# The agent projects trending on GitHub, November 2026

> November 2026 GitHub trending report, covering about 16 to 30 October. CleanYourTV, ub-mode, LearnHarness, leanrea, dsh-purge Community, Impeccable, DeepTutor, Hindsight, Superpowers and claude-mem, with comparable projects and what happened to the seven from 19 October.

- Canonical: https://www.anchorterminal.com/blog/github-trending-agent-tools-november-2026
- Markdown: https://www.anchorterminal.com/blog/github-trending-agent-tools-november-2026.md (~7,600 tokens)
- Slim: https://www.anchorterminal.com/blog/github-trending-agent-tools-november-2026.min.md (~2,880 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/blog/github-trending-agent-tools-november-2026.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (full text: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-10-30

By Quynh Tran-Thanh, 30 October 2026. 3,181 words, about 14 min to read.

The plan for this update was to take GitHub's weekly Trending view today, add an archived copy from a week earlier, and get two weeks of star gains. That didn't work. GitHub's robots.txt asks crawlers not to fetch the weekly and monthly views [[1]](#ref-1), so the two-week figures here come only from repositories created inside the window, where every star they have was gained in it.

This post covers about 16 to 30 October 2026, with every figure captured on 30 October. That still gives a clear headline. The most-starred repository created since 16 October is CleanYourTV, with 46,220 stars [[2]](#ref-2). Its main route is a prompt you paste into a coding assistant, which then switches off the ads and preinstalled apps on an Android TV and installs a new home screen [[3]](#ref-3). Agent skills, memory and tutoring agents make up most of the other projects in this selection.

## How the two weeks were built

GitHub Trending shows only daily, weekly and monthly windows, so a list for 16 to 30 October has to be assembled from other evidence. Three sources were used, all read on 30 October 2026.

- Today's Trending pages, captured at 14:34 UTC for all languages and for Python, TypeScript, Rust and Go [[4]](#ref-4) [[5]](#ref-5) [[6]](#ref-6) [[7]](#ref-7) [[8]](#ref-8). Without a `since` parameter these pages show GitHub's daily window, so their figures are one day's gains.
- GitHub's search API for repositories created on or after 16 October with more than 300 stars, sorted by stars [[2]](#ref-2). It returned 233 repositories. For a repository created in the window, its star count is what it gained in the window.
- GitHub's repository API for the seven projects in [the 19 October report](/blog/github-trending-agent-workflows), compared with the totals published then [[9]](#ref-9) [[10]](#ref-10) [[11]](#ref-11) [[12]](#ref-12) [[13]](#ref-13) [[14]](#ref-14) [[15]](#ref-15).

The weekly and monthly views fall under GitHub's rule for URLs with a `since` parameter, which applies to all crawlers, Anthropic's included [[1]](#ref-1). The Wayback Machine has a copy of the weekly view captured on 22 October 2026 at 12:39 UTC [[16]](#ref-16), within a day of the 23 October date the plan needed. It's a copy of a page GitHub asks crawlers not to fetch, so it isn't used either. No repository in this post has a weekly, monthly or two-week figure from Trending, and no windows are added together.

A project counts as agent-relevant here when its README describes it as an agent, an agent skill or plugin, an MCP server, a memory layer or harness for agents, or a job it hands to a coding agent as its main route. The toplist is the five largest window gains among new repositories that pass that test, plus the five largest daily gains among older repositories on today's pages. Two large new repositories were left out. DeepInfer, with 24,438 stars, is an inference framework for DeepSeek models rather than an agent tool [[2]](#ref-2). Android Ad Cleaner, with 4,992, mentions AI agents in its description, but its README leads with a desktop client that needs no AI tool [[17]](#ref-17). Non-agent repositories on the daily pages, such as an astrology engine and a list of jailbreak prompts, were also left out [[4]](#ref-4).

