convez.ai
SkillDevOps & InfraestruturaIntermediário
bendourthe

Nexus-Hub

bendourthe/Nexus-Hub

Harness de skills para cada assistente de IA coding

Ver no GitHub
BomPluginScore · de 0 a 100

Catálogo upstream de skills enterprise para assistentes de IA coding, com 252 skills curadas, 14 comandos e 22 hooks. Funciona nativamente em Claude Code, GitHub Copilot, Cursor e outros agentes, com instalação cross-platform e zero telemetria.

Instalação

$npx convez-ai bendourthe/Nexus-Hub

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
0
Acompanhando
0
Saúde
100%
Último commit
há 1 dia
Versão
v4.15.3
Issues abertas
0
Contribuidores
1

Casos de uso

  • Enriquecer assistentes de IA com 250+ skills curadas para coding
  • Unificar workflow de desenvolvimento entre múltiplos agentes (Claude, Copilot, Cursor)
  • Automatizar tarefas repetitivas de engenharia com hooks e comandos

Veredito

Solução sólida e madura para teams que usam múltiplos agentes IA — oferece cobertura real (252 skills) e transparência (reverse-eng first, zero telemetria). A instalação é simples, mas exige familiaridade com VS Code/terminal.

README

A tradução para o português deste README ainda está na fila.

Nexus-Hub

Nexus-Hub is the upstream skill catalog for AI coding assistants: 252 skills, 14 commands, 22 hooks, 23 agents, and 4 language rule families. It installs in one step on Windows, macOS, and Linux, and it works the same across Claude Code, OpenAI Codex, Gemini (via Antigravity), GitHub Copilot, Cursor, GitHub CLI, and the sibling Nexus desktop app and VS Code extension. The catalog is reverse-engineering-first by policy: zero third-party data processors, zero outbound calls from skills / commands / hooks, zero telemetry.

Interactive Guide -- start here

New to Nexus-Hub? Open the interactive guide. It is a self-contained, click-through walkthrough of the entire workflow -- install, onboard an unfamiliar codebase, plan, implement, harden, and ship -- with simulated VS Code / terminal sessions and the artifact each command produces. It is the fastest way to get a teammate productive, and it doubles as a live-demo-quality presentation.

  • File: guide/nexus-hub-guide.html -- one HTML file, fully offline, no server or install required.
  • To view it: GitHub does not render HTML inline. Open the file above and click Download raw file (top-right of the file view), then open the downloaded .html in any browser. Or clone the repo and double-click it.
  • To share it: send that single file to anyone on the team. See guide/README.md for maintainer notes.

Renamed from DevAI-Hub at v2.0.0 to align with the sibling project Nexus, a local-first desktop AI Studio that consumes Nexus-Hub as its upstream skill feed. Existing ~/.devai-hub/ installs are migrated in place by the v2.0.0 installer on first run; see docs/archive/v2/v2.0.0/RELEASE_NOTES.md for the full migration story.


How Nexus-Hub fits with Nexus

Nexus-Hub and Nexus are two halves of the same idea, split along a deliberate seam.

  • Nexus-Hub (this repo) is the catalog: 252 curated skills, 14 commands, 22 hooks, 23 agents, 4 rule families, plus 4 internal MCP servers (nexus-skill-server, nexus-code-search, nexus-web-fetch, nexus-context-compressor). It is content-only, platform-agnostic, and shipped via an installer that writes to ~/.nexus-hub/ and into each AI assistant's per-platform config locations.
  • Nexus is a local-first desktop AI Studio that consumes Nexus-Hub as its skill feed. Nexus's AGENTS.md names this repo as "the only external project we deliberately link to" -- the upstream feed for its skill harness.

The two projects are designed to be useful independently: you can install Nexus-Hub into any supported agent platform without touching Nexus, and Nexus can run with or without the upstream catalog wired in. The combination is what gives a single curated skill set to every agent surface a developer touches: terminal, IDE, desktop app, and CLI.


What's New in v3.3.0

v3.3.0 adds a skill-native loop-engineering layer. Every change is local-first: zero new outbound calls, dependencies, credentials, or third-party processors, and no new slash command.

