A web dashboard that orchestrates CLI-based AI coding agents. You bring your own tools and infrastructure — VibeSwarm turns ideas into code.
VibeSwarm is an agentic CI/CD system for turning ideas into application code. It provides a responsive web UI for managing projects and orchestrating AI coding agents running on your host machine. Think of it as a self-hosted control panel for your AI workforce.
- Self-hosted — runs on a Raspberry Pi, VPS, laptop, or cloud instance.
- Bring your own agents — install the CLI tools you already use and VibeSwarm detects them automatically.
- No API keys stored — authentication lives in your host environment; VibeSwarm just calls CLI commands.
| Agent | CLI Command | Verified against | Minimum | Install |
|---|---|---|---|---|
| Claude Code | claude |
2.1.322 | 2.0.0 | claude.ai/code |
| OpenCode | opencode |
1.18.31 | 1.0.0 | opencode.ai |
| GitHub Copilot | copilot |
1.0.86 | 1.0.0 | copilot CLI |
Agents are auto-detected at startup. If a supported CLI tool is on your PATH, VibeSwarm registers it as a provider automatically.
These CLIs change fast. ProviderVersionReference (in src/VibeSwarm.Shared/Providers/) is the
single source of truth for the release each integration was built and verified against — last
reviewed 2026-11-11. Newer CLI features are version-gated, so an older CLI keeps working
with those features switched off rather than failing.
At startup VibeSwarm compares each detected CLI against this reference and logs the result: a warning when the CLI is below the supported minimum, and an informational line when it is older or newer than the verified release. When re-verifying against a new CLI, update that file so the gap between "what we tested" and "what is installed" stays visible.
Claude Code and GitHub Copilot meter usage against a subscription; VibeSwarm tracks both and pauses work when a window is exhausted.
- Claude Code reports limits as structured
rate_limit_eventmessages during a normal headless run. VibeSwarm reads the rolling 5-hour (shown as Session) and 7-day (shown as Weekly) windows, plus whichever limit is currently binding. Warnings printed to stderr are parsed as a fallback. - GitHub Copilot meters premium requests and, increasingly, GitHub AI Credits.
VibeSwarm reads the structured usage report written by
--usage-output-filewhere the CLI supports it, and otherwise falls back to parsing the session summary on stderr. - OpenCode has no quota of its own — metering belongs to whichever model provider you configure. Self-hosted open-source models (Ollama, LM Studio, llama.cpp, vLLM and similar) are reported as Unmetered: no meter is drawn, and they are never paused for exhaustion, since there is no window to wait for.
Claude Code providers offer the CLI's aliases (sonnet, opus, haiku, fable), which always
run the newest model of each line, followed by pinned model IDs. The list ships with VibeSwarm and
is refreshed shortly after each start, so new models appear without a manual refresh.
A Claude Code provider can also name an Advisor (--advisor): a stronger model the agent
consults at hard moments, billed at that model's rate. It applies to every run of the provider
(jobs, planning, review and ideas). Running them on Sonnet with Opus or Fable as advisor keeps
most of the work on the cheaper model. Claude Code skips an advisor weaker than the run's model,
and skips Fable until it has been enabled once in an interactive session. Leave it unset to use
Claude Code's own advisorModel setting, or pick Off to turn it off for that provider.
Each agent keeps its own memory in its own place, so VibeSwarm keeps one project memory
that every agent shares. Edit it under Instructions and memory in a project's settings
(up to 20,000 characters). Before each job VibeSwarm writes it to
.vibeswarm/project-memory.md in the working directory, tells the agent to read it and add
lasting lessons (not progress, plans or todo lists), and saves the agent's edits back to the
project when the job ends.
Session files never reach your commits. At every job start VibeSwarm lists the agents' own
folders and files (.vibeswarm/, .claude/, .opencode/, .copilot/, CLAUDE.local.md,
tasks/todo.md), Playwright and test output, logs and temp files in the repository's
.git/info/exclude. That file is local to your clone, so your .gitignore is never changed,
and a run that only leaves such files behind doesn't count as a code change. Agents are also
told to keep plans, todo lists and scratch work out of the repository.