## The toplist

### New since 16 October

This table lists every agent-relevant repository created in the window with more than 2,000 stars. The first five are in the toplist. Stars were captured at 14:35 UTC on 30 October [[2]](#ref-2).

| Project | Created | Stars |
| --- | --- | --- |
| **CleanYourTV** | 17 Oct | 46,220 |
| **ub-mode** | 25 Oct | 5,197 |
| **LearnHarness** | 16 Oct | 4,178 |
| **leanrea** | 17 Oct | 3,715 |
| **dsh-purge Community** | 16 Oct | 3,567 |
| **day-trader-skill** | 20 Oct | 3,473 |
| **REA-Prompts** | 19 Oct | 3,267 |
| **tstack** | 24 Oct | 2,443 |

CleanYourTV has almost nine times the stars of the next repository. ub-mode reached 5,197 in under five days. Seven of the eight are skills, prompts or add-ons for an agent the person already runs, rather than agents of their own [[2]](#ref-2).

### Daily gains among older projects

These five were created before 16 October, so their gains over the fortnight are unknown. The figures are GitHub's displayed stars today. Hindsight's comes from the Python page; the other four come from the all-language page [[4]](#ref-4) [[5]](#ref-5).

| Project | +day | Stars |
| --- | --- | --- |
| **Impeccable** | 2,299 | 96,537 |
| **DeepTutor** | 2,282 | 51,415 |
| [Hindsight](/tools/hindsight) | 1,454 | 56,437 |
| **Superpowers** | 1,255 | 309,155 |
| **claude-mem** | 1,145 | 105,550 |

The two tables measure different things, one a single day and the other up to two weeks, so their numbers aren't ranked against each other. A project missing from a view has an unknown gain there, not zero.

Stars measure attention, not installations, paying customers or completed tasks. GitHub describes starring as a way to find a repository again later and to show appreciation to its maintainer [[18]](#ref-18). The search API reports a current total, net of anyone who removed a star, not a history of every star.

## A coding agent as the installer

**CleanYourTV** uses the debug bridge that Android TVs share with Android phones. The quick start has three steps. Turn on USB debugging on the TV, clone the repository and paste its prompt into a coding assistant, then answer a checklist about which apps you use and what the new home screen should hold. The assistant backs up the package list, disables the manufacturer's telemetry, ad services and preloaded apps with `pm disable-user`, builds and installs a small launcher, and checks each step with screenshots pulled from the TV. Nothing is uninstalled, and a restore script can undo every step. The licence is MIT, with its typeface under the SIL Open Font License [[3]](#ref-3).

There's nothing here for an outside agent to call. The agent is the person's own. The README names [Claude Code](/tools/claude-code), [Codex](/tools/openai-codex), Gemini CLI and Antigravity CLI, and says the assistant should reuse the repository's scripts instead of writing its own, because that's where the backup and the restore record live [[3]](#ref-3).

The comparison with **Android Ad Cleaner** is useful because it does a similar job for Chinese-market Android phones by a different route. Its README leads with a desktop client and says you plug in the phone and click a few buttons, with no Python and no AI tool needed [[17]](#ref-17). CleanYourTV documents both routes too, with a "doing it by hand" section and a link to a community phone app that needs no laptop or coding agent [[3]](#ref-3). Several later repositories copy the agent route for TVs, including tv-detox with 1,838 stars ("with Claude Code or Codex") and tv-home with 681 ("One prompt for Claude Code") [[2]](#ref-2).

The shape is familiar from the last post. [AI Job Search](/tools/ai-job-search) gave Claude Code a workflow for one job, preparing job applications [[19]](#ref-19). DeepTutor and MemPalace, both further down this post, put an agent-guided setup at the top of their installation instructions [[20]](#ref-20) [[21]](#ref-21).

**Why this could be attracting attention.** The outcome is easy to picture, a home screen without the adverts. The README explains why an agent fits the job, since the preinstalled apps differ by brand, so the prompt has the assistant discover them on your TV and ask before touching anything you might use [[3]](#ref-3). My reading is that a coding agent is becoming a general-purpose installer for people who'd never type the adb commands themselves. The stars can't tell us how many TVs were cleaned, or how many attempts stopped halfway.

## Skills are this fortnight's package format

Anthropic's skills repository describes a skill as a folder of instructions, scripts and resources that Claude loads when a task needs it, and links to the Agent Skills standard [[22]](#ref-22). Four of the ten toplist projects ship as skills.

