SwarmKit
A specialist agent swarm for Claude Code, OpenCode and Antigravity, built from one source.
How it works
A task comes in.
It is classified. The tier sets how much ceremony it gets.
Routed to the specialist that owns the work: backend-specialist.
Quality gates run in parallel, only where relevant.
Fix a typo, T1: Edit directly; No plan, no gate.
Add an API route, T2: backend-specialist, security-auditor, release-tester.
Refactor auth, T3: explore, backend-specialist, security-auditor, code-proofreader.
- T1 Trivial edits
- T2 Contained domain work
- T3 Cross-cutting work
The roster
33 agents, one cell each. Hover or focus a cell.
- Core 1
- Context and fast work 3
- Domain specialists 19
- Quality gates 4
- Showroom 6
Numbers come from runs, or they do not appear.
- ponytail diff
- net: pending
- benchmarks
- pending
- harness
- bench: Antigravity (agy) gemini-3.8-flash-high. site: Claude Code.
- build log
- 30 human interventions, 4 agent dispatches
Install
./install.sh --all- Claude Code --claude
- OpenCode --opencode
- Antigravity --agy
- 33 agents
- 27 skills
- 11 MCP servers
- 5 slash commands
- --n8n
- --cloudflare
- --all
- --free
- --uninstall
- --help
Reference
The detail behind the scroll. You write every agent, skill and rule once in core/; a small compiler turns them into each CLI's native format and the installer links the results into place.
01 / Install
Install
Clone the repo and run one script; it links everything into place.
git clone https://github.com/ha-lun/swarmkit.git
cd swarmkit
./install.sh --allAnything it replaces is backed up first. Running it again changes nothing. --all is also the default when no flag is given.
Platforms 3
Claude Code
--claudeSpecialists install as subagents. Tool limits are enforced by tool lists and guard hooks.
OpenCode
--opencodeSpecialists install as agents. Tool limits are enforced by permission blocks.
Antigravity
--agySpecialists ship as skills in the swarmkit plugin. Antigravity has no custom subagents, so limits are stated in each skill.
Installer flags 9
| Flag | What it does |
|---|---|
--opencode | Install OpenCode config |
--agy | Install Antigravity (agy) Swarm config |
--claude | Install Claude Code Swarm config |
--n8n | Configure local self-hosted n8n credentials (optional add-on, not in --all) |
--cloudflare | Install Cloudflare skills and configure auth (optional add-on, not in --all) |
--all | Install all agent configs (opencode, agy, claude) |
--free | Enable free mode for OpenCode (uses default models, no keys) |
--uninstall | Uninstall all configurations |
--help | Show this help message |
02 / How it routes
How it routes
A request is classified into a tier, explored if needed, executed, then checked.
Routing tiers
| Tier | Kind of task | What happens |
|---|---|---|
T0 | Questions and reviews | Answered directly. |
T1 | Trivial edits | Typos, renames, version bumps. Done directly, with no plan and no gate. |
T2 | Contained domain work | Plan, your approval, then execution by the domain specialist. |
T3 | Cross-cutting work | Same as T2, in an isolated git worktree. |
T0 to T3 are routing tiers (how much ceremony a task gets). fast, standard and deep are model tiers (which model class runs an agent).
Example routes
Illustrative: the actual route depends on what the change touches.
Fix a typo
T1- Edit directly
- No plan, no gate
Add an API route
T2- Plan and approval
- Implement
backend-specialist - If input handling changed
security-auditor - If tests were not run
release-tester
Refactor auth
T3- Isolated worktree
- If the codebase is unfamiliar
explore - Plan and approval
- Implement
backend-specialist - If auth changed
security-auditor - If the diff is large
code-proofreader
Quality gates 3
Run after the work, in parallel where the CLI allows, and only when relevant. Worktrees are never merged or removed without an explicit instruction.
| Agent | Runs when |
|---|---|
security-auditor | auth, secrets or input handling changed |
code-proofreader | the diff is large |
release-tester | tests were not run |
03 / Roster
Roster
33 agents, grouped by the job they do.
Core 1
Plans and dispatches. Does not touch files.
lead-devdeepPrimary orchestrator
Context and fast work 3
Read-only context, git, mechanical edits.
explorefastRead-only context-gathering pre-flight for the lead-dev swarm
git-specialistfastGit workflow specialist — commit/branch review (default) AND...
junior-devfastJunior dev for light, mechanical code edits that don't need a domain...
Domain specialists 19
Where T2 and T3 work lands.
android-capacitor-specialiststandardAndroid specialist for Capacitor apps with React + Vite
animation-specialiststandardAnimation, 2D, and 3D specialist for web — Motion, GSAP, Anime.js...
backend-specialiststandardBackend specialist focused on API design, service boundaries...
blender-specialiststandard3D modeling, mesh generation, asset staging, geometry nodes, spatial...
db-specialistdeepDatabase specialist for schema design, migrations, query...
devops-specialiststandardDevOps specialist for CI/CD pipelines, infrastructure as code...
docker-specialiststandardDocker specialist for containerization, Dockerfiles, Compose stacks...
electron-specialiststandardElectron specialist for wrapping existing React + Vite web apps as...
frontend-specialiststandardFrontend specialist focused on production-ready UI implementation...
ios-capacitor-specialiststandardiOS specialist for Capacitor apps with React + Vite
linkedin-specialiststandardLinkedIn content specialist
lovable-specialiststandardFrontend specialist for Lovable-made projects
monitoring-specialiststandardMonitoring and observability specialist for Prometheus, Grafana...
n8n-debuggerstandardSystematic debugging and diagnosis of broken n8n workflows
n8n-workflow-builderstandardBuild and design n8n workflows from requirements
seo-specialiststandardSEO specialist — makes sure websites actually get seen by Google and...
seo-workerstandardPost-build SEO & sharing pass worker
server-specialiststandardUbuntu server administration expert for system configuration...
swarm-architectdeepSwarm architect specialist for designing, extending, and refactoring...
Quality gates 4
Review, verify and test.
code-proofreaderstandardCode proofreader that finds dead code, redundant logic, unused...
release-testerfastFinal quality gate that runs tests, linters, type checkers, and...
security-auditordeepSecurity reviewer that scans code for secrets leakage, hardcoded API...
test-writerstandardWrites unit and integration tests for new code
Showroom 6
A coordinator and its workers for scroll-driven product pages.
showroom-art-directorstandardShowroom Art Director
showroom-asset-processorfastShowroom Asset Processor Worker
showroom-frontend-builderstandardShowroom Frontend Builder Worker
showroom-intakestandardShowroom Intake Worker
showroom-motion-engineerstandardShowroom Motion Engineer Worker
showroomstandardShowroom Coordinator
04 / Extend
Extend
MCP servers, skills and slash commands that ship with the kit.
MCP servers 11
Registered for Claude Code and Antigravity from core/mcp.json.
- blender
- cloudflare
- cloudflare-bindings
- cloudflare-builds
- cloudflare-docs
- cloudflare-observability
- firecrawl
- gemini-mcp-tool
- google-search-console
- google-trends
- playwright
Skills 27
Shared by all three CLIs, from core/skills/.
- backend-quality
- capacitor-mobile-quality
- caveman
- caveman-commit
- caveman-compress
- caveman-help
- caveman-review
- curated-resources
- frontend-quality
- git-workflow
- img2threejs
- n8n-api
- n8n-debugging
- ponytail
- ponytail-audit
- ponytail-debt
- ponytail-help
- ponytail-review
- premium-frontend-system
- release-testing
- scroll-craft
- security-review
- seo-engineering
- seo-sharing-pass
- showroom
- swarm-handoff
- web-design-guidelines
Slash commands 5
OpenCode.
- /ponytail-auditAudit the whole repo for over-engineering, what can be deleted
- /ponytail-debtHarvest ponytail: comments into a tracked debt ledger
- /ponytail-helpQuick reference for ponytail levels, skills, and commands
- /ponytail-reviewReview changes for over-engineering, what can be deleted
- /ponytailSwitch ponytail intensity level (lite/full/ultra/off)
05 / Proof
Proof
What is measured, what is not yet, and how this site was made.
Ponytail diff
net: pending
Ponytail is the always-on discipline against over-engineering. No before-and-after diff has been recorded yet, so no figure is shown.
Benchmarks
benchmarks pending
No benchmark data has been recorded yet. When it exists it will show medians over at least three runs per fixture per agent type, with the raw file linked. Losses will be shown too.
Harness disclosure
Benchmark runs: Antigravity (agy) with gemini-3.8-flash-high. This site: built with Claude Code, using the agents described above. The scene is procedural: no external models, textures or volume data, and no AI-generated media.
06 / Build log
Build log
Who did what, read from the repo at build time. It separates what the human did from what agents did.
- Human interventions
- 30
- 4 gate approvals, 18 decisions and answers, 8 art-direction rounds with a human verdict. Measurements and agent-applied changes are not counted.