Highlights:

  • loop-engineering workflow skill (catalog/skills/workflow/loop-engineering/): a connective layer that composes Nexus-Hub's existing primitives into named, goal-terminated agentic loops. It ships a loop-definition schema (name, goal, iteration_cap, check_command, exit_condition, driver, maturity, agents, tags), a seeded local loop library (three full archetypes -- ship-pr-until-green, build-until-green, e2e-until-green -- plus two mapped examples that defer to the first-class /test and /review changes commands), a five-pieces-to-primitive mapping, and a scheduled-triage recipe expressed entirely with owned primitives. The loop driver remains the host platform's /loop and /goal; Nexus-Hub references those commands and never reimplements them. The hosted-gallery pattern was reverse-engineered into a purely local, service-free registry (no install counts, no remote fetch).
  • Goal-based stopping + independent-evaluator methodology: agent-orchestration-primitives and verification-before-completion now teach that a loop's exit condition must be a falsifiable check_command evaluated by a checker that did not produce the work -- and name two loop human-cost anti-patterns (cognitive surrender, comprehension debt) with mitigations, cross-linking session-teach-back as the comprehension-debt countermeasure.

Catalog: 252 skills across 21 categories. See CHANGELOG.md for the full entry and docs/v3.3.0/ for the plan and known gaps.


What's New in v3.2.0

v3.2.0 lands two reverse-engineer-first adoptions and removes the v3.x command shims. Every change is local-first: zero new outbound calls, credentials, or third-party processors.

Highlights:

  • Internal nexus-context-compressor engine (extensions/nexus-context-compressor/): a local-first, reversible, zero-outbound replacement for the external rtk Rust binary. It routes tool output to deterministic strategies (JSON-array dedup, AST-aware code-body elision reusing the nexus-code-search tree-sitter extractors, KV-cache prefix stabilization), persists every dropped span behind a reversible <<ccr:HASH N_rows>> marker (a local SQLite store resolves it back on demand), and ships an opt-in PreToolUse hook + an internal MCP context_compress / context_retrieve tool. An offline, deterministic accuracy-regression gate guards fidelity in CI, and an optional default-off ML token-dropper (public pre-trained ONNX weights, offline-capable) is available behind an extra. This is the 4th internal MCP server.
  • session-teach-back skill: a Socratic mastery-confirmation loop that quizzes the operator on what a session built until every concept is confirmed. Skill-native, no new code or outbound call.
  • Removed the 40 v3.x deprecation command shims (breaking): the old command names no longer resolve. The consolidated 14 commands + 3 permanent aliases (/commands, /commit, /constitution) remain; the old-to-new mapping is in docs/v3.0.0/command-migration.md.

Catalog: 251 skills across 21 categories. See CHANGELOG.md for the full entry and docs/v3.2.0/ for the plans, architecture writeup, and known gaps.


What's New in v2.1.0

v2.1.0 makes Spec-Driven Development (SDD) a first-class workflow in Nexus-Hub by adopting 11 capabilities surfaced by the v2.0.0 cross-project comparison with GitHub Spec Kit. The release is fully additive -- every change is opt-in, and pre-existing command and skill behavior is preserved.

Highlights:

