← research

The Top-0.1% Toolbelt (wild tools vs our stack)

Jul 14, 2026

Date: 2026-07-07 · Method: 5 parallel research lanes (sonnet subagents) over gh search/gh api (star counts verified live), WebSearch/WebFetch, and opencli twitter search — which was rate-limited (HTTP 429) in every lane, so X endorsements are web-sourced snippets, not primary metrics. Flagged per entry.

Dedupe base: RESEARCH_frontier2.md, RESEARCH_cli_capabilities.md, RESEARCH_fable_rush_github.md (notch/island category, ccusage/CodexBar/statusline ecosystem, merge-queue/agentbox/token-diet/compact-plus, cmux ecosystem, fable-skills genre, native CLI capabilities incl. Seatbelt sandbox, /batch, agent teams).

Judged against the standing stack (never recommended, only compared): Mission Control, cmux, gbrain/vault, OpenClaw, lavish, hyperframes, gstack (careful/freeze/guard), no-mistakes, fleet (Codex/GLM/Fugu), loop-engineering, browser-harness, Browserbase, agent-reach, pxpipe (~86% measured input cut), Wispr Flow, context7 (installed), headroom MCP (installed).

Verdict scale: ADOPT (install this week) · WATCH (real, not yet worth the slot) · SKIP (redundant/dead/worse).

1. Memory / context systems (beyond gbrain + vault)

claude-mem — thedotmack/claude-mem · 86,176★ (verified; created 2025-08-31, pushed Jul 6) Automatic session memory: captures everything the agent does, AI-compresses it, injects relevant context into future sessions; works across Claude Code, OpenClaw, Codex, Gemini, Copilot, Hermes. Endorsed: viral wave — @charliejhills video ("95% fewer tokens per session… 46k stars in 48 hours", x.com/i/status/2043930556657029252, 666♥/120k views), @VaibhavSisinty at the 84k-star mark (x.com/i/status/2070363935690346982). Vs ours: gbrain is user-curated knowledge; extract-approach captures only hard-won insights. Nothing in the stack does automatic agent-session→session memory. Complementary, not overlapping — and it speaks OpenClaw too. VERDICT: ADOPT (trial) — the one whole missing layer in the memory stack; watch its token overhead against pxpipe economics.

zilliztech/claude-context — 12,073★ (created 2025-06, pushed Jun 22; +3.7k★/week trending run) Hybrid semantic+BM25 code-search MCP (Milvus team); AST-aware chunking, ~40% token-reduction claim. Endorsed: independent review at andrew.ooo; PulseMCP top pick. Vs ours: native grep/Read + Fable's own navigation cover small/medium repos; this earns its keep only on huge/polyglot codebases, and needs a vector backend + embedding keys. VERDICT: WATCH — install the week a 100k-file repo enters the picture, not before.

getzep/graphiti — 28,438★ (pushed Jul 6) · Temporal knowledge-graph agent memory (bi-temporal edges); Zep paper arXiv:2501.13956, Neo4j dev-blog coverage. Vs ours: gbrain has no temporal-conflict model, but the user's corpus (~600 pages) is below the scale where this pays (even Garry Tan's "useful at 10k+ files" logic applies here). VERDICT: SKIP (for now) — right idea, wrong scale for this vault.

mem0ai/mem0 (60,246★), letta-ai/letta (23,676★), basicmachines-co/basic-memory (3,385★), GreatScottyMac/context-portal (763★) — generic memory SDKs/servers, no fresh practitioner endorsement found. Vs ours: all fill the slot gbrain already owns, with less structure and no MECE vault. VERDICT: SKIP (all four) — redundant with gbrain.

scrypster/muninndb — 305★ (created Feb 2026, pushed Jul 4) · "Cognitive database": Ebbinghaus decay + Hebbian strengthening as memory primitives, single binary, MCP-native. No field validation yet. VERDICT: WATCH — the only conceptually novel memory design in the sweep; unproven.

