swarmery

Vendor-neutral agent-development core: generic agents, skills, commands, hooks, and the project-aware agent-work CLI.

37 skills · 8 commands · 13 agents · hooks

Infrastructure & delivery domain pack: Kubernetes/Helm, GitOps promotion, IaC, GitLab CI, cloud CI auth (GCP + AWS), Keycloak. Opt-in; requires core.

11 skills · 1 command · 3 agents

Issue-tracker pack: /jira-fix drives any Jira ticket end-to-end — access preflight, defect-or-change triage, reproduction or test-first evidence, delegated fix/implementation, evidence comment, QA transition. Opt-in; requires core and an Atlassian MCP provider enabled on the machine.

7 skills · 1 command · 1 agent

UAV/drone domain pack: MAVLink-style telemetry, mission planning, embedded/edge runtime.

3 skills · 4 agents

Claude-engineering pack: skills for building/auditing Claude agent systems — agent architecture, tool/MCP design, prompt engineering, Claude Code config, context reliability. Opt-in; requires core.

5 skills

Design-handoff pack: /design-implement takes an exported design and re-expresses it in the project's stack pixel-accurately — computed-style token inventory, reuse-vs-create recon, an approval gate, and a measured pixel diff as the completion criterion. Opt-in; requires core.

3 skills · 1 command · 1 agent

Web/marketing domain pack: SEO, i18n, landing-page CRO, Figma-to-code styling.

1 command · 3 agents

Repo-wide architecture map: /architecture-map generates a machine-readable JSON contract (layers, modules, file-anchored flows) + self-contained HTML viewer, freshness-stamped per commit; the swarmery dashboard serves both on its Architecture page. Also /visualize: turn a document or an explanation into one interactive, self-contained HTML explainer page.

2 skills · 1 command

Multi-account pack: bind a project to one of several Claude Code accounts and run every session under it — /account, a shell wrapper, an optional shell function, and a wrong-account warning. The binding is machine-local, and a binding file committed to git is ignored. Opt-in; requires the swarmery CLI.

1 command · hooks

Context-graph pack: the graft skill plus two hooks — coverage-gated file:line locators injected into a prompt before the first grep, and a blast-radius note after every edit. Opt-in; requires the graft CLI on the machine.

1 skill · hooks

Knowledge-graph pack: the /graphify skill — repo/folder → persistent knowledge graph with query/path/affected tools and HTML/JSON/Neo4j exports. Opt-in; requires the graphify CLI on the machine.

1 skill

IoT domain pack: BLE communication, device telemetry, health-metrics processing.

1 agent

Semantic code-navigation pack: Serena LSP MCP server. Opt-in; requires the serena binary (uv) on the machine.

1 MCP server

Pull-request review pack: /pr-review reviews a GitHub PR and posts verified inline findings, or triages the reviewer comments on it — verify each against the code, fix the valid ones behind the repo's gates, rebut the rest with file:line evidence, reply in-thread. Opt-in; requires the gh CLI.

1 command

Guard mod pack (function hooks): keeps file tools inside the project's roots, redacts secret-shaped values in tool results before Claude sees them (the raw output may still be kept in the local transcript; see the pack README's Limits), routes production-deploy shell commands to the native ask dialog, and can hide agents of packs the project has not enabled. Policy comes from .claude/project.json at runtime. Opt-in; needs neither core nor the swarmery daemon.

Live-session sensor mod (function hooks): reports turn boundaries, tool calls by name and input hash, context fill, rate limits and cost to a local swarmery daemon over its token-guarded link API. The sensor never blocks or rewrites anything; silent when the daemon's token file is absent. With the opt-in control option the dashboard can stop the session's running turn and send it prompts; with the opt-in forecast option one haiku call per user turn predicts what the turn will touch and the daemon scores the forecast against the actual (paths and numbers only); with the opt-in approvals option a permission ask goes to the swarmery approvals queue first and an AskUserQuestion can be answered from the dashboard, the shim or the native dialog taking over whenever the daemon does not decide; with the opt-in swarm option it logs messages between sessions, keeps a shared registry of the files each session is editing (an advisory line when two live sessions edit one file), and lets the dashboard message a session; with the opt-in planBand option a band above the prompt shows the session's swarmery plan phase and its checklist (/phase opens it, criteria tick from the terminal) and a line under the answer flags an editing turn that looks off the running phase. Opt-in; requires the swarmery daemon.