  • 4 new slash commands: /constitution (project-governance file with versioned MUST / SHOULD principles), /analyze-spec (read-only cross-artifact analyzer with severity-tagged findings), /clarify-spec (sequential 5-question ambiguity-reduction loop), /tasks-to-issues (convert strict-format task lines into linked GitHub issues via the local gh CLI).
  • 3 new skills: project-constitution, cross-artifact-analyzer, tasks-to-issues.
  • 3 new templates: constitution-template.md, spec-template.md (with P1 / P2 / P3 user stories + FR-### / SC-### IDs), spec-quality-checklist.md.
  • /generate-plan updates: opt-in --specs-layout flag for the specs/<NNN>-<slug>/ directory layout, strict - [ ] T### [P?] [US?] file_path task-line format with phase organization (Setup / Foundational / User-Story / Polish), Constitution Check + Complexity Tracking sections in every generated plan.
  • [NEEDS CLARIFICATION] marker discipline with a 3-marker hard cap prioritized scope > security/privacy > UX > technical, wired into spec-driven-development, ambiguity-detector, and idea-refine skills.
  • Integration Registry (G12, pulled forward from v2.2.0): a Python class hierarchy under scripts/lib/integrations/ that owns per-platform install logic. Adding a new AI assistant is now one ~30-line subclass + one register line instead of editing 5+ instruction templates in lockstep. See docs/archive/v2/v2.1.0/adr/adr-001-integration-registry.md for the architecture decision record.
  • Expanded platform support (3 new platforms): Antigravity 2.0 (Google), Gemini CLI (Google), and Nexus-AI (https://github.com/bendourthe/Nexus-AI) join the existing Claude Code / Codex / Cursor / Gemini / Antigravity 1.0 / GitHub Copilot / OpenCode list. The installer auto-dispatches to all 10 platforms so switching assistants is seamless.

All 11 SDD adoption items are classified skill-native under the MCP Registry Policy -- no new outbound calls, no new credentials, no new third-party data processors, no new runtime dependencies. The Integration Registry is also stdlib-only Python.

See CHANGELOG.md and docs/archive/v2/v2.1.0/RELEASE_NOTES.md for the full v2.1.0 entry, and docs/archive/v2/v2.1.0/plans/adoption-spec-kit.md for the 10-phase plan.


What's New in v2.0.0

v2.0.0 is a rename + brand modernization release. There are three headline changes:

1. Renamed from DevAI-Hub to Nexus-Hub (breaking)

The repository, distributed artifact, plugin metadata, installer, MCP servers, extensions, scripts, skills, commands, hooks, agents, rules, templates, and every documentation surface have been renamed to Nexus-Hub. The complete list of breaking changes is in CHANGELOG.md under [2.0.0] -- Breaking changes. The headline items:

  • Installed root: ~/.devai-hub/ -> ~/.nexus-hub/
  • Plugin name: devai-hub -> nexus-hub
  • MCP server names: devai-skill-server -> nexus-skill-server (and the other two)
  • Environment-variable prefix: DEVAI_* -> NEXUS_*
  • GitHub URL: bendourthe/DevAI-Hub -> bendourthe/Nexus-Hub (GitHub's automatic rename redirect handles in-flight links)

Existing DevAI-Hub installs are migrated in place by a one-shot prompt at the top of the v2.0.0 installer. No symlinks, no compatibility shims -- a single user-visible event.

2. Modernized installer

Both scripts/installer.sh and scripts/installer.ps1 now print an ASCII-art NEXUS-HUB banner at startup, followed by the tagline, version, and repo URL. The legacy-install migration logic runs immediately after the banner and before the welcome prompt.

3. Explicit Nexus brand linkage

This README and the top-level agent instruction files (AGENTS.md, CLAUDE.md) now describe the relationship between Nexus-Hub and Nexus directly, so the two-project shape is obvious to anyone landing on either repo.

See CHANGELOG.md and docs/archive/v2/v2.0.0/RELEASE_NOTES.md for the full v2.0.0 entry.