2. Transcript / session tooling

claude-devtools — matt1398/claude-devtools · 3,669★ (pushed May 13) · Visual DevTools for sessions; uniquely deep at raw MCP request/response payload debugging. Endorsed: 500+ upvote Reddit launch, Product Hunt, dev.to comparison (dev.to/gonewx/i-tested-4-tools-for-browsing-claude-code-session-history-17ie). Vs ours: Mission Control owns session forensics; the MCP-payload inspector is the one lens it lacks. VERDICT: WATCH — open it the next time an MCP server misbehaves; otherwise Mission Control subsumes it.

sniffly — chiphuyen/sniffly · 1,244★ (dormant since Aug 2025) · Error-pattern mining across sessions ("Content Not Found" = 20-30% of failures). Endorsed: Chip Huyen herself (x.com/chipro/status/1945527700808184115) + HN. Vs ours: dormant tool, but the error-taxonomy lens is a Mission Control feature idea, free to steal. VERDICT: SKIP the tool, steal the idea (error-pattern panel on our own parsed data).

claude-historian-mcp — Vvkmnn/claude-historian-mcp · 180★ · MCP that lets the agent query its own past sessions at runtime. No endorsement found. Vs ours: Mission Control is human-facing; gbrain is knowledge, not transcripts. Agent-queryable history is a real third axis — but claude-mem (above) delivers most of it automatically. VERDICT: SKIP — superseded by claude-mem if adopted; else a Mission Control MCP-facade feature.

claude-code-log (1,130★, active), claude-code-trace (341★), tail-claude (147★), claude-conversation-extractor (634★), superview.sh (2,106★ but quiet since April, zero verifiable endorsement) — viewers/exporters over the same JSONL Mission Control already parses. VERDICT: SKIP (all).

3. MCP servers actually worth running

context7 — upstash/context7 · 58,678★ · Still the consensus "install first" across every 2026 practitioner roundup found. Already installed here. VERDICT: KEEP (confirmed, no action).

serena — oraios/serena · 26,155★ (pushed Jul 6, active) · Semantic code retrieval/editing over a symbol graph (find-references, replace_symbol_body) instead of grep+read. Endorsed: 2026 reviews converge on "try it on your biggest repo, not your smallest"; maintainers notably removed their regex-edit tool on Claude Code because the native editor already wins there — honest engineering. Vs ours: the only MCP in the sweep that adds a capability the stack genuinely lacks (symbol-level navigation on big codebases). Marginal on small repos. VERDICT: ADOPT (trial) on the largest active repo; drop it if native search keeps winning.

github-mcp-server — github/github-mcp-server · 31,252★ · Official GitHub MCP. Vs ours: gh CLI is already muscle memory here and costs zero context window; the MCP mostly re-skins it. VERDICT: SKIP — gh covers it cheaper.

chrome-devtools-mcp (46,101★) / playwright-mcp (34,783★) / browser-use MCP — Vs ours: browser-harness is purpose-built CDP control with field-tested gotcha docs; adding a browser MCP is a third layer over the same compositor. Simon Willison himself now prefers raw Playwright scripts over the MCP for coding agents. VERDICT: SKIP (all browser MCPs) — browser-harness is the equivalent, and better-fitted.

pal-mcp-server (ex zen-mcp) — BeehiveInnovations/pal-mcp-server · 11,646★ · multi-model second-opinion MCP. Last push 2025-12-15 — 6+ months stale. Vs ours: the fleet skill + routing doctrine is a hand-rolled, maintained version of the same thing. VERDICT: SKIP.

Linear MCP (official, hosted mcp.linear.app) — Vs ours: Mission Control already exports the deadline board to Linear; the hosted MCP would let any session touch issues directly, OAuth 2.1, zero maintenance. VERDICT: ADOPT (cheap) — one claude mcp add away, real workflow fit.

sentry-mcp (753★ live — blog claims of ~5k are stale) — SKIP unless Sentry enters the stack. apple-mcp (3,120★, dead 11 months) — SKIP. Standalone sequential-thinking — fragmented, eaten by native extended thinking — SKIP.

4. Skill ecosystems

obra/superpowers — 247,803★ (verified — real, ~5x anything else in this report) · Full agentic-dev methodology as skills (brainstorming gate, TDD enforcement, systematic-debugging, subagent-driven development); in Anthropic's own plugin marketplace since Jan 2026. Vs ours: CLAUDE.md + fable-judgment + unknowns-discipline independently encode much of the same philosophy; the delta is mechanical enforcement and battle-tested wording. VERDICT: ADOPT (mine, don't wholesale-install) — read end-to-end, cherry-pick 2-3 skills, diff against our own doctrine. The most load-bearing single read in this sweep.

vercel-labs/skills CLI — 25,254★ (pushed Jul 7) + skills.sh leaderboard (top skill: Remotion at 117k weekly installs) · "npm for agent skills" across 27+ runtimes; sibling collection vercel-labs/agent-skills (28,746★). Vs ours: the ~50-skill private library has no discovery/distribution layer; this is the standard forming around it. VERDICT: ADOPT (cheap) — use for discovery + as the publish channel if any private skill ever goes public.

