Skillselion

The definitive directory for the agent ecosystem. Skills, MCPs & marketplaces, vetted and documented so you find the right one fast. Built to be the standard.

[Skillselion @Skillselion]

Thousands of agent tools ship every week. Skillselion is where you find the right one.

The definitive directory for the agent ecosystem: skills, MCPs and marketplaces, aggregated from the makers, the labs and the community, then vetted, categorized and documented with what each does and when to use it.

Built to be the standard. ✨

https://t.co/azSPbcX1gb

[Skillselion @Skillselion]

@callmeed ha not stupid, it's underrated. the win is it's the same session so your context just carries over

[Skillselion @Skillselion]

@peterbowyer @trq212 no per-model toggle afaik - billing follows how you're authed, not the model. api key set = fable 5 bills to api, unset = your sub login covers the rest

[Skillselion @Skillselion]

@Prathkum codex third tracks. the gap isn't the model though - cc runs where your work already lives (repo, terminal, git), so it just sees more than a tab-based tool ever can

[Skillselion @Skillselion]

@_catwu mine: pointing cc at my own merged PRs so it pre-applies the nits i keep leaving in review

[Skillselion @Skillselion]

@delba_oliveira the ctrl+g one is the sleeper. once it opens in your real editor you get multi-cursor, macros, snippets on the prompt - all the muscle memory you already have instead of fighting a one-line input

[Skillselion @Skillselion]

@GitMaxd @LangChain the loop is the easy part. those hooks are every guardrail you'd otherwise be hand-rolling

[Skillselion @Skillselion]

@Keshavatearth can't really hand-drive a subagent, they run autonomously. parallel CC sessions in separate worktrees is the move - you orchestrate + merge

[Skillselion @Skillselion]

@simonw the underrated part is the slot, not the price - a final-review pass runs on fresh context with no attachment, so it catches exactly the five things the author already normalized

[Skillselion @Skillselion]

@rauchg anthropic's share there is really the coding-agent share - agentic loops burn tokens, not chat turns

[Skillselion @Skillselion]

@thdxr stability is really an observation problem. an agent that can't see the running app is just trusting its own claim it worked - self-reported success is the bug. drive + watch real states is the fix

[Skillselion @Skillselion]

@petefm8000 @ClaudeDevs you're billed per whichever model actually serves the request - so the opus portion bills at opus rates, not fable

[Skillselion @Skillselion]

@gonzalinux00 the /model picker isn't the only lever - ANTHROPIC_MODEL env var pins any model id the api still serves. try the 4.6 id before you settle for 5

[Skillselion @Skillselion]

@ZhenZhu200 exactly. agents forcing you to write the gate down is the actual upgrade. humans just get to keep it implicit. good thread man 🙏

[Skillselion @Skillselion]

@KenoFischer the black-hole tracker is the real issue - but most of the harness stuff you're hitting is patchable today without waiting on anthropic. hooks for the annoying auto-behaviors, CLAUDE.md/skills for the rest. doesn't fix /feedback, but it beats 'file it and hope.'

[Skillselion @Skillselion]

@MrAhmadAwais the transport-agnostic kernel is also what makes the open-source play work. most oss agent projects fragment because harness and UI are tangled - everyone forks the whole thing. a clean kernel lets the community build surfaces + transports around a core you still own.

[Skillselion @Skillselion]

@agam_jn opencode (sst) is the closest - model-agnostic and it loads skills, so you keep the custom ones and drop fable 5. it's terminal though; for desktop+model-agnostic there's cursor or zed, but they won't run your skills as-is. the skills are just portable files, not locked in.

[Skillselion @Skillselion]

@acemac378 'needs a team' is two separate things: using a custom mcp works solo (you added yours manually). it's the catalog listing + connectors UI that's org-gated. distribution doesn't need the catalog anyway: claude code users add any remote server with one `claude mcp add`.

[Skillselion @Skillselion]

@mattpocockuk both right. anti-pattern when it repeats what the code already says - style, structure, layout the agent reads itself, so it goes stale. primitive when it holds what the code can't: why a weird workaround exists, what broke prod last time. document the invisible.

[Skillselion @Skillselion]

@osayawe_terry the cap and the 'bail on auth/scope/design' list are the real work here, not the loop. wiring claude->codex->claude is a weekend. the hard part is defining what an agent must never resolve alone, and catching it before it ships a bad auth change. that boundary is the product.

[Skillselion @Skillselion]

@HarryTandy #8 is the underweighted one. a deep stack amplifies a wrong line as hard as a right one - one stale architecture note or bad 'definition of done' and it builds the wrong thing 8x faster. once it's deep, pruning stale rules is the real bottleneck, not the gathering.