Supported Agentic Platforms

PlatformInstall targetPathPer-platform surface
Claude Code (Anthropic)~/.claude/ + project .claude/legacy + registryFull: skills, commands, hooks, agents, rules, MCP configs
OpenAI Codex CLI~/.codex/ + project .codex/ + AGENTS.mdlegacy + registryFull: skills (under skills/), commands (under prompts/), agents, rules
Gemini (IDE / Antigravity 1.0)~/.gemini/ + project .gemini/GEMINI.mdlegacy + registryFull: skills, commands (under workflows/), agents, rules
Gemini CLI (Google, ENTERPRISE-ONLY post-2026-06-18)~/.gemini/commands/*.toml + project .gemini/commands/*.tomlregistry (new in v2.1.0; gated behind --enterprise / -Enterprise flag in v2.2.0)TOML-format custom commands generated from catalog/commands/*.md. Non-enterprise users transition to Antigravity CLI before 2026-06-18 per the 2026-05-21 Google announcement.
Antigravity 2.0 + CLI (Google)~/.gemini/antigravity-cli/ + project .agents/registry (new in v2.1.0, CLI coverage added v2.2.0; paths verified v2.3.0)Full: skills, commands (under workflows/), subagents, rules. Single integration covers both the desktop IDE and the standalone Antigravity CLI (agy binary), verified 2026-05-29 against Google's public Antigravity CLI docs.
GitHub Copilot (VS Code)project .github/copilot-instructions.mdlegacy + registryBehavioral guardrails (skill index embedded as text); merge semantics if the file already exists
Cursorproject .cursor/rules/*.mdc + AGENTS.mdregistryPer-rule .mdc files + behavioral guardrails (skill index embedded as text)
OpenCodeproject AGENTS.md + .opencode/registryBehavioral guardrails + skills mirror
Nexus-AI (Local Studio)~/.nexus-ai/ + project .nexus-ai/registry (new in v2.1.0)Full mirror: skills, commands, agents, rules, hooks, MCP configs, templates
GitHub CLI (gh)via gh copilot extensionindirectSkill / command references via AGENTS.md open standard
Nexus desktop appupstream consumerindirectReads the same catalog as its skill feed
Nexus VS Code extensionupstream consumerindirectReads the same catalog as its skill feed

Coverage caveat: the registry path (introduced in v2.1.0 Phase 10) dispatches install / teardown through scripts/lib/integrations/runner.py and supports a --dry-run mode. The legacy path (the long-standing in-installer copy blocks) continues to be the canonical install for Claude / Gemini / Codex / Copilot until v2.2.0 parity migration (tracked as DF-001 in docs/archive/v2/v2.1.0/known-gaps.md). Both paths produce the same end-state on disk for those platforms; the per-platform installer logic lives in scripts/installer.sh, scripts/installer.ps1, and the per-platform subclasses under scripts/lib/integrations/. Per-platform capability specs (install surface, distributed content, instruction file, quirks) are documented under docs/specs/.

Branch-based install (v2.4.0): pass --branch <name> (Bash) or -Branch <name> (PowerShell) to install the catalog from a pushed branch instead of the current checkout. The installer shallow-clones the repo at <name> into a deterministic cache directory (~/.nexus-hub/branches/<sanitized-name>/) and runs the install from that checkout, so the user's working copy is never touched. The branch name is sanitized for filesystem safety (path-traversal sequences are neutralized). Combine with --check / -Check for a clone-free probe that prints the resolved cache path and clone source.


Quick Start (The 30-Second Setup)

  1. Clone or download this repository.
  2. Run the installer:
    • Windows: double-click install.bat.
    • macOS / Linux: run ./install.sh in your terminal.
  3. Drag and drop your target project folder when asked.
  4. Confirm to install global skills.
  5. (Optional) Select a project to configure workspace-specific rules.

Done. The installer writes to ~/.nexus-hub/ (the user-global catalog) and into each supported assistant's per-platform config locations. If a legacy ~/.devai-hub/ install is detected, you will see a single migration prompt at the top of the run -- answer Y (default) to migrate in place.

After the installer completes:

  • Globally: your user profile has all 252 skills, 14 commands, 22 hooks, 23 agents, plus Gemini and Codex instructions.
  • Locally: your project has copilot-instructions.md and AGENTS.md tailored to your language.

What is Nexus-Hub?

Most AI assistants are generic by default: they know a lot but specialize in nothing. Nexus-Hub is the layer that turns a generic assistant into a specialist for the work you actually do.

It does three things:

  1. Behavioral rules -- per-language code-style and security rules that tell the assistant how to write code in your project (not just whether the code compiles).
  2. Autonomous skills -- 208 curated capability prompts grouped into 22 categories. Each skill has a 3-tier loading model (always-loaded summary, body on trigger, deeper references on demand) so context cost stays proportional to what the agent actually needs.
  3. Workflow awareness -- 36 slash commands that chain skills into multi-step processes (plan generation, phase implementation, deep review, version bump, release notes, session history).