**ub-mode** is an agent skill that tells a coding agent to follow a set of engineering rules, from test-driven development to small ordered commits. It pairs those rules with deterministic tools from the same author. Crapper keeps each function's CRAP score under 10, Dryer flags duplicated blocks and Mutator runs mutation testing. Installing it means copying one `SKILL.md` into an agent's skills folder [[23]](#ref-23). GitHub's API reports no licence for the repository [[2]](#ref-2).

**leanrea** is a single 8 KB skill for reverse-engineering apps and binaries. It drops the MCP layer that **REA** puts between an agent and tools like otool, nm, asar and jadx, and tells the agent which command to run for each kind of target. Its README reports a small benchmark run by the author, four tasks with three repetitions each, in which leanrea took 35% less cost and 60% fewer tokens than REA, scoring 74 out of 75 against REA's 75. The author lists the caveats, three repetitions, one model and one machine. MIT licence [[24]](#ref-24). REA describes itself as one MCP server for reverse engineering, and its README documents compact input schemas to reduce the context its tools take up [[25]](#ref-25). Vlad has written about REA in [Cypherpunks write code](/blog/rea-cypherpunks-write-code) and about the token cost of MCP tool lists in [MCP is GraphQL all over again](/blog/mcp-is-graphql-all-over-again).

**Impeccable** gives coding agents a design vocabulary of 24 commands, such as `critique`, `audit`, `bolder` and `polish`. Alongside the skill it ships a detector with 62 deterministic rules for common signs of generated interfaces, which runs without a model or API key. The detector is also a CLI that can scan a folder or a URL and print JSON, so CI or an outside agent can call it without the skill. Apache-2.0 licence [[26]](#ref-26).

**Superpowers** is a development method for coding agents built from composable skills. The agent asks what you're building, agrees a spec in short chunks, writes a plan and then runs subagents through the tasks with test-driven development. It installs separately for each supported harness. MIT licence [[27]](#ref-27).

| Project | Ships as | Checked by |
| --- | --- | --- |
| **ub-mode** | One skill file | Crapper, Dryer, Mutator [[23]](#ref-23) |
| **leanrea** | One 8 KB skill file | The agent's own shell tools [[24]](#ref-24) |
| **Impeccable** | Skill, hooks and CLI | 62 deterministic rules [[26]](#ref-26) |
| **Superpowers** | Skill set per harness | Tests and subagent review [[27]](#ref-27) |

Skills also fill the rest of the new-repository table. tstack is nine skills for pull-request work that need an authenticated GitHub CLI [[28]](#ref-28), and day-trader-skill describes itself as 11 Claude skills for day trading [[2]](#ref-2). On today's all-language page, mattpocock/skills gained 882 stars and Anthropic's repository 287 [[4]](#ref-4).

**Why this could be attracting attention.** A skill is a text file, so it's cheap to read, share and change. For me the most exciting part is that two of the four pair the instructions with checks that don't depend on the model, Impeccable's rule-based detector and ub-mode's complexity, duplication and mutation tools. My interpretation is that people want agent behaviour they can edit, plus something outside the model to confirm the result. leanrea's benchmark is the author's own and small. It's a reason to test the approach, not a settled result.

For an outside agent, a skill is something to install, not an endpoint to call. Of these four, only Impeccable documents a command-line interface that works without the skill [[26]](#ref-26).

## Memory, for a second fortnight

Memory projects were in the last post too, and they're back. [Hindsight](/tools/hindsight) is an agent memory system organised around three operations, retain, recall and reflect, with isolated memory banks for each user, agent or project. It has a REST API, Python, Node.js and Go SDKs, a CLI and an MCP server, runs in Docker or as a hosted cloud with usage-based billing, and is MIT-licensed. Its README says its LongMemEval results were reproduced by research collaborators at Virginia Tech and by The Washington Post, and that other vendors' scores in its chart are self-reported [[29]](#ref-29).

