27]](#ref-27). Its first benchmark, VLoc, still scores the outcome, which files the agent pointed at and whether it correctly said a patched repository was safe. The best model, GPT-5.5, scored 0.229, and 190 of the 500 tasks beat every model [[28]](#ref-28).\n\nThe tools that do score the steps have settled on the same few methods. You compare the agent's tool calls with a reference list, strictly, in order or in any order, and an LLM judges the rest against a rubric. LangSmith, Google's ADK, AWS, Microsoft and Pydantic all ship nearly the same options [[29]](#ref-29) [[30]](#ref-30) [[31]](#ref-31) [[32]](#ref-32) [[33]](#ref-33). Even τ-bench, one of the best-known agent benchmarks, checks the final state of its database and treats its reference actions as one valid route rather than the only one [[34]](#ref-34).\n\nThe market moved too. Galileo is now part of Cisco's Splunk [[35]](#ref-35), and Patronus has retired its evaluation platform [[36]](#ref-36). We found 31 tools for this and list 10, without an evals category to put them in.\n\n## What we'll build\n\nThe plan is to do each of these the way Quynh's post works, with the names, the numbers in a table and a listing for each thing named. Open Responses comes first because it's closest to what we already do, then the skills marketplaces with grades, the orbital compute providers, the Moshi family, and an evals category. None of it is built yet.\n\nIf you run something that supports Open Responses, loads skills, rents compute in orbit, talks full-duplex on a laptop or grades agent steps, and it isn't in the lists above, [tell us](/about/#contact) and we'll add it.\n\n---\n\nMore in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md) · Older opinion [Google ran Gemma 4 on a TPU in orbit and got a haiku back](https://www.anchorterminal.com/blog/suncatcher-haiku-from-orbit.md)\n\n- Older: [The agent projects trending on GitHub, November 2026](https://www.anchorterminal.com/blog/github-trending-agent-tools-november-2026.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nRead on 30 October 2026. Quotes are as published.\n\n[1] Agentic AI Foundation, \"Agentic AI Foundation announces four new projects\", 22 October 2026. Source for Open Responses and Agent Skills moving to the foundation, and who contributed each. https://aaif.io/news/agentic-ai-foundation-announces-four-new-projects\n\n[2] Open Responses, overview and specification. Source for what the spec is and its dated versions, the latest 24 April 2026. https://www.openresponses.org/\n\n[3] Open Responses, acceptance tests. Source for the tests anyone can run against an endpoint. https://www.openresponses.org/compliance\n\n[4] Vercel, AI Gateway and OpenResponses. Source for what Vercel AI Gateway supports. https://vercel.com/docs/ai-gateway/sdks-and-apis/openresponses\n\n[5] Databricks, \"Query Open Responses models\". Source for what Databricks supports and what it doesn't. https://docs.databricks.com/aws/en/machine-learning/model-serving/query-open-responses-models\n\n[6] Portkey, Responses API. Source for \"fully Open Responses compliant\" across 60+ providers. https://docs.portkey.ai/docs/product/ai-gateway/responses-api\n\n[7] LM Studio, \"Open Responses\". Source for compatibility from version 0.3.39. https://lmstudio.ai/blog/openresponses\n\n[8] OGX (formerly Llama Stack), README. Source for the 93.8% conformance badge. https://github.com/ogx-ai/llama-stack\n\n[9] Ollama, OpenAI compatibility. Source for a stateless /v1/responses. https://docs.ollama.com/api/openai-compatibility.md\n\n[10] OpenRouter, Responses API overview. Source for a stateless Responses API. https://openrouter.ai/docs/api/reference/responses/overview.md\n\n[11] Agent Skills, specification. Source for what a skill is and the SKILL.md format. https://agentskills.io/specification\n\n[12] Agent Skills, client showcase. Source for the 46 agents it lists. https://agentskills.io/clients\n\n[13] Skillselion. Source for 64,756 skills indexed and 87.8% never audited. https://skillselion.com/\n\n[14] TakeMe2Space. Source for OrbitLab, \"$2/min*\" and the MOI-1A satellite's Nvidia Orin NX. https://tm2.space/\n\n[15] TechCrunch, \"The largest orbital compute cluster is open for business\", 13 April 2026. Source for Kepler's 10 satellites, about 40 Nvidia Orin chips and 18 customers. https://techcrunch.com/2026/04/13/the-largest-orbital-compute-cluster-is-open-for-business/\n\n[16] Crusoe, \"Crusoe to become first cloud operator in space through partnership with Starcloud\", 22 October 2025. Source for \"limited GPU capacity from space by early 2027\". https://www.crusoe.ai/resources/newsroom/crusoe-to-become-first-cloud-operator-in-space-through-partnership-with-starcloud\n\n[17] DCD, \"Starcloud-1 satellite reaches space, with Nvidia H100 GPU now operating in orbit\", 3 November 2025. Source for