The catalog itself is content; the harness around it is the per-platform installer plus a small set of local MCP servers that surface the catalog to any agent that speaks MCP.


Recommended Workflows

Nexus-Hub provides two opinionated end-to-end workflows. Use these as a starting point and adapt to your project.

New Project Workflow (5 phases)

Build from scratch with an AI coding agent as your primary partner.

1. Planning

Open an AI chatbot (Claude.ai or ChatGPT) and brainstorm: problem, users, core features, tech stack, constraints. End the session by asking the chatbot to produce a structured Markdown implementation plan -- phases with subtasks, each subtask carrying a self-contained prompt the agent can execute.

2. Project setup

  1. Create the Git repo with a three-tier branching model: main / develop / feature/*.
  2. Install the Nexus-Hub toolkit: ./install.sh (macOS / Linux) or install.bat (Windows).
  3. In Claude Code, run /setup project -- bootstraps CLAUDE.md, the directory structure, .gitignore, README.md, DEVLOG.md, and CHANGELOG.md in 8 guided phases.
  4. Save the implementation plan from step 1 to docs/<version>/plans/<slug>.md.
  5. Commit with /commit.

3. Development (core loop)

For each plan phase:

  1. Create a feature branch: feature/phase-N-short-description.
  2. Open a fresh Claude Code session.
  3. Run /implement <slug> <phase> -- walks every subtask, generates and runs tests, applies fixes, runs /update gitignore + /update docs, generates a session-history file, and produces a commit message.
  4. Commit and push the feature branch.
  5. Merge into develop. Repeat for the next phase.

4. Quality assurance (pre-release)

  1. Run /review full -- a 12-phase orchestrator that chains known-gaps collection, health gates, dependency scan, docs / git hygiene, project validators, codebase description (/describe full), and the security, pentest, and full codebase-review scopes.
  2. Read the synthesis report -- it produces a P0 / P1 / P2 / P3 ranked list of findings with a GO / GO-WITH-CONDITIONS / NO-GO verdict.
  3. Address all P0 and P1 findings before release. P2 findings can be deferred to a follow-up patch release; P3 findings are advisory.
  4. Run /review sbom for compliance documentation.

5. Release

  1. Run /update release -- orchestrates version detection, layout cleanup, .gitignore audit, version-bump across all configuration files, CHANGELOG migration, doc sync, and DEVLOG entry.
  2. Merge develop into main, tag the release, and push.

Inherited Project Workflow (2 phases)

For projects you have inherited or need to audit.

1. Primary analysis and deep review

  1. Clone the repo, open it in VS Code, start a Claude Code session.
  2. Run /review full -- the same 12-phase orchestrator from Phase 4 of the New Project Workflow. The synthesis report's prioritized roadmap (P0 / P1 / P2 / P3) becomes your initial backlog.
  3. If documentation is sparse, backfill it: /update docs (README, if missing), /update changelog (from git history), /update devlog, /update refactor (only when the repo has structural issues).
  4. Establish the develop branch if not already present.
  5. Commit the analysis artifacts.

2. Making changes

For each change:

  1. Brainstorm in a chatbot, then run /plan to produce a structured implementation plan saved to docs/<version>/plans/<slug>.md.

  2. Run /implement <slug> <phase> per phase -- identical to the New Project Workflow's development loop.

  3. (Optional) Use git worktrees for parallel work:

    git worktree add ../project-fix feature/security-fix
    # work in a separate Claude Code session, then:
    git worktree remove ../project-fix
    
  4. After all changes land on develop, run /review full again to verify nothing regressed, then /update release and merge to main.

The QA and release steps are identical to the New Project Workflow.


Manual setup (if you do not want to run the installer)

If you prefer to copy things yourself, here is how the repo is organized.

Claude Code (Anthropic)

The most powerful integration -- adds autonomous agent capabilities.

  • CLAUDE.md: the "brain". Copy catalog/CLAUDE.md to your project root and customize.