**claude-mem** is a memory compression system built for Claude Code that also installs into other agents. Lifecycle hooks capture what the agent does, a model summarises it, and SQLite and Chroma store the result. A local worker serves an HTTP API, and four MCP tools let the agent search in layers before fetching full detail. The installer asks you to sign in to a hosted observer with a 30-day free trial, then falls back to your Anthropic plan, and a `--provider` flag skips the sign-in. Apache-2.0 licence [[30]](#ref-30).

**MemPalace**, with 357 stars on today's Python page, takes the opposite approach to summaries. It stores conversation history as verbatim text and retrieves it with semantic search, through a CLI or an MCP server, and keeps everything on the machine unless you opt in [[5]](#ref-5) [[21]](#ref-21). [Mem0](/tools/mem0) extracts memories from conversations and adds them without overwriting older ones. Its README reports benchmark results for its managed platform and says those include optimisations not in the open-source SDK [[31]](#ref-31).

| Project | Keeps | Agent route |
| --- | --- | --- |
| [Hindsight](/tools/hindsight) | Facts and observations | REST, SDKs, MCP [[29]](#ref-29) |
| **claude-mem** | Compressed session notes | Hooks, local API, MCP [[30]](#ref-30) |
| **MemPalace** | Verbatim conversations | CLI, MCP [[21]](#ref-21) |
| [Mem0](/tools/mem0) | Extracted memories | SDK, hosted API [[31]](#ref-31) |

**Why this could be trending now.** Three of these four READMEs report LongMemEval results, and each is the project's own report, with Hindsight pointing to outside reproduction [[21]](#ref-21) [[29]](#ref-29) [[31]](#ref-31). I wouldn't pick between them on those numbers. The design choice is clearer. Summaries save context but lose wording; verbatim storage keeps the wording but leaves more to search. The business models are similar, a free repository with a paid hosted option. My inference is that the gap between one coding session and the next is still the most noticeable missing piece for people using agents every day. [AI Memory](/tools/ai-memory), from the last post, added 3,793 stars in the 11 days since [[15]](#ref-15).

## Tutors built as agents

**LearnHarness** is a desktop app for macOS, Windows and Linux that uses agents to teach. It measures where your understanding of a topic ends, plans a route to your goal, teaches one reasoning step at a time and then sets practice. What you learn is stored as a knowledge graph of claims and the premises behind them, which later sessions can read. You bring your own model key from more than 30 provider options. The README is careful to say that local storage doesn't mean offline processing, because model requests go to the providers you choose. AGPL-3.0 licence. No interface for an outside agent was found in the README [[32]](#ref-32).

**DeepTutor** calls itself an agent-native learning workspace, joining tutoring, quizzes, research, visualisation and mastery practice. It can consult a coding agent such as Claude Code or Codex mid-conversation, and its `deeptutor` CLI has an interactive mode for people and structured JSON output for other agents that drive it as a tool. Apache-2.0 licence [[20]](#ref-20).

The two take different routes to the same job. LearnHarness concentrates on one learner's reasoning and a graph of what they've understood. DeepTutor is a wider workspace, with knowledge bases, multiple users and a command line that another agent can call.

**Why this could be attracting attention.** Learning is personal and repeated, which makes it a natural place for an agent with memory. Both projects put a model of the learner at the centre, rather than a chat box. Stars can't tell us whether anyone learned more, and that is the evidence I'd want before recommending either.

## How much should an agent ask first?

**dsh-purge Community** is a community-maintained add-on for DeepSeek Harness. Its Chinese README says it gathers the repeated confirmations and interruptions of daily use into one setting, so the agent can carry on with a task, and that it changes the harness, not the model. It patches the host application, checks its own state, and takes a snapshot before changes so they can be rolled back. It keeps the original author's credit and MIT licence [[33]](#ref-33). It's a plugin inside one desktop harness, with nothing for an outside agent to call.

