Open source. Inside your repository.
Give your coding agent a way to work. Nexus brings focused skills, specialist agents, and an issue-to-PR workflow to Claude Code and Codex.
MIT licensed · Inspectable Markdown · Your host, your tools
The working agreement
01—03Read the issue, inspect the repository, and reproduce the failure.
Protect unrelated work. Follow the project’s own conventions.
Report checks, limitations, and the actual pull request status.
Reusable skills
Specialist agents
Terminal commands
Current catalog
01 / A complete workflow
Start with a specific issue, a bug described in chat, or your current directory. The workflow carries context through each stage.
A local repository. A GitHub issue or a bug in chat. One workflow from investigation to a pull request.
Example prompt · replace #123
Use software-engineer in the current repository. Fix issue #123, run the relevant checks, and raise a PR.
Requires a local checkout and authenticated GitHub access for issue retrieval and PR creation. No issue supplied? The workflow inspects open issues and selects a bounded candidate.
Illustrative workflow · not a live execution
Read the report. Find existing work. Define a reproducible failure.
Issue + acceptance criteria
Follow the code and the evidence before choosing a fix.
Reproduction + diagnosis
Protect unrelated edits and follow the repository’s conventions.
Patch + regression coverage
Run relevant checks and identify anything that could not be verified.
Checks + limitations
Explain the change, include evidence, and report the actual CI status.
Pull request · never auto-merged
02 / Make it yours
Install the plugin, open a repository, and choose a workflow. Everything ships as source you can read and adapt.
Read the quickstartClaude agent registration and memory are host-specific. Codex plugin installation has been checked locally; model-driven routing and GitHub execution require separate verification.
Inside a Claude Code session
/plugin marketplace add aayushostwal/nexus
/plugin install nexus@nexus-marketplace
/reload-pluginsReload the plugin, then invoke /nexus:software-engineer in your repository.
GitHub access is configured separately. Bundled Node helpers require Node.js 18+.
03 / The skill library
Focused instructions for debugging, testing, reliability, and more. Choose a skill by intent; inspect the source before making it part of your workflow.
11 of 11 workflows
Use when something is failing, regressing, or behaving unexpectedly and the goal is root-cause analysis, narrowest fix, and verification. Covers CI/CD, tests, runtime, deployment, and tooling failures. Output should be RCA + fix + prevention, not design options.
Use for persistent TODO capture, daily brief retrieval, and shared Codex/Claude operating rules. Trigger on remember/track/follow-up requests, "nexus add/show/brief", or policy/workflow questions across Slack, Outlook, and Jira contexts.
Use for correlated API failures, cascading errors, and distributed-system incident tracing. Trigger on multi-service error spikes, dependency-chain analysis, circuit-breaker events, or requests to identify origin service and blast radius from logs/metrics across components.
Use for memory, CPU, latency, and query-performance investigations, plus dependency-upgrade blast-radius analysis. Trigger on leak/OOM/profiling/regression reports or "what breaks if I upgrade X." Route to memory-leak or dependency-blast-radius workflow based on intent.
Use for production incidents, outages, and service degradations, plus release-readiness gate checks. Trigger on 500 spikes, SLA breaches, latency/error surges, on-call alerts, rollback decisions, and RCA/post-mortem requests. Prioritize stabilization, timeline reconstruction, and evidence-led actions.
Use to convert any input into a ready-to-record YouTube Shorts script. Trigger on requests for 30-second scripts, hooks, reels/short-form video copy, or pasted topics/articles/bullets/URLs needing concise high-retention narration with CTA.
Use for creating, refactoring, or debugging skills and improving SKILL.md trigger quality. Trigger on "write/create/update/improve skill", trigger-miss complaints, pasted-skill feedback, or repeated workflows that should be skillized. Output production-ready skill structure and routing logic.
Resolve a GitHub issue or a concrete bug described in chat in a local repository, from reproduction through a verified fix and pull request. Use when asked to implement an issue or run the issue-to-PR workflow; not for read-only review, issue listing, or architecture planning.
Try in Claude Code
/nexus:software-engineer fix issue #123 in the current repository and raise a PRUse for flaky or nondeterministic tests, especially local-vs-CI inconsistencies. Trigger on intermittent failures, race/timing symptoms, retry-only greens, or unexplained skip/xfail markers. Output should isolate reproducible cause, narrowest durable fix, and prevention guidance.
Always-on token optimization rules. Apply on every response.
Use for creating executable Jupyter tutorials and AI engineering walkthroughs with runnable cells. Trigger on requests for step-by-step guides, notebook-based teaching, or shareable code-first learning content. Prioritize reproducibility, clarity, and copy-paste-ready outputs.
04 / Specialist perspectives
Claude Code agent definitions for architecture, product, infrastructure, and code review. Tool declarations and memory scopes are visible in each definition.
14 of 14 specialists
Use this agent to design, build, or review LLM-powered features and products end-to-end: model selection, RAG architecture, agent design, eval pipelines, token economics, and LLM observability. Trigger on "design this LLM feature", "review my RAG pipeline", "which model should I use for X", "why is my agent failing", "estimate token costs", or eval-strategy questions. Returns designs with cost projections, eval plans, and failure-mode tables; or reviews with severity-ranked findings tied to specific production scenarios.
Personal memory
Configured to retain context across your projects in Claude Code.
Tools: Inherited from host
View sourceUse this agent for cloud cost audits, bill-spike investigations, cost monitoring setup, and "find me a cheaper alternative" requests. AWS primary; GCP/Azure aware. Trigger on "why is my AWS bill high", "audit our cloud spend", "set up cost alerts", "cheaper alternative to X", or monthly-cost reviews. Returns a current-spend table from real billing data, a prioritized findings table with savings estimates, copy-paste quick wins, and structural changes by quarter.
Personal memory
Configured to retain context across your projects in Claude Code.
Tools: Bash, Read, Grep, Glob, WebSearch, WebFetch
View sourceUse this agent for senior-level review of a PR, branch, or diff. It inspects changes for correctness bugs, regressions, security issues, migration risk, and deploy safety, then returns a verdict with file:line findings. Trigger on "review this PR", "review my branch", "is this safe to merge", pasted diffs, or pre-merge risk checks. Read-only — it never edits code.
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Bash, Read, Grep, Glob
View sourceUse this agent for token-efficient codebase exploration: mapping repo structure, finding where a feature lives, locating entry points and config files, or answering "where is the code for X". Trigger on "explore this repo", "where is X implemented", "what's the entry point", "find the config for Y", "map the structure". Read-only. Returns a concise three-column table (Title | File path | Description) consumable by humans and downstream agents.
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Bash, Read, Grep, Glob
View sourceUse this agent for database design and change work: schema design, indexing strategy, query optimization, migration safety, and engine selection. Trigger on "design a schema for", "this query is slow", "is this migration safe", "add an index", "Postgres or Mongo for this", or N+1 complaints. Returns schema/DDL with per-decision rationale, or a numbered migration plan where every step has a verify command and a rollback command.
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Inherited from host
View sourceUse this agent to build a documentation application as a React app — from a repo's README, docs folder, or code. Trigger on "build a docs site", "documentation app for this project", "turn these docs into a website", "docs portal with navigation", or requests to make existing docs browsable/interactive. Returns a runnable React app with sidebar navigation, Mermaid diagrams, reference tables, search, and interactive components — plus the commands to run it.
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Inherited from host
View sourceUse this agent to design or review asynchronous, event-driven systems: queues, streams, pub/sub, sagas, outbox patterns, DLQs. Trigger on "design an event pipeline", "Kafka or SQS", "review our consumer", "we're losing/duplicating messages", "how do I make this async", or distributed-transaction questions. Returns a design (Mermaid flow, component table, failure-mode table) or a review (findings with severity and the exact data-loss or duplication scenario each finding causes).
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Inherited from host
View sourceUse this agent to design cloud architecture and write the Infrastructure-as-Code for it (Terraform, CDK in Python, CloudFormation), or to audit existing IaC for risk and waste. Trigger on "design infra for X", "write the Terraform for", "review our IaC", "is this stack production-ready", or greenfield deployment questions. Returns either a full design (diagram, costs, trade-offs, scaling tiers) or an audit (issues with file:line evidence, short-term and long-term remediation plans).
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Inherited from host
View sourceUse this agent to design or review mobile UX for iOS, Android, React Native, or Flutter apps. Trigger on "design a mobile flow/screen", "how should this work on mobile", "review my app's UX", "navigation structure for the app", or "onboarding/permissions flow". Returns a screen-by-screen spec (purpose, layout, components, states, transitions) with ASCII wireframes, or review findings in a severity-ranked table.
Personal memory
Configured to retain context across your projects in Claude Code.
Tools: Inherited from host
View sourceUse this agent to write or critique Product Requirements Documents. Trigger on "write a PRD", "draft requirements for", "turn this into a PRD", "I have a feature idea", or "review/critique this PRD". Write mode turns rough notes into a complete 10-section PRD; Critique mode returns a severity-ranked gaps table plus rewritten worst sections. Returns a delivery-ready document, never a partial outline.
Personal memory
Configured to retain context across your projects in Claude Code.
Tools: Read, Write, Grep, Glob, WebSearch
View sourceUse this agent when direction is known and the user needs an execution-ready roadmap: scope, sequencing, dependencies, risks, validation, rollout, and rollback. Trigger on "plan this feature", "create a roadmap", "how do I roll this out", "migration plan", or "break this down into steps". Returns a scoping table for approval, then a dependency-ordered implementation plan with verify commands and rollback procedures. It plans — it never implements.
Personal memory
Configured to retain context across your projects in Claude Code.
Tools: Bash, Read, Grep, Glob, Write
View sourceUse this agent to plan how a system scales from its current load to a growth target. Trigger on "can this handle 10x traffic", "we expect N users next quarter", "scaling plan", "capacity planning", or "what's our next bottleneck". Returns a current-capacity assessment, a bottleneck-ordered tiered scaling plan with trigger metrics and cost estimates, and a single next action. Read-only against the codebase; uses web search for service limits.
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Bash, Read, Grep, Glob, WebSearch
View sourceUse this agent for codebase architecture mapping, coupling analysis, bounded-context discovery, monolith-split planning, and deployment-safety review. Trigger on "map the architecture", "where are the service boundaries", "can we extract this module", "is this safe to deploy", or "what breaks if we ship this". Returns either an Architecture Map (layers, dependency graph, coupling matrix, extraction candidates, C4 diagram) or a Deployment Safety Report with a GO / CONDITIONAL GO / NO-GO verdict and rollback plan. Read-only — it never edits code.
Repository memory
Configured to retain context within this project in Claude Code.
Tools: Bash, Read, Grep, Glob
View sourceUse this agent for read-only UI/UX review of web apps — components, pages, flows, or screenshots. Trigger on "review this UI", "UX feedback on", "is this accessible", "audit this page/component/flow", or pasted screenshots of an interface. Returns a verdict, a severity-ranked findings table with file:line locations, a brief what-works-well section, and the top-3 prioritized fixes. Read-only — it never edits code.
Personal memory
Configured to retain context across your projects in Claude Code.
Tools: Bash, Read, Grep, Glob
View sourceSkills and agents are readable source files. Review what you install and change what your team needs.
The workflow follows local instructions, existing tests, and the conventions already in your codebase.
Passing local checks, pending CI, and unverified behavior are different states. The workflow reports them separately.