SkillsMP (2M+ skills indexed, 2-star floor, no curation) — SKIP for quality; browse-only. wondelai/skills (1,567★ cross-agent skill spec incl. OpenClaw) — WATCH — relevant the day the private library needs to run on OpenClaw/Hermes natively. samber/cc-skills-golang (2,428★) — SKIP (no Go here).

5. Worktree / parallel-agent managers

BloopAI/vibe-kanban — 27,291★ but pushed Apr 24 — cooling · kanban over per-task worktrees. Vs ours: cmux + native /batch + --tmux -w cover parallelism; kanban UX isn't worth a staling dependency. VERDICT: SKIP.

smtg-ai/claude-squad — 8,046★ (active) · tmux+worktree TUI with a built-in diff/review gate before merge. Vs ours: closest direct cmux competitor; the review gate is its only delta, and no-mistakes + /code-review cover review. VERDICT: SKIP — switching cost exceeds the delta.

kbwo/ccmanager — 1,177★ (very active) · cross-CLI session manager that copies Claude Code session state into new worktrees for context continuity. VERDICT: WATCH — the continuity trick is novel; revisit if multi-vendor CLI switching becomes routine.

imbue-ai/sculptor — 194★ (pushed today) · parallel agents in Docker containers with "Pairing Mode" to sync a container back into the local repo. VERDICT: WATCH — different isolation model, tiny traction.

stravu/crystal (3,098★ → deprecated into closed Nimbalyst) — SKIP (dead end). Conductor (conductor.build, closed, macOS GUI) — SKIP — scriptability downgrade. d-kuro/gwq (439★) — SKIP — a nicer worktree picker, nothing more. Terragon — shut down, SKIP. uzi — dead 13 months, SKIP.

6. Hook libraries

nizos/tdd-guard — 2,244★ (pushed yesterday; npm + marketplace plugin) · PreToolUse hook that blocks Edit/Write when the agent skips failing-test-first or over-implements past the current test. Endorsed: ClaudeLog, Claude Directory write-ups. Vs ours: no-mistakes gates after the fact; gstack guard skills gate danger, not discipline. Test-first enforcement is a genuinely missing gate — and it's the "hook enforcement > prompt prose" doctrine, productized. VERDICT: ADOPT (trial) on one TS project; keep only if it doesn't fight the workflow.

disler/claude-code-hooks-mastery — 3,816★ · IndyDevDan's canonical every-hook-event teaching repo. VERDICT: SKIP as install, keep as reference when authoring the fan-out-cap PreToolUse hook CLAUDE.md still owes.

karanb192/claude-code-hooks (439★, fresh) — copy/paste cookbook; cherry-pick source, not a dependency. SKIP. GowayLee/cchooks (131★, Python hook SDK) — SKIP unless we start writing Python hooks. beyondcode/claude-hooks-sdk (PHP) / webdevtodayjason/claude-hooks (dead ~1yr) — SKIP.

7. Sandboxing

stacklok/brood-box — 48★ (created Feb 2026, pushed 2 days ago; Stacklok engineering blog: "Your AI Agent Doesn't Deserve Root Access") · Runs Claude Code/Codex/OpenCode in hardware-isolated microVMs (libkrun / Hypervisor.framework) with COW workspace snapshots (agent never touches real files until reviewed), DNS-aware egress firewall, Cedar-based MCP-tool authorization. Vs ours: native Seatbelt is syscall-level; gstack guard is prompt-level. VM boundary + snapshot-before-review is a strictly stronger envelope for unattended /goal-style runs. Low stars, credible vendor. VERDICT: ADOPT (trial) for one nightly unattended run; the fit with the unattended-runs doctrine is exact.

anthropic-experimental/sandbox-runtime — 4,592★ (pushed yesterday) · The literal open-source guts of Claude Code's native sandbox, runnable standalone via npx @anthropic-ai/sandbox-runtime <cmd>. Vs ours: zero delta inside Claude Code; the play is wrapping non-Claude processes (OpenClaw agents, fleet CLIs) in the same Seatbelt profile. VERDICT: WATCH — adopt the day an OpenClaw agent needs OS-level confinement.

e2b (12,869★), daytonaio/daytona (72,277★), microsandbox (6,851★), buildkite/cleanroom (55★) — cloud/generic sandbox infra; right tools, wrong scale for a single power user. VERDICT: SKIP (e2b/daytona) / WATCH (microsandbox as a generic local microVM backend). claudebox (dead 10mo), textcortex/claude-code-sandbox (archived → unverified "Spritz" successor) — SKIP.