Agents can open, click through and screenshot the web apps they build. Every job gets the
Playwright MCP server, driving a headless
Chromium on the host, whether Claude Code, Copilot or OpenCode runs it. Set it up under
Settings > General > Browser for agents: it shows whether Node.js and Chromium are
installed, and Install Playwright and Chromium installs whatever is missing. Chromium goes
into Playwright's shared cache (~/.cache/ms-playwright on Linux), with its system libraries
when VibeSwarm can use sudo without a password. A system Chromium (such as Debian's
chromium package on a Raspberry Pi) works too. Turn off Give every job a browser to keep
it to projects with a web environment.
To batch changes and ship them when you choose, set a Development branch and a Release branch under When a job finishes in a project's settings. Jobs then work on the development branch, and the project gets a Release tab listing the changes that are waiting. Deploy merges the development branch, as it is on origin, into the release branch and pushes it, so your CI/CD can ship it from there.
If the release branch has changes of its own that conflict, such as a hotfix, nothing is pushed. VibeSwarm queues a job instead: it merges the release branch into the development branch and leaves the conflicts for a provider to resolve, and the result is committed as a merge only once no conflict markers remain. When that job finishes, press Deploy again. Deploys are always manual.
Two optional helpers sit above single jobs. Both are set per project in its settings, and both run read-only: Claude Code or Copilot with their write tools blocked.
- Review (off by default) checks each job before its work is delivered. When the agent finishes, the project's build and tests run. Then a second model reads the change against the goal and the plan's acceptance criteria. The reviewer is a Claude or Copilot CLI that can open files, or an inference provider that sees only the diff. Problems go back to the agent in the same session, at most twice. Review never holds work back: if the reviewer is off, near its usage limit, fails or gives no verdict, the work is delivered and the job says why it wasn't reviewed.
- Lead suggests what to build next. A Lead review copies the branch jobs deliver to, as it is on origin, and reads it alongside the project memory, recent job outcomes, unresolved review findings and an optional Direction you write. It returns a short report and up to eight ideas, each with a goal and acceptance criteria. They wait under Suggested by Lead on the project page until you approve them. Dismissed ones are never suggested again.
On the Auto-Pilot tab, Ideas → Only ones I approve makes auto-pilot run only ideas you typed or approved. When none are left, it asks the Lead for more, tells you once, and waits.
Download from dot.net/download. Install the .NET SDK for your platform.
git clone https://github.com/NorthRiverDesign/VibeSwarm.git
cd VibeSwarmdotnet run --project src/VibeSwarm.WebThe default setup uses SQLite and the browser-based setup wizard, so you can start the app without creating a config file first.
cp src/VibeSwarm.Web/.env.example .envOnly copy .env when you want to customize ports, pre-create the first admin account, or switch databases.
The application will:
- Run database migrations (SQLite by default, zero configuration)
- Auto-detect installed CLI agents
- Start on http://localhost:5000
Navigate to http://localhost:5000
To also serve HTTPS, turn on Settings → Connection → Serve over HTTPS and restart VibeSwarm. It then generates a self-signed certificate and listens on https://localhost:5001 as well. Your browser will show a certificate warning until you trust it (devices can install it from /cert):
- Chrome/Edge: Click "Advanced" → "Proceed to localhost"
- Firefox: Click "Advanced" → "Accept the Risk and Continue"
Browsers only allow installing VibeSwarm as an app and keeping the screen awake over HTTPS (or on localhost).
On first launch you will be redirected to a setup wizard to create your admin account. For automated deployments, set DEFAULT_ADMIN_PASS and optionally DEFAULT_ADMIN_USER in your .env file before starting the app.
Prefer the portable release asset from GitHub Releases for production installs, or publish one locally:
dotnet publish src/VibeSwarm.Web/VibeSwarm.Web.csproj -c Release -o ./build/vibeswarm /p:UseAppHost=false
cp src/VibeSwarm.Web/.env.example ./build/vibeswarm/.envThen run the published app from ./build/vibeswarm with:
dotnet VibeSwarm.Web.dllThe .env file is the only configuration you need. Place it in the repo root or in src/VibeSwarm.Web/.
| Variable | Default | Description |
|---|---|---|
ASPNETCORE_URLS |
http://localhost:5000 |
Bind addresses. Use 0.0.0.0 for remote access. https:// ones need HTTPS on in Settings. |
DEFAULT_ADMIN_USER |
admin when only a password is provided |
Optional admin username for automated setup. |
DEFAULT_ADMIN_PASS |
(empty — setup wizard) | Admin password. Min 8 chars, upper + lower + digit. |
DATABASE_PROVIDER |
sqlite |
Database engine: sqlite or mysql. |
ConnectionStrings__Default |
Data Source=vibeswarm.db |
Connection string for the chosen provider. |
You can also set these as system environment variables instead of using .env.
SQLite (default) — zero configuration. The file vibeswarm.db is created next to the app.
Migrations run automatically on startup and build the schema from empty, so a new install works on either provider with no manual setup.
MySQL:
DATABASE_PROVIDER=mysql
ConnectionStrings__Default=Server=localhost;Database=vibeswarm;User=vibeswarm;Password=secretProvider aliases are supported: mariadb / mysql.
You can also switch databases from Settings → Database. That flow copies your current VibeSwarm data into an empty target database, writes the new provider and connection string to a runtime config file outside the repo, and takes effect after restart.
cp vibeswarm.db vibeswarm.db.backupEF Core bakes provider-specific SQL into a migration when it is generated — SQLite emits
TEXT columns where MySQL needs char(36) — so one migration set cannot serve both
providers. VibeSwarm therefore keeps a set per provider:
src/VibeSwarm.Web/Data/Migrations/
├── Sqlite/ # owned by SqliteVibeSwarmDbContext
└── MySql/ # owned by MySqlVibeSwarmDbContext
VibeSwarmDbContext itself owns no migrations. Everything that applies them goes through
DataServiceExtensions.CreateMigrationContext, which picks the set matching the configured
provider.
Every schema change must be generated for both providers:
dotnet ef migrations add <Name> --project src/VibeSwarm.Web \
--context SqliteVibeSwarmDbContext --output-dir Data/Migrations/Sqlite
dotnet ef migrations add <Name> --project src/VibeSwarm.Web \
--context MySqlVibeSwarmDbContext --output-dir Data/Migrations/MySqlMigrationSetsStayInStepTests fails the build if one set drifts from the other, so a
forgotten provider is caught by dotnet test rather than by a failed deployment.
On MySQL, string columns with a MaxLength of 1000 or more are mapped to LONGTEXT.
InnoDB caps a row at 65,535 bytes and utf8mb4 costs 4 bytes per character, so without this
the AppSettings, Jobs and CriticalErrorLogs tables exceed the limit and cannot be
created. MaxLength is still enforced by EF validation.
Create /etc/systemd/system/vibeswarm.service:
[Unit]
Description=VibeSwarm
After=network.target
[Service]
Type=notify
WorkingDirectory=/opt/vibeswarm
ExecStart=/opt/vibeswarm/VibeSwarm.Web
Restart=always
RestartSec=10
KillSignal=SIGINT
Environment=ASPNETCORE_URLS=http://0.0.0.0:5000;https://0.0.0.0:5001
Environment=ASPNETCORE_ENVIRONMENT=Production
[Install]
WantedBy=multi-user.targetsudo systemctl daemon-reload
sudo systemctl enable vibeswarm
sudo systemctl start vibeswarm# Download NSSM from nssm.cc
nssm install VibeSwarm "C:\path\to\publish\VibeSwarm.Web.exe"
nssm set VibeSwarm AppDirectory "C:\path\to\publish"
nssm start VibeSwarm- VibeSwarm serves plain HTTP by default. Turning on Settings → Connection → Serve over HTTPS adds HTTPS with a self-signed certificate. For production, place it behind a reverse proxy with a real certificate.
- The application calls CLI tools on the host system. It does not store API keys — your agents authenticate through their own configurations.
- All user passwords are hashed with ASP.NET Core Identity defaults.
Expected with self-signed certificates, which VibeSwarm only uses while Settings → Connection → Serve over HTTPS is on. Turn it off and restart to serve HTTP only, accept the warning in your browser, or put a real certificate in front with a reverse proxy.
# Windows
netstat -ano | findstr :5000
# Linux/macOS
lsof -i :5000Change the port in .env via ASPNETCORE_URLS.
- Verify the CLI tool is on your PATH:
claude --version,opencode --version,copilot --version - Check the application logs for detection results, including how each CLI's version compares to the release VibeSwarm was verified against.
- You can always add agents manually through the web UI under Providers.
- .NET 10 SDK
- Windows, Linux, or macOS
- Git
dotnet build VibeSwarm.slncd src/VibeSwarm.Web
dotnet rundotnet testdotnet publish src/VibeSwarm.Web/VibeSwarm.Web.csproj -c Release -o ./build/VibeSwarm/
├── src/
│ ├── VibeSwarm.Client/ # Blazor WebAssembly front-end
│ ├── VibeSwarm.Shared/ # Shared models, services, utilities
│ └── VibeSwarm.Web/ # Server — API, SignalR, Identity, CLI orchestration
│ └── .env.example # Configuration template
├── build/ # Published output
└── README.md
- .NET 10.0 - Web framework
- Blazor WebAssembly - UI
- SignalR - Real-time communication
- Entity Framework Core - ORM (SQLite, MySQL)
- Provider CLIs - Claude Code 2.1.276, OpenCode 1.18.31, GitHub Copilot 1.0.86
- ASP.NET Core Identity - Authentication
We are working on an update notification system and providing pre-built solutions. Things change fast. For now, we recommend checking out the main branch and running dotnet publish -c Release -o ./build/. If you have VibeSwarm installed as a service, restart the process.
VibeSwarm is open sourced under the MIT license and developed by the company, North River Design LLC. The company retains all rights to claim intellectual property of VibeSwarm.
Because this whole application is vibe coded, feel free to fork it. We just need to absolutely specify we are NOT responsible for ANY claims of damage.
If you would like to contribute to VibeSwarm, open a GitHub issue with any problems you are facing. We are an extremely small team (1 person) and rapidly developing this application in a changing landscape.
Any attempt at submitting contributions with nefarious intent will be denied. Be a good person.
This application is being developed to serve our needs and is distributed for free. Any negativity will not be tolerated or obliged. This project is open source and you are not entitled to any support or reimbursement for any damages.
VibeSwarm is provided with absolutely no warranty. Use at your own risk. The maintainers are not responsible for any damages or losses resulting from the use of VibeSwarm. Users are encouraged to thoroughly test the application in their own environments before deploying it in production settings.
VibeSwarm is for expert developers who know the limitations and risks of using AI coding agents. It is the user's responsibility to ensure that the generated code meets their quality, security, and compliance standards.
Any and all use of VibeSwarm must comply with the terms of service and usage policies of the AI providers whose agents are utilized. Users are responsible for understanding and adhering to these policies to avoid violations that could lead to account suspension or other penalties.
By using VibeSwarm, users agree to indemnify and hold harmless the maintainers from any claims, damages, or liabilities arising from their use of the application. In short, we are NOT responsible if the AI agent decides to delete important information on your system — it's just a wrapper application.
This application is free for anyone to use. If you like it and want to donate, we have a donation link set up.
Donate to North River Design with Stripe.
Thank you!