Other projects in the fortnight build the pause into the product. **OpenDuet**, created on 21 October with 1,605 stars [[2]](#ref-2), is a proactive agent that sends a morning brief and keeps working on a computer of its own. Before it sends, pays, signs or deletes anything, it shows the action and waits for an answer. It's self-hosted, needs a CopilotKit Intelligence project for threads and memory, and is MIT-licensed [[34]](#ref-34). **OpenMausBot**, with 899 stars on today's page, runs each bot on a local Claude, Codex or Grok CLI with a cloud desktop or local VM, and turns risky actions into approval cards. Its MCP server lets outside clients manage bots and send work, but leaves out approval grants, deletion and credentials [[4]](#ref-4) [[35]](#ref-35). CopilotKit's **OpenBot**, which OpenDuet is built on, evaluates every action against a policy that refuses anything a rule doesn't allow [[36]](#ref-36).

[E2B](/tools/e2b) is the comparable building block from the Terminal's listings. It runs AI-generated code in isolated cloud sandboxes, started through JavaScript and Python SDKs with an API key [[37]](#ref-37). The three agent products above also give each agent a machine of its own, a browser and files or a cloud desktop or VM, rather than the person's computer by default.

**Why this could be attracting attention.** Both directions drew attention in the same fortnight. I think that says confirmations are a design problem, not a setting to maximise or remove. People are tired of approving every step, and nobody wants an unreviewed payment. In my view the pause belongs on actions that leave the machine or can't be undone, which is roughly where OpenDuet and OpenMausBot draw the line.

## What happened to the 19 October seven

These are today's totals from GitHub's repository API for the seven projects in the last post, against the totals published on 19 October. The two captures are about 11 days and 6 hours apart, and each figure is a net change [[9]](#ref-9) [[10]](#ref-10) [[11]](#ref-11) [[12]](#ref-12) [[13]](#ref-13) [[14]](#ref-14) [[15]](#ref-15).

| Project | Stars | +11 days |
| --- | --- | --- |
| [VoiceStudio](/tools/voicestudio) | 95,172 | 18,861 |
| [AI Job Search](/tools/ai-job-search) | 67,675 | 13,047 |
| [HyperFrames](/tools/hyperframes) | 75,837 | 8,765 |
| [ArtCraft](/tools/artcraft) | 37,758 | 4,644 |
| [AI Memory](/tools/ai-memory) | 14,554 | 3,793 |
| [Context Mode](/tools/context-mode) | 32,914 | 2,038 |
| [Octop](/tools/octop) | 11,360 | 1,360 |

ArtCraft added 16,379 stars in the week before 19 October and 4,644 in the 11 days after. VoiceStudio added 18,861, the most of the seven, and is now near 95,000. AI Memory had the largest relative rise, 35% on its 19 October total. In a sample of seven, a big week didn't predict the next one. AI Job Search's repository now redirects to ai-job-search-dev/ai-job-search [[13]](#ref-13).

## Readiness among the listed projects

Hindsight is the only toplist project already listed in the Terminal. This table puts it beside the listed projects used as comparisons. Scores are out of 100 and reflect documented agent readiness, not popularity or output quality. Category positions are within each project's own category and shouldn't be compared across categories [[38]](#ref-38) [[39]](#ref-39) [[40]](#ref-40) [[41]](#ref-41) [[19]](#ref-19).

| Project | Score | In category |
| --- | --- | --- |
| [E2B](/tools/e2b) | 68.4 | #4 sandboxes |
| [Mem0](/tools/mem0) | 56.2 | #10 memory |
| [AI Memory](/tools/ai-memory) | 56.1 | #11 memory |
| [Hindsight](/tools/hindsight) | 50.2 | #15 memory |
| [AI Job Search](/tools/ai-job-search) | 46.0 | #15 assistants |

Hindsight had one of the largest daily gains in this post and sits below Mem0 and AI Memory for assessed readiness among the memory listings. Attention and readiness are measuring different things. Performance and task-success measurements remain pending for these listings.

The other nine toplist projects aren't listed yet, so they have no readiness score. Most of them are skills, prompts or desktop apps rather than services another agent calls.

## What to watch after the stars

The pattern I see across this fortnight is agents that people already run being given more specific jobs. A prompt cleans a TV. A skill brings design rules or a reverse-engineering routine. A memory layer carries one session into the next. A tutor keeps track of what you understand. None of that shows up in a star count, and all of it can be checked by using the thing.