8. Eval / testing harnesses for agent output

promptfoo — 22,984★ · CI eval runner + red-team probes; acquired by OpenAI Mar 2026 (stays MIT), 350k+ devs. Vs ours: native claude plugin eval + ultrareview + no-mistakes + judge-based loop conditions cover the daily gate; promptfoo adds CI-pinned regression suites and a red-team probe library. VERDICT: WATCH — reach for it when a prompt/skill needs a versioned regression suite; note the OpenAI ownership.

langfuse — 30,574★ · self-hosted trace observability + LLM-judge evals + datasets. Vs ours: loop judges have no persistence layer — but standing up langfuse for one user is heavy, and Mission Control is the natural home for trace history. VERDICT: WATCH.

inspect_ai (2,310★, UK AISI; endorsed by Hamel Husain, hamel.dev) — rigor tier above daily needs. SKIP. evalite (1,611★, Matt Pocock, x.com/mattpockuk/status/1979531172968374300) — TS-native local evals with git-diffable snapshots. WATCH — nicest fit if evals ever move into a TS repo's test suite. deepeval (16,680★) / autoevals (962★) — overlap no-mistakes / cheap judge diversity. SKIP. edonadei/caliper (27★ — eval harness for Claude Code skills specifically) — WATCH, exactly our shape, unproven.

9. Model routers / gateways

claude-code-router — musistudio/claude-code-router · 35,611★ (pushed Jul 6; 1,194+ issues = heavy real usage) · Local proxy intercepting Claude Code's own requests → any provider (OpenRouter, DeepSeek, GLM, Kimi, Ollama) with per-request routing rules (e.g. background/subagent traffic → cheap model). Vs ours: fleet + OpenClaw route at the skill/session layer; nothing in the stack routes inside Claude Code per-request. Natural fit for the "GLM overflow when Anthropic budget runs low" lane, and composable with pxpipe (both are BASE_URL proxies — chain order matters, test it). VERDICT: ADOPT (trial) — the one router with a real slot in this stack.

9Router — decolua/9router · 20,356★ · free-tier arbitrage across 40-60 providers with auto-fallback; endorsed by @VaibhavSisinty ("trending #5 on GitHub"). Vs ours: quota arbitrage is ToS-fragile and duplicates CCR minus the discipline. VERDICT: SKIP.

LiteLLM (52,793★) — team-scale gateway; noted as a possible fix for OpenClaw's fallbacks: [] outage-stall TODO, but heavy for one user. WATCH. Plano/archgw (6,678★, DigitalOcean-acquired), Bifrost (6,298★, 5k-RPS-grade), Portkey (12,337★, Palo-Alto-acquired, cadence slowing), 1rgs/claude-code-proxy (3,671★, strictly narrower than CCR), y-router (381★, stalling), RouteLLM (abandoned 2024) — SKIP (all).

10. Cost / token tooling (benchmark: pxpipe's measured ~86%)

headroom — chopratejas/headroom · 57,175★ · AST-aware semantic compression of tool outputs/logs/RAG chunks, 60-95% claimed (self-reported). Already installed in this environment as the headroom MCP — the finding is that we haven't dogfooded it. Vs pxpipe: different mechanism (structural vs render-to-PNG), works on all models where pxpipe only pays off on Fable/GPT-5.6. Complementary pair: headroom for cheap-model runs, pxpipe for Fable runs. VERDICT: ADOPT (dogfood what's installed) — run headroom_stats after a week of real use; measure, don't trust the 92% claim.

claude-code-cache-fix — cnighswonger/claude-code-cache-fix · 348★ (pushed Jul 6) · patches a prompt-cache invalidation regression causing up to 20x cost spikes on resumed sessions. Vs ours: orthogonal to all compression — it preserves cache hits rather than shrinking input. Directly relevant to the "$20 hey" context-tax lane from RESEARCH_frontier2. VERDICT: ADOPT (verify first) — confirm the regression exists on CLI 2.1.202 before patching anything.

claw-compactor (2,188★, quiet since Apr) — 14-stage reversible AST compression as a skill; the lossless complement for the exact-recall content pxpipe explicitly can't carry. WATCH. tamp (86★, lossless BASE_URL proxy, ~52.6%) / squeez (162★, bash-output-only) — under pxpipe's bar. SKIP. LLMLingua (6,415★) — perplexity pruning mangles code identifiers; wrong fit for coding agents. SKIP.

