Whiskers

[Whiskers logo]

Whiskers

Controlled infrastructure operations for humans and AI agents.

Whiskers is a self-hosted control plane for Docker hosts and Kubernetes workloads. Instead of handing operators or AI agents unrestricted SSH access, it exposes explicit tools governed by permissions, code-enforced guardrails, approvals and a complete action history.

Bootstrap a host once, then operate over mesh + mTLS without a standing private SSH key.

[Release] [Status] [License: Apache 2.0] [.NET] [Blazor] [MCP Badge]

⚠️ Beta. Whiskers is under active development and not yet API-stable. Run it on a trusted network, review the security policy, and expect breaking changes before 1.0.

Why Whiskers

AI agents can operate infrastructure — but the usual way to let them is to hand over an SSH key or a shell, which grants far more power than any single task needs. Whiskers puts a governed layer in between:

• Limit what an operator or agent can do. API keys, roles and explicit tool access define the available capability set — no arbitrary shell.
• Require approval before impact. Code-enforced guardrails allow, block, or pause sensitive operations for human confirmation.
• Know exactly what happened. Agent History and the Audit Log record actor, decision, target, parameters (secrets redacted) and result.

The same control plane reaches Docker, Kubernetes, logs, deployments, services and CVE context — but that reach is evidence of coverage, not the headline. The headline is governed, accountable operations. See docs/product/POSITIONING.md for the full positioning.

How governance works

flowchart LR    H[Human operator]    A[AI agent]    CP[Whiskers control plane]    AUTH[Authentication & permissions]    GR[Code-enforced guardrails]    AP[Human approval]    TOOLS[Explicit infrastructure tools]    INFRA[Docker / Kubernetes / hosts]    AUDIT[Agent history & audit log]
    H --> CP    A --> CP    CP --> AUTH    AUTH --> GR    GR -->|allow| TOOLS    GR -->|confirm| AP    AP -->|approved| TOOLS    GR -->|block| AUDIT    TOOLS --> INFRA    TOOLS --> AUDIT    AP --> AUDIT

Table of contents

• Quick start
• The hero workflow
• Capabilities
• Architecture & security
• MCP server
• AI agent
• Configuration
• Screenshots
• Roadmap & beta
• Project structure
• Development
• Contributing
• License

Quick start

Prebuilt multi-arch images (amd64 / arm64) are published to ghcr.io/lupusmalusdeviant/whiskers from v0.12.0. Before that release, use the From source path below.

Quick — one command

# Installer: pulls the image, writes a small compose file in ./whiskers/, waits until healthy, prints the URL.curl -fsSL https://raw.githubusercontent.com/LupusMalusDeviant/Whiskers/main/deploy/install.sh | bash
# ...or a bare docker run:docker run -d --name whiskers -p 127.0.0.1:5100:8080 \  -v whiskers-data:/app/data -v /var/run/docker.sock:/var/run/docker.sock \  ghcr.io/lupusmalusdeviant/whiskers:latest

Open http://127.0.0.1:5100. The installer accepts --port, --bind, --data, --yes (install.sh --help); re-running it updates in place (pull + up).

Standard — Docker Compose

Download docker-compose.yml from the latest release (already pinned to the image), then docker compose up -d. Optional settings go in a .env beside it — everything is optional, Whiskers runs with none (see .env.example).

The app listens on 127.0.0.1:5100 by default (HOST_BIND / HOST_PORT) and never binds publicly unless you change the bind.

First sign-in

On first launch a guided setup wizard in the browser creates the admin account — no .env and no config file required. Because whoever completes the wizard becomes Admin, it first asks for a one-time setup token that only someone with access to the server can read: docker exec whiskers cat /app/data/setup-token.txt (the installer prints it; Helm users kubectl exec the same path, or preset it with WHISKERS_SETUP_TOKEN). From then on you sign in with that local login (email + password; 5 wrong passwords lock the account for 15 minutes, and login attempts are rate-limited per client). Alternatives: seed the admin unattended with WHISKERS_ADMIN_EMAIL + WHISKERS_ADMIN_PASSWORD_FILE; use Google or generic OIDC (see Configuration); or, for a trial on a trusted network, AUTH_DISABLED=true — every request is then an Admin, so Whiskers only answers clients from private networks, with an IP or localhost as host name and without a foreign Origin (see Configuration).

Connect an MCP client

Whiskers is self-hosted: there is no hosted or cloud endpoint. Your MCP client talks to your own instance at /mcp (Streamable HTTP) with an API key created under Settings → MCP:

{  "mcpServers": {    "whiskers": {      "type": "http",      "url": "http://localhost:5100/mcp",      "headers": { "Authorization": "Bearer <WHISKERS_MCP_API_KEY>" }    }  }}

Details, scopes and the guardrail behaviour are in MCP server.

More install options

location /whiskers/ {    proxy_pass http://127.0.0.1:5100/whiskers/;    proxy_http_version 1.1;    proxy_set_header Upgrade $http_upgrade;    proxy_set_header Connection $connection_upgrade;    proxy_set_header Host $host;    proxy_set_header X-Real-IP $remote_addr;    proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;    proxy_set_header X-Forwarded-Proto $scheme;    proxy_set_header X-Forwarded-Prefix /whiskers;    proxy_read_timeout 300s;    proxy_buffer_size 128k;    proxy_buffers 4 256k;}

helm install whiskers oci://ghcr.io/lupusmalusdeviant/charts/whiskers \  --set vault.key="$(openssl rand -hex 32)"kubectl port-forward svc/whiskers 8080:8080   # → http://localhost:8080 (setup wizard)

git clone https://github.com/LupusMalusDeviant/Whiskers.gitcd Whiskersdocker compose up -d          # builds the image locally

The hero workflow

The flow that defines the product — Observe → Propose → Check policy → Request approval → Execute → Verify → Audit:

1. A container / workload is unhealthy.
2. An agent analyses status and logs using a read-scoped key — no write access needed.
3. The agent proposes a restart.
4. The active guardrail rates the restart tool as Confirm, so nothing runs yet.
5. Whiskers raises an approval request: actor, tool, target server, target workload, redacted parameters, rationale and expiry.
6. A human approves it.
7. Whiskers executes only the approved action, with the approved parameters.
8. Whiskers verifies the workload afterwards.
9. Agent History and the Audit Log show the full chain, and the UI deep-links between result, approval and history.

This is the reference workflow for AI-operations onboarding and the acceptance test for governance.

Capabilities

Docker, Kubernetes and the rest are targets the control plane governs — grouped by what they do for you:

Governance

• Per-key permissions — MCP API keys are scoped Read / Write / Admin, or to an explicit tool list.
• Code-enforced guardrails — per-tool Allow / Confirm / Block, evaluated at the tool boundary (not in the prompt), "most-restrictive wins".
• Human-in-the-loop approvals — a Confirm tool pauses execution until an authorized user approves the exact action + parameters.
• Agent History — every agent tool call recorded: actor/key, redacted parameters, decision, result, duration, target.
• Audit Log — every user action recorded, with a filterable history.

Operations

• Containers — live dashboard (CPU / memory / health) across all hosts; start / stop / restart / remove; logs, stats, health reports; grouping by server, Compose project and free-form tags.
• Kubernetes — add a k3s/Kubernetes cluster via kubeconfig (stored encrypted in the vault); pods appear next to containers with owner grouping, logs and honest scale/rollout actions.
• Host management — firewall (ufw), Nginx sites, systemd services, SSL certificates and an integrated web terminal (host and container).
• Deployments — a deploy form, Compose upload, curated app templates, image search across registries, and git deploy (repo → cloned/built/composed on the target host; push-triggered redeploys via HMAC-signed webhooks).

Security

• CVE scanning — OS and container images (Trivy) with a deduplicated findings dashboard, severity and age.
• Secrets vault — stored secrets encrypted at rest (AES-256-GCM); sensitive values redacted before logging and persistence.
• Mandatory HMAC webhook secrets (X-Hub-Signature-256-compatible) and fail-closed authorization — every endpoint requires authentication unless it explicitly opts out.
• SSH host-key verification (trust-on-first-use) on every ssh path.
• Mesh + mTLS steady-state operation (see Architecture & security).

Resilience

• Self-backup & restore of Whiskers' own data — optionally VAULT_KEY-encrypted (AES-256-GCM), schedulable, crash-safe deferred-swap restore.
• Image-update rollback — a snapshot before each update; one-click revert on the dashboard.
• Monitoring & alerting — historical CPU/RAM/disk metrics, health reports, and notifications (Mattermost, Matrix, Telegram, ntfy, Discord, Slack, Email/SMTP, generic webhook) on unhealthy/stopped/OOM containers, restart loops, sustained-load anomalies, new CVEs and available image updates, with a persistent in-app feed.

Integrations

• MCP server — the authenticated interface for external AI clients (see MCP server).
• Cloud control (out-of-band) — provider-agnostic power / snapshot / metrics (Hetzner, Hostinger) via the provider API, even when SSH/Docker is temporarily unreachable.
• Registries — private-registry credentials in the vault, used automatically for authenticated pulls.
• Database — SQLite (zero-config) or PostgreSQL via Entity Framework Core.
• UI — light / dark / system theme and a fully localized English (default) / German interface.

Architecture & security

Whiskers manages each onboarded host across three separate planes. None of them uses SSH in steady state, and none of them needs a stored SSH key:

SSH is a bootstrap channel only. Onboarding opens a single SSH connection — with host-key verification (trust-on-first-use, pinned afterwards) — to join the host to a private WireGuard mesh (Tailscale/NetBird), roll out the mTLS certificates and the Docker proxy, and install telemetry. The bootstrap credential (uploaded key or a password held in memory only) is then deleted. From that point on the host is operated key-free over the three mTLS/mesh planes; Whiskers holds no standing SSH key for it. (A server that was added as a plain SSH connection and never onboarded keeps using SSH until it is migrated.)

What these planes are, honestly: the Docker-API and host-control planes are root-equivalent on the managed host — creating a container or running nsenter is root by construction. mTLS and the mesh decide who can reach that power; scopes, guardrails and approvals decide what a given caller may do with it. The full design, PKI (step-ca), onboarding flow and trade-offs are in docs/ARCHITECTURE.md; the threat model is in SECURITY.md.

Security summary

• No standing SSH key for managed hosts in steady state; management ports are mesh-bound, nothing management-related is exposed to the public internet by design.
• Deterministic policy enforcement, not prompt instructions. Every MCP tool call — from a signed-in web user, an external MCP client or the in-process agent — passes the same scope check in code (Read / Write / Admin, or an explicit per-key tool list). Every one of them then also passes the same guardrail engine: action classification (Allow / Confirm / Block, protected resources, forbidden arguments, read-only kill switch) and human-in-the-loop approval for Confirm. See How the policy layer applies.
• MCP API keys are created in the web UI with an explicit scope; the ServerWatch-era unscoped key file api-keys.json no longer exists (see the MCP server section).
• Secrets (.env, the encrypted vault, the DB) live under /app/data / .env, both gitignored and volume-mounted — never baked into the image or committed.
• Full-repo security review & remediation (see the security-fixes changelog): all critical and high findings, plus most medium/low issues, are fixed — per-key RBAC on every path (including the agent and MCP), the vault on AES-256-GCM + PBKDF2, schema via EF Core migrations, trust-critical images pinned by digest, restricted forwarded-header trust, and a scrape-token-gated /metrics.

Supply chain

Every published release image is, from v0.12.1 on:

• scanned before it ships — the release pipeline builds the image, runs a Trivy scan, and fails the whole run on any CRITICAL so a vulnerable image is never pushed;
• built multi-arch with provenance + an SBOM attestation (SLSA provenance and an SPDX SBOM attached to the image, and a standalone SBOM as a release asset);
• keyless-signed with cosign (Sigstore) — the signature is bound to the release workflow's GitHub OIDC identity and recorded in the public Rekor transparency log; no signing key is held by anyone.

Verify the signature before you run it:

cosign verify ghcr.io/lupusmalusdeviant/whiskers:0.15.0 \  --certificate-identity-regexp '^https://github.com/LupusMalusDeviant/Whiskers/.github/workflows/release.yml@refs/tags/v' \  --certificate-oidc-issuer https://token.actions.githubusercontent.com

A valid result means the image was built by this repository's release workflow from a tag — not rebuilt or tampered with. (Images published before v0.12.1 are provenance-attested but not cosign-signed.)

If you discover a security issue, please report it privately — see SECURITY.md.