  • Skills: the "hands". Copy folders from catalog/skills/ to your project's .claude/skills/ folder.

    Example: copy catalog/skills/research/trend-research to enable the trend-research skill.

Gemini (Google) and Antigravity

Optimized instructions for Google's Gemini models, including the Antigravity workspace layout.

  • Gemini instructions: copy templates/ai-instructions/base-gemini.md (or templates/ai-instructions/generic-instructions.md for the legacy template) to .gemini/GEMINI.md in your project or user profile.
  • Skills and workflows: the installer mirrors these to .gemini/skills/ and .gemini/antigravity/global_workflows/ so they appear globally in Antigravity.

GitHub Copilot (Microsoft)

Instructions for VS Code's Copilot Chat.

  • Copy templates/ai-instructions/coding-instructions/{language}.md to .github/copilot-instructions.md.

Codex (OpenAI)

OpenAI Codex CLI integration. Codex reads AGENTS.md at the project root (the open standard, also honored by Cursor / Aider / Jules) plus its user-level config in ~/.codex/.

  • AGENTS.md: copy templates/ai-instructions/base-codex.md content into your project's AGENTS.md.
  • Skills and prompts: the installer mirrors catalog/skills/ to ~/.codex/skills/ and catalog/commands/ to ~/.codex/prompts/. For manual setup, copy each tree to those destinations.

Cursor

Cursor IDE integration.

  • Project rules: copy templates/ai-instructions/base-cursor.md content into .cursor/rules/nexus-hub.mdc at your project root. Use alwaysApply: true in the frontmatter so Cursor applies the rule on every prompt.
  • Open-standard AGENTS.md: Cursor also reads AGENTS.md at the project root, so the Codex setup above covers Cursor too.

OpenCode

OpenCode IDE integration. OpenCode reads AGENTS.md per the open standard.

  • Copy templates/ai-instructions/base-opencode.md content into your project's AGENTS.md.

Development setup

For contributors working on Nexus-Hub (not consumers of the installer), the repo ships a .devcontainer/ at the root. Open the repo in VS Code with the Dev Containers extension installed (or click "Reopen in Container" when prompted) and the post-create hook will install Python tooling (pytest, ruff), the GitHub CLI (gh), and the Claude Code CLI (claude). Authenticate gh and claude once the container is up, then run make validate to confirm the catalog is clean.

The devcontainer is opt-in -- the standard Quick Start above does not require it. It exists for first-touch contributor onboarding and for reproducing the maintainer's environment across machines.


Featured Skills

SkillWhat it does
Architecture DesignSystem decomposition, ADRs, C4 diagrams, and fitness functions.
AI Agent DevelopmentBuild agents with tool use, memory systems, and multi-agent orchestration.
RAG ImplementationEnd-to-end RAG pipelines with chunking, embeddings, and evaluation.
API DesignREST, GraphQL, and gRPC design with versioning and error handling.
Code ReviewA 6-step deep dive (security, performance, logic) before you merge.
Test GenWrites comprehensive unit tests using AAA pattern and mocks.
E2E TestingPlaywright / Cypress automation with page objects and CI integration.
ComplianceChecks code against SOC2, GDPR, and ISO standards.
Trend ResearchResearches Reddit / X for the last 30 days to find trends and write prompts.

The full catalog is at data/SKILL_INDEX.md. Per-category landing pages live under catalog/skills/.


Usage Monitoring

Three complementary ways to track your Claude Code usage limits.

CLI Usage Display (Automatic)

A Stop hook that shows your usage limits directly in the terminal after each Claude Code response. Color-coded and silent when usage is healthy (below 50%).

Usage: Session 72% | Weekly 15% | Sonnet 3%  (Session resets in 28m)

Installed automatically by the Nexus-Hub installer. Requires curl and jq.

VS Code Extension

Monitor usage from the VS Code status bar with a full dashboard.