11. Voice / dictation (user runs Wispr Flow)

Handy — cjpais/Handy · 25,897★ · free, fully-offline cross-platform STT (Whisper/Parakeet/Moonshine); HN front page (item 46628397). Vs Wispr Flow: the offline/free counter — a fallback, not an upgrade. WATCH. VoiceMode — mbailey/voicemode · 1,258★ · bidirectional voice conversation with Claude Code (local Whisper + Kokoro), /voicemode:install. Different lane than dictation. WATCH — trial if hands-free driving ever matters. VoiceInk (5,437★, $39-once Mac dictation) — SKIP (mid-tier vs Wispr Flow). Spokenly (closed) — inverts the flow: the agent calls ask_user_dictation when it needs input. The one mechanically novel voice idea in the sweep. WATCH. talkito (55★) / mcp-voice-hooks (119★) / voice-mcp (15★) — SKIP (niche variants).

12. Ambient presence delta (beyond the known island category)

Claude HUD — jarrodwatts/claude-hud · 26,197★ · not a notch app: a statusline plugin rendering a 4-line in-terminal HUD (context %, active tools, running subagents, todo progress, rate-limit countdowns) — "htop for Claude Code," called the most popular Claude Code plugin in existence; multiple independent write-ups + forks. Vs ours: zero-conflict with Mission Control (in-terminal vs app) and with the banked "judgment strip" plan — also a design reference for it. VERDICT: ADOPT — cheapest ambient-awareness win in the whole sweep.

MioIsland (510★, independent notch entrant) · claude-island (2,435★ — but previously named "vibe-notch"; possible rebrand of known territory, flagged) · atoll (0★ Open Island fork with inline diffs) · pookify (31★, 6 days old) — the island lane stays saturated; nothing changes the frontier2 verdict. SKIP/WATCH (category noted). tmux-agent-status (222★) / tmux-agent-sidebar (299★) / tmux-agent-indicator (70★) — live tmux sub-genre; we live in cmux, not tmux. SKIP. ClaudeBar (1,297★, quota menubar) — CodexBar territory. SKIP.

13. The "can't work without X" genre (tools named in high-signal posts)

  • lazygit — jesseduffield/lazygit · terminal git UI; "the real MVP — staging individual hunks while claude code rewrites your codebase is a lifesaver" (@bellman.pub). Vs ours: a real gap — no hunk-level review tooling in the stack while agents edit. VERDICT: ADOPT.
  • Ghostty — repeatedly the named terminal for long streaming sessions (Morgan @morganlinton cheatsheet thread; multiple blogs); GPU rendering kills stream lag. Vs ours: orthogonal to cmux; cheap to try. WATCH.
  • Warp agent toolbelt — commercial terminal-with-agents; duplicates existing hooks workflow. SKIP.
  • Intercom's internal platform (@brian_scanlan, amplified by Intercom CEO): 13 plugins / 100+ skills / hooks as "a full-stack engineering platform" — no OSS names extractable; validates the private-skill-library strategy.
  • Counter-signal worth keeping: Boris Cherny (Claude Code's creator) — "my setup might be surprisingly vanilla" (x.com/bcherny/status/2007179832300581177). The maximalist toolbelt is not the only 0.1% posture.

ADOPTION QUEUE (ranked — install command + first-use test)

Install commands are best-effort from this sweep; confirm against each README before running (standing verify-packages-before-install rule: npm view <pkg> first).