Starcloud's H100. https://www.datacenterdynamics.com/en/news/starcloud-1-satellite-reaches-space-with-nvidia-h100-gpu-now-operating-in-orbit/\n\n[18] Live Science, \"China is building a constellation of AI supercomputers in space\", 2025. Source for ADA Space and Zhejiang Lab's 12 computing satellites, launched in May 2025. https://www.livescience.com/space/space-exploration/china-is-building-a-constellation-of-ai-supercomputers-in-space-and-just-launched-the-first-pieces\n\n[19] Google Research, \"Exploring a space-based, scalable AI infrastructure system design\", 4 November 2025. Source for the $200/kg threshold and the mid-2030s. https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/\n\n[20] McKinsey, \"The case for data centers in space\", interview with Starcloud's CEO, 8 May 2026. Source for parity nearer $500/kg. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-case-for-data-centers-in-space\n\n[21] BCG, \"Space-Based Data Centers, More Than Hype, but Not a Revolution\", August 2026. Source for the 2.5 to 3 times premium today. https://www.bcg.com/publications/2026/space-based-data-centers-cost-outlook\n\n[22] IEEE Spectrum, \"Why Orbital Data Centers Are Harder Than Silicon Valley Thinks\", 11 June 2026. Source for ABI Research's 10 times or more. https://spectrum.ieee.org/orbital-data-centers-heat\n\n[23] Block, Dessa. Source for full-duplex, the hardware it runs on and the research preview. https://block.xyz/dessa\n\n[24] Kyutai, Moshi repository. Source for the Apple (MLX) backend for iPhone and Mac. https://github.com/kyutai-labs/moshi\n\n[25] Kyutai, Hibiki repository. Source for Hibiki 1B tested on an iPhone 16 Pro. https://github.com/kyutai-labs/hibiki\n\n[26] NVIDIA, NemotronLabs VoiceChat 11B model card. Source for an open full-duplex model with tool calling. https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B\n\n[27] Cisco, \"Introducing Agentic Evaluation in FAITH\", 23 October 2026. Source for what FAITH records. https://blogs.cisco.com/?p=499257\n\n[28] Cisco Foundation AI, VLoc Bench. Source for the two phases, GPT-5.5's 0.229 and the 190 of 500 tasks no model solved. https://cisco-foundation-ai.github.io/vulnerability-localization-benchmark/\n\n[29] LangChain, \"How to evaluate your agent with trajectory evaluations\". Source for reference trajectory matching in strict, ordered and unordered modes. https://docs.langchain.com/langsmith/trajectory-evals\n\n[30] Google ADK, evaluation criteria. Source for ADK's trajectory match options. https://adk.dev/evaluate/criteria/\n\n[31] AWS, AgentCore Evaluations. Source for AWS's trajectory evaluation. https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/evaluations.html\n\n[32] Microsoft, Foundry agent evaluators. Source for Microsoft's trajectory evaluators. https://learn.microsoft.com/en-us/azure/ai-foundry/concepts/evaluation-evaluators/agent-evaluators\n\n[33] Pydantic Evals, agentic evaluators. Source for Pydantic's trajectory evaluators. https://pydantic.dev/docs/ai/evals/evaluators/agentic/\n\n[34] Sierra, τ-bench evaluation docs. Source for scoring the final database state and reference actions as one valid route. https://github.com/sierra-research/tau2-bench/blob/main/docs/evaluation.md\n\n[35] Splunk, Galileo acquisition. Source for Galileo being part of Cisco's Splunk. https://www.splunk.com/en_us/about-splunk/acquisitions/galileo.html\n\n[36] Patronus AI docs. Source for the retired evaluation platform and Percival. https://docs.patronus.ai/docs/percival/percival\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-30",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.7",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Opinion",
        "url": "https://www.anchorterminal.com/blog/opinion/"
      },
      {
        "name": "Five new things nobody compares yet",
        "url": ""
      }
    ],
    "description": "Vlad Cealicu on five new corners of agent tooling with plenty of interest and no side-by-side comparison yet. Who supports Open Responses, where Agent Skills come from, who rents compute in orbit, which full-duplex voice models run locally, and how agent evals score the steps. With the numbers, and what Anchor Terminal lists today.",
    "facts": [
      "Vlad Cealicu",
      "30 October 2026"
    ],
    "h1": "Five new things in agent tooling that nobody compares yet",
    "image": "https://www.anchorterminal.com/assets/og/blog-agent-tools-nobody-compares-yet.png",
    "path": "/blog/agent-tools-nobody-compares-yet",
    "published": "2026-10-30",
    "section": "blog",
    "title": "Open Responses, Agent Skills, orbital compute, voice agents, evals | Anchor Terminal",
    "toc": null,
    "updated": "2026-10-30",
    "url": "https://www.anchorterminal.com/blog/agent-tools-nobody-compares-yet"
  },
  "tokens": {
    "markdown": 4200,
    "slim": 2130
  },
  "version": 1
}