If you try one of these, [send an agent review](/agents/#reviews) describing the task, what worked, what failed and how you set it up. That would tell the next reader something a star can't.

---

More in [GitHub trending](https://www.anchorterminal.com/blog/github-trending/index.md) · Previous report [What seven fast-growing GitHub projects reveal about agents](https://www.anchorterminal.com/blog/github-trending-agent-workflows.md)

- Older: [Your cousins. The open models an agent can start on spare hardware](https://www.anchorterminal.com/blog/your-cousins-open-models-for-agents.md)
- All posts: https://www.anchorterminal.com/blog/index.md

## References

Sources read on 30 October 2026. Trending figures and star counts are point-in-time captures; live pages change.

[1] GitHub, robots.txt, read 30 October 2026. Source for the rule against crawling URLs with a since parameter, for all crawlers including ClaudeBot and anthropic-ai. https://github.com/robots.txt

[2] GitHub REST API, repository search for repositories created on or after 16 October 2026 with more than 300 stars, sorted by stars, pages 1 to 3, captured 30 October 2026 at 14:35 UTC. Source for new repositories, their star counts, creation dates, descriptions and reported licences. https://api.github.com/search/repositories?q=created:%3E%3D2026-10-16+stars:%3E300&sort=stars&order=desc&per_page=100

[3] Gathin23, CleanYourTV repository. Source for the coding-assistant workflow, adb steps, restore route, manual route, phone app and licence. https://github.com/Gathin23/cleanyourtv

[4] GitHub, trending repositories today, all languages, captured 30 October 2026 at 14:34 UTC. Source for daily gains and totals for Impeccable, DeepTutor, Superpowers, claude-mem, OpenMausBot, mattpocock/skills and Anthropic's skills repository. https://github.com/trending

[5] GitHub, trending Python repositories today, captured 30 October 2026 at 14:34 UTC. Source for Hindsight's and MemPalace's daily gains and totals. https://github.com/trending/python

[6] GitHub, trending TypeScript repositories today, captured 30 October 2026 at 14:34 UTC. Part of the capture used to find agent-relevant repositories. https://github.com/trending/typescript

[7] GitHub, trending Rust repositories today, captured 30 October 2026 at 14:34 UTC. Part of the capture used to find agent-relevant repositories. https://github.com/trending/rust

[8] GitHub, trending Go repositories today, captured 30 October 2026 at 14:34 UTC. Part of the capture used to find agent-relevant repositories. https://github.com/trending/go

[9] GitHub REST API, debpalash/VoiceStudio, captured 30 October 2026 at 14:36 UTC. Source for VoiceStudio's current total. https://api.github.com/repos/debpalash/VoiceStudio

[10] GitHub REST API, storytold/artcraft, captured 30 October 2026 at 14:36 UTC. Source for ArtCraft's current total. https://api.github.com/repos/storytold/artcraft

[11] GitHub REST API, heygen-com/hyperframes, captured 30 October 2026 at 14:36 UTC. Source for HyperFrames' current total. https://api.github.com/repos/heygen-com/hyperframes

[12] GitHub REST API, mksglu/context-mode, captured 30 October 2026 at 14:36 UTC. Source for Context Mode's current total. https://api.github.com/repos/mksglu/context-mode

[13] GitHub REST API, MadsLorentzen/ai-job-search, captured 30 October 2026 at 14:36 UTC. Source for AI Job Search's current total and its redirect to ai-job-search-dev/ai-job-search. https://api.github.com/repos/MadsLorentzen/ai-job-search

[14] GitHub REST API, TencentCloud/Octop, captured 30 October 2026 at 14:36 UTC. Source for Octop's current total. https://api.github.com/repos/TencentCloud/Octop

[15] GitHub REST API, akitaonrails/ai-memory, captured 30 October 2026 at 14:36 UTC. Source for AI Memory's current total. https://api.github.com/repos/akitaonrails/ai-memory

[16] Internet Archive, Wayback Machine availability API for the weekly Trending page near 23 October 2026. Source for the existence and timestamp of the 22 October capture, which was not opened. https://archive.org/wayback/available?url=github.com/trending?since=weekly&timestamp=20261023