- Agent dispatches
- 4
- animation-specialist 2, frontend-specialist 1, showroom 1
Build timeline 44 events
Phase 0 benchmark deferred: user will supply benchmark file later
Dev server binds to 0.0.0.0 (global rule overrides PLAN.md 127.0.0.1)
showroom finished a task
BRIEF.md approved; five-chapter structure per PLAN section 6 confirmed
Phase 2 plan approved: worktree site/previs, animation-specialist builds grey-box previs; go/no-go after
animation-specialist finished a task
Round 1: 1 change applied by the agent
Phase 2 frame rate reported by the human
grey-box previs approved; fps within budget on laptop and phone
Phase 3 plan approved: ink cool blue-black; Space Grotesk + JetBrains Mono; accent = 3 candidates in round 1
animation-specialist finished a task
Round 2: 1 verdict from the human
Round 2: 6 changes applied by the agent
Human: focus on desktop, skip phone fps checks for now. Phone budget and medium tier are deferred, not dropped from PLAN.md
human: it all looks good; all elements and tokens locked (violet accent, cool ink, Space Grotesk + JetBrains Mono). Desktop fps sweep not reported; phone/medium tier deferred
Round 3: 6 verdicts from the human
Phase 4 plan approved: ponytail diff + benchmarks pending; repo github.com/ha-lun/swarmkit; horology placeholder; responsive at 1440/1024/768/390 layout only
frontend-specialist finished a task
human: the page looks good; continue. Review items 1-5 (interventions count definition, parallel-gates qualifier, terse roster blurbs, tier legend, benchmark schema) not yet answered
Phase 5 plan approved; interventions counter counts only the human own actions
Human feedback after Phase 5: animation should be the centre, small captions only. Plan 5b approved: detail text moves into the world (cell label cards) plus a Reference section; pinned scenes; slim Proof HUD; finale is the climax
Human: animation-first build is better; wants the honeycomb as a sphere/globe. Plan 5c approved: Goldberg sphere, scroll orbits camera, satellites as orbiting moon cluster; new art-direction round 4 to follow
Human: yes to round-4 tuning (finer globe 252 cells, lower columns, softer rim, pulled-back fly-over). Material = pearlescent ceramic; surroundings = atmosphere halo + sparse dust + moving studio light; text = captions 16px, large wordmark/tagline at start. Look values unlocked for round 4 only
Round 4: 1 verdict from the human
Round 4: 1 change applied by the agent
Human round-4 feedback: dislikes the glow (all of it removed), comet animation not up to par, transitions not smooth, globe has no texture and panels misaligned. Chose speckled stone surface and remove all glow. Round 5 plan approved
Human after round 5: still some work to do but very good. Asked to commit everything and resume in a new session. Round-5 look NOT locked; site/integration NOT merged
Round 5: 2 verdicts from the human
Round 6 plan approved: elevation terrain + rock shader, polished cut facets replace rings, seamless cells-proof dissolve (live-pose outgoing frame, post stays on, one ease), scene text fixed and fade-only. Look values remain unlocked
Round 6: 1 verdict from the human
Round 6: 1 change applied by the agent
Round 7 plan approved: basalt columns (quantised stepped heights, flat smooth tops, dark sides), palette tokens UNLOCKED for this round (graphite, iron-blue, bone candidates via ?palette=), accent stays violet. Human asked whether an AI-generated image would improve the texture, then said go ahead with the plan as written: no AI media used
Round 7: 1 verdict from the human
Round 7: 1 change applied by the agent
Palette LOCKED to iron-blue (ink #080b10, ink-2 #121824, wax #9aa6b4, wax-dim #46505e, text #d9dee5; accent stays violet). Round 8 plan approved: piston columns, slow spin parked during hive and fly-over, agents steady
Round 8: 1 verdict from the human
Round 8: 1 change applied by the agent
Round 9 plan approved: agents as fixed towers above the moving rods with polished caps; richer top and wall shading and lighting; key-light cast shadows on high tier only (medium gets AO stand-in). Palette stays locked (iron-blue)
Round 9: 1 verdict from the human
Round 9: 1 change applied by the agent
Round 10 plan approved: moon spin, drift and light piston pulse parked home for hive and fly-over; Reference redesigned as index rail plus six focused sections (Install, How it routes, Roster, Extend, Proof, Build log), honest pending states kept
Round 10: 1 change applied by the agent
Human ended the session and will continue tomorrow. Round 10 (moon motion, Reference rebuild) built, no verdict yet; look values still unlocked; site/integration NOT merged; HANDOFF.md updated
Round 11: 1 change applied by the agent
Horology demo: link coming.Source: github.com/ha-lun/swarmkit