  • Auto-fetch: reads your OAuth token and fetches live usage data from the Anthropic API.
  • Status bar: shows session and weekly usage percentages with a custom Claude icon.
  • SVG tooltip: hover for theme-aware progress bars with per-metric breakdown and reset timers.
  • Dashboard: click for a full usage dashboard with model recommendations and optimization tips.

See extensions/claude-usage-monitor/ for setup instructions.

/usage Command

On-demand detailed usage report with model-switching recommendations. Auto-fetches from the API (falls back to manual entry if credentials are unavailable).


Safety and Use in Regulated Industries

Nexus-Hub is built on a reverse-engineering-first principle: the catalog ships zero third-party data processors, zero outbound calls from skills / commands / hooks, and zero telemetry. The full threat-model breakdown, industry compatibility matrix, and reporting policy is in SECURITY.md.

Short version:

  • Open-source / hobby / internal commercial software: green. No restrictions.
  • Regulated industries (healthcare, finance, government, life sciences, automotive, industrial): green WITH caveats. Nexus-Hub itself is safe; the caveat is that your chosen LLM provider is where prompts go (use a regulated-cloud option like AWS Bedrock, GCP Vertex AI, Azure OpenAI, or a self-hosted model consistent with your data-protection obligations).
  • Defense / classified / air-gapped: outside Nexus-Hub's threat model. Do your own assessment.

What Nexus-Hub does NOT do: telemetry, analytics, phone-home, third-party data processors, model downloads, API-key requirements. The MCP Registry Policy in AGENTS.md categorically rejects search-as-service, embeddings-as-service, scraping-as-service, and generation-as-service. The authoritative classification of every MCP server ever shipped or considered is at docs/policy/mcp-reverse-engineering-matrix.md.

What is OUT of Nexus-Hub's control: your chosen LLM provider, any MCP server you add outside the Nexus-Hub registry, user-initiated outbound calls (gh, git push, curl), and your own user-authored hooks and rules. See SECURITY.md section 3 for the full caveats.

To report a security issue: email benjamin.dourthe@gmail.com or open a private security advisory at github.com/bendourthe/Nexus-Hub/security.


Roadmap

Nexus-Hub evolves in versioned slices. Each upcoming line item below traces to a concrete plan file under docs/<version>/plans/ (the durable source) and resolves once its [<version>] block lands in CHANGELOG.md. No star gates, no sponsor tiers, no paid features -- the catalog is reverse-engineering-first and stays that way.

FocusTargetStatusSource
Rename DevAI-Hub to Nexus-Hub, modernize installer with ASCII banner, integrate Nexus brand linkagev2.0.0In progressdocs/archive/v2/v2.0.0/plans/nexus-hub-rename.md
Cross-OS CI matrix for installer smoke tests (closes the cumulative DF-003 / DF-005 / DF-006 / DF-007 / DF-008 cluster from v1.1.5 known-gaps)v2.1.0Planneddocs/archive/v1/v1.1.5/ known-gaps cluster
Skill-eval-loop integration into pre-commit (assertion-graded regression guard for high-traffic skills before they ship)v2.1.0Plannedcatalog/skills/workflow/skill-eval-loop/SKILL.md
MCP registry expansion under the existing 5-step policy (reverse-engineer-first; hard-no on search / embeddings / scraping / generation as a service)continuousIn progressdocs/policy/mcp-reverse-engineering-matrix.md

For narrative-style updates on what changed and why, see docs/DEVLOG.md. For the formal Keep-a-Changelog log of every release, see CHANGELOG.md. For the per-version unfinished-work tracker that the next plan reads to decide what carries forward, see docs/<version>/known-gaps.md.


Collaboration

Nexus-Hub is a curated open-source project. While pull requests are typically not accepted from outside contributors, suggestions, feedback, and recommendations are more than welcomed. If you have a better prompt, a smarter rule, or a pattern you would like to see in the catalog, please reach out directly:

I am happy to discuss skill / command / hook proposals, integration ideas for new platforms, or specific use cases -- especially when the proposal aligns with the policy direction of this project (reverse-engineering-first, no third-party data leaks).


License

See LICENSE.

Ver o original no GitHub

Bom para

Relacionados