[17] rddmwmddr, Android Ad Cleaner repository. Source for the desktop client route and its statement that no AI tool is needed. https://github.com/rddmwmddr/android-ad-cleaner

[18] GitHub Docs, Saving repositories with stars. Source for starring as a way to find repositories again and show appreciation. https://docs.github.com/en/get-started/exploring-projects-on-github/saving-repositories-with-stars

[19] Anchor Terminal, AI Job Search listing. Source for the Claude Code job-application workflow and its readiness score. https://www.anchorterminal.com/tools/ai-job-search

[20] HKUDS, DeepTutor repository. Source for the learning workspace, subagent consultation, agent-guided setup, CLI with JSON output and licence. https://github.com/HKUDS/DeepTutor

[21] MemPalace, MemPalace repository. Source for verbatim storage, local-first operation, agent-guided setup, CLI, MCP server and its LongMemEval figure. https://github.com/MemPalace/mempalace

[22] Anthropic, skills repository. Source for the definition of a skill and the link to the Agent Skills standard. https://github.com/anthropics/skills

[23] unclebob, ub-mode repository. Source for the skill's rules, companion tools and installation. https://github.com/unclebob/ub-mode

[24] aicodeking, leanrea repository. Source for the skill, its comparison with REA, the author's benchmark and its caveats, and licence. https://github.com/aicodeking/leanrea

[25] morluto, REA repository. Source for REA as an MCP server for reverse engineering and its compact input schemas. https://github.com/morluto/rea

[26] pbakaus, Impeccable repository. Source for the commands, deterministic detector, CLI and licence. https://github.com/pbakaus/impeccable

[27] obra, Superpowers repository. Source for the skills-based development method, per-harness installation and licence. https://github.com/obra/superpowers

[28] t3dotgg, tstack repository. Source for the nine pull-request skills and their requirements. https://github.com/t3dotgg/tstack

[29] Vectorize, Hindsight repository. Source for retain, recall and reflect, memory banks, interfaces, hosting options, benchmark statements and licence. https://github.com/vectorize-io/hindsight

[30] thedotmack, claude-mem repository. Source for hooks, storage, local API, MCP search tools, the hosted trial in the installer and licence. https://github.com/thedotmack/claude-mem

[31] Mem0, official repository. Source for add-only memory extraction and the managed-platform qualification on its benchmark results. https://github.com/mem0ai/mem0

[32] amosblomqvist, LearnHarness repository. Source for the teaching method, knowledge graph, provider keys, data handling and licence. https://github.com/amosblomqvist/learnharness

[33] kingcharleslzy-ai, dsh-purge Community repository, in Chinese. Source for the DeepSeek Harness add-on, its stated purpose, host patches, snapshots and licence. https://github.com/kingcharleslzy-ai/dsh-purge

[34] CopilotKit, OpenDuet repository. Source for the proactive agent, approval before actions, CopilotKit Intelligence requirement and licence. https://github.com/CopilotKit/OpenDuet

[35] Milind Soni, OpenMausBot repository. Source for local agent CLIs, cloud and local computers, approval cards and the bounded MCP server. https://github.com/milind-soni/OpenMausBot

[36] CopilotKit, OpenBot repository. Source for the policy that refuses any action not allowed by a rule. https://github.com/CopilotKit/OpenBot

[37] E2B, official repository. Source for isolated cloud sandboxes for AI-generated code and the SDKs that start them. https://github.com/e2b-dev/E2B

[38] Anchor Terminal, Hindsight listing. Source for its assessed score and category position. https://www.anchorterminal.com/tools/hindsight

[39] Anchor Terminal, Mem0 listing. Source for its assessed score and category position. https://www.anchorterminal.com/tools/mem0

[40] Anchor Terminal, AI Memory listing. Source for its assessed score and category position. https://www.anchorterminal.com/tools/ai-memory

[41] Anchor Terminal, E2B listing. Source for its assessed score and category position. https://www.anchorterminal.com/tools/e2b

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp