Needt

[Needt home] [Needt tasks] [Needt calendar] [Needt habits] [Needt mail]

_{Real Needt screens. /needt calls Needt's own tools over MCP — the same ones the app uses.}

Install

Claude Code — one line:

claude plugin marketplace add teenxgrailed/needt-plugin && claude plugin install needt@needt

(or inside Claude Code: /plugin install needt --marketplace teenxgrailed/needt-plugin; update later with claude plugin update needt@needt)

Claude (web, Desktop) — download needt.plugin, then Customize → Plugins → Add → Upload plugin.

Then start a new chat and just talk — "what's on today?", "help me think about pricing", "find the best way to…", "roadmap for my app". You never have to name a skill.

The skills

/needt and /needt:roadmap need a Needt Pro account — it connects to your planner over MCP. Every other skill works without an account.

How it works

Widgets. In Claude Desktop, where the chat can render widgets, the skills show what has a shape as one — today's list with Done buttons, option cards, money metrics, finding cards with Accept fix, a subject map, a project roadmap as a metro map. Everywhere else the same answer arrives as text. Rules: plugins/needt/skills/setup/references/widgets.md.

A first run that asks. The first time a skill fires, setup asks up to four clickable questions — which plan this AI app is on, what to connect, language and tone, memory — and saves the answers as one memory line. Say "у меня Max" or "switch memory off" later and it updates. /needt:setup runs it on demand.

Fits your plan. The readiness line carries budget: pro | max | ultra | …, and every skill scales to it: on Pro, research runs inline and no sub-agents start unless you ask; on Max, independent sweeps run in parallel on Sonnet; on Ultra, up to eight agents with the main model synthesising. Table: plugins/needt/skills/setup/references/delegation.md.

One readiness line. The first skill you use in a chat prints a line like Needt · ready · project: Acme · tone: casual · search: web, youtube, github · questions: click · agents: yes · memory: 12. Every other skill reads it instead of checking again; the first time it appears it says in one line what it is.

Memory that follows you. When you say "remember…", "never do X again", or correct how something was done, the plugin saves one line to your Needt memory (Saved: …) and every skill treats it as a standing instruction — in any chat, on any device. Secrets, health details and one-off facts are never saved. Turn it off any time in the Needt app.

The skills call each other. One skill owns each kind of request and the others hand over in the same reply: a YouTube link → youtube, an unsure fact → search, a missing answer → asking, a sequence that spans sessions → roadmap, work for today → needt. The map of owners and hand-offs: plugins/needt/skills/setup/references/skills-map.md.

Uses your own skills too. The readiness line counts the other skills in your catalog (skills: n), and every Needt skill hands the part they own to them — a deck to your pptx skill, a deploy to your runbook, a campaign to your marketing skill — briefed from what is already known and delivered in the Needt skill's voice. Rules: plugins/needt/skills/setup/references/user-skills.md.

Uses what your AI already has. Skills route to the tools your host exposes — web search, Firecrawl, Exa, GitHub, Context7, YouTube — and say so when something is missing instead of faking it. On first run, setup lists free ways to add the missing ones in a few lines.

Parallel and cheap. Independent research runs in parallel sub-agents on a faster model; synthesis, judgement and verdicts stay with the main model.

Checks what it installs — and itself. Every release passes a 30-scenario eval (eval/), its own SkillSpector scan (risk 25/100, no high findings) and size caps (each skill under 8 KB, each description under 600 characters). Before recommending a new skill, plugin or MCP server, the plugin scans it with NVIDIA SkillSpector when it is installed (risk 51+ means it is not installed), or reads its files for known red flags.

Asks only when it matters. Reversible actions run and get reported in one line. Sending, posting, paying and deleting always wait for your yes.

What's in this repo

plugins/needt/                 the plugin (skills, manifests, MCP connection)eval/                          release gate: scenarios and results per versionrelease.sh, build-standalone.py  gate + build  skills/<name>/SKILL.md       one folder per skill  .mcp.json                    https://mcp.needt.app/mcp.claude-plugin/marketplace.json   Claude Code marketplace.agents/plugins/marketplace.json  Codex marketplaceserver.json                    MCP Registry entryglama.json                     Glama listingassets/                        icon and screenshots

Privacy

The plugin is Markdown instructions, one small Python script for YouTube transcripts (runs only if yt-dlp is installed), and one MCP connection to https://mcp.needt.app/mcp. No hooks, no telemetry. Your assistant reads and writes your own Needt data through that server after you sign in with OAuth; nothing else is sent anywhere by the plugin.

Privacy policy · Terms

Support and contributing

Questions and bugs: open an issue or write to [email protected]. Ideas for a skill, a better trigger phrase in your language, a sharper rule — pull requests are welcome. Run claude plugin validate plugins/needt before sending one.

Changelog: plugins/needt/README.md.