  1. claude-mem — auto cross-session memory, the one missing layer. Install: npm view claude-mem && npm i -g claude-mem && claude-mem install Test: open a fresh session in a repo touched yesterday, ask "what were we doing in the last session?" — it should answer from injected memory, then check the token overhead it added.
  2. Claude HUD (jarrodwatts/claude-hud) — in-terminal agent HUD. Install: per README (plugin marketplace: /plugin marketplace add jarrodwatts/claude-hud → /plugin install claude-hud). Test: run a subagent fan-out and confirm context %, live tools, and subagent count render in the statusline.
  3. lazygit — hunk-level git review while agents edit. Install: brew install lazygit Test: after an agent edits 3 files, stage individual hunks from one file and discard one hunk.
  4. headroom (already installed — dogfood) — no install. Test: compress one 50k-token log via headroom_compress, verify retrieval fidelity via headroom_retrieve, read headroom_stats after a week; keep only if measured savings are real on cheap-model runs.
  5. claude-code-router (musistudio/claude-code-router) — per-request routing inside Claude Code (GLM overflow lane). Install: npm view @musistudio/claude-code-router && npm i -g @musistudio/claude-code-router Test: route background/subagent traffic to GLM on one disposable session; verify main-loop stays on Anthropic OAuth and nothing double-proxies with pxpipe (chain order test).
  6. tdd-guard (nizos/tdd-guard) — mechanical test-first gate. Install: npm view tdd-guard && npm i -g tdd-guard + PreToolUse hook per README. Test: on a Vitest repo, ask for a feature without writing a test first — the Edit should be blocked; then do it test-first and confirm it passes through.
  7. brood-box (stacklok/brood-box) — microVM isolation for unattended runs. Install: per Stacklok README (fresh project — verify binary/tap). Test: run one nightly-loop-shaped task inside it; confirm COW snapshot review shows the diff before anything touches the real repo, and the egress firewall logs blocked domains.
  8. serena (oraios/serena) — symbol-graph code navigation on the biggest repo. Install: claude mcp add serena -- uvx --from git+[github.com/oraios/serena](https://github.com/oraios/serena) serena start-mcp-server Test: "find all references to X and rename the symbol" on the largest active codebase; keep only if it beats native search on the first three asks.
  9. claude-code-cache-fix — verify-then-adopt. Test first: reproduce the resumed-session cache-invalidation cost spike on CLI 2.1.202 with a receipts diff (Mission Control has the data); only patch if the regression is real here.
  10. vercel skills CLI — discovery/distribution layer for the skill library. Install: npx skills (browse) — Test: search the leaderboard for one skill worth cribbing; evaluate publish flow for any private skill going public.
  11. superpowers (obra/superpowers) — mine, don't wholesale-install. Test: read the methodology end-to-end, diff against fable-judgment + unknowns-discipline, cherry-pick ≤3 skills into the private library with attribution.
  12. Linear MCP (official hosted) — claude mcp add --transport http linear [mcp.linear.app/mcp](https://mcp.linear.app/mcp`) Test: create one issue from a session and confirm it lands in the same project the deadline board exports to.

HONEST NULLS

  • X was rate-limited (429) in all five lanes, every attempt. Endorsements above are web-indexed snippets/blogs/HN, not primary like/view counts. The two solid named practitioner endorsements: Chip Huyen → sniffly, Hamel Husain → inspect_ai (plus Matt Pocock for his own evalite).
  • Nothing verifiably beats pxpipe's measured ~86% — headroom (60-95%) and LLMLingua (~95%) are self-reported; the honest movers are complements (lossless claw-compactor, cache-preservation) not replacements.
  • Voice: nothing beats Wispr Flow for the actual use case (fast dictation into agents). The only novel mechanic is Spokenly's agent-initiated ask_user_dictation.
  • No SketchyBar agent widget and no Hammerspoon agent-presence Spoon with traction — TermGrid has no wild sibling.
  • No skill marketplace has a money layer yet — the land-grab is package-manager infrastructure, not revenue.
  • "hns" (voice lead) could not be identified as a real tool. Spritz (claude-code-sandbox successor) named-but-unlocatable. braintrust has no canonical open repo (closed SaaS).
  • Zero X toolbelt post in-window yielded exact engagement numbers (X fetch 402s + opencli 429s); the Intercom thread (13 plugins / 100+ skills) is the highest-credibility genre artifact and names no OSS tools.
  • Dead/deprecated confirmed this pass: crystal (→ closed Nimbalyst), Terragon (shut down), uzi, claudebox, webdevtodayjason/claude-hooks, apple-mcp, RouteLLM, textcortex/claude-code-sandbox (archived).