MCP server

Whiskers ships an integrated MCP server (Streamable HTTP, endpoint /mcp) so AI agents (e.g. Claude Code) can operate your infrastructure. Add it to your MCP client:

{  "mcpServers": {    "whiskers": {      "type": "http",      "url": "https://whiskers.example.com/mcp",      "headers": { "Authorization": "Bearer <WHISKERS_MCP_API_KEY>" }    }  }}

Behind a reverse proxy on a subpath (PATH_BASE=/whiskers) the URL becomes https://example.com/whiskers/mcp; for a local instance it is http://localhost:5100/mcp.

API keys and scopes

• Keys are created in the web UI under Settings → MCP by an Admin. Each key carries exactly one scope — Read, Write or Admin — or, instead of a scope, an explicit list of allowed tools. Keys can be disabled or deleted there at any time; a disabled key is rejected on the next request.
• Every tool declares its minimum scope in code and the check runs on every call. A scoped key never reaches a tool above its scope; a tool without a configured level is denied (fail-closed); an Admin can switch individual tools off for every key.
• Keys are stored only as hashes. A key is 256 bits from a cryptographic RNG; Whiskers shows it once, right after creation, and persists only its SHA-256 hash plus the last four characters for recognition. A presented key is hashed and compared in constant time. Neither the key nor any part of a rejected token is written to the log. Someone who reads the data directory or a backup gets hashes, which do not authenticate. Requests without a valid key are rejected before any tool runs.

Note for older index entries (e.g. MCP directories that crawled the ServerWatch-era README): Whiskers was called ServerWatch before the rebrand in v0.12. The old flat, unscoped key file /app/data/api-keys.json is gone: its last compatibility fallback was removed in v0.15.0, no code path reads it, and a leftover copy is deleted on start. MCP access works only with a scoped key created in the web UI, and those keys are stored as SHA-256 hashes, never in plaintext.

Tools span:

• Containers: list/inspect/logs/metrics/env, start/stop/restart/update
• Server & host: info, logs, metrics, health summary, execute_command (Admin)
• Deployment: deploy_app, deploy_compose
• Infrastructure: firewall, Nginx, systemd, SSL
• Databases: detect, list, schema, query, backup
• Networks: list/create/remove, connect/disconnect containers
• Logs & alerts: search, list/create log alerts
• Scheduler: list/create/delete/run scheduled tasks
• CVEs & updates: server/container CVE summaries, update status
• Cloud (out-of-band): Hetzner & Hostinger power/snapshot/metrics
• Agent: instruct_agent (delegate a natural-language task to the in-process agent)

The complete, current list with permission levels is in docs/mcp-tool-catalog.md and in the web UI under Settings > MCP.

Prompts (user-invokable workflows; they only return instructions, every tool call they lead to is still scope- and guardrail-checked):

Resources (read-only context about the MCP surface itself, no infrastructure data):

How the policy layer applies

Whiskers' AI guardrails are a deterministic policy-enforcement layer (Read / Write / Admin scopes, action classification and human-in-the-loop approvals), evaluated in code at the tool boundary — never in a prompt, so no model output can talk its way past it.

Approvals are resolved on the Approvals page; approving requires at least the level the action needs. The guardrail engine is part of the agent module (on by default). With that module switched off there is no Approvals page either, and direct calls are bounded by their scope alone.

AI agent (optional)

Beyond the MCP server (which serves external agents), Whiskers has an in-process acting agent: you describe an operations task in natural language and it plans and executes using Whiskers's own tools. It supports multiple LLM providers (OpenAI, OpenRouter, Ollama, Gemini, Anthropic, and Claude Code) selectable in the UI. An optional read-only advisor chat is also available — it explains and suggests but runs nothing.

Its safety model is enforced in code, not in the prompt:

• Guardrails (a separate, admin-only guardrails.json) are a deterministic policy-enforcement layer: Allow/Confirm/Deny rules evaluated at the tool boundary, "most-restrictive wins", identical whether the agent was started from the web UI or via MCP.
• The agent inherits the rights of whoever triggered it (web user or MCP key) and can never exceed them.
• Hybrid autonomy: reads run autonomously; writes/admin actions require confirmation.

See src/Whiskers/Services/Agent/ for the implementation.

Configuration

Whiskers is configured entirely through environment variables (.env); everything is optional. The most important groups:

See .env.example for the full, commented list.

Notable runtime details

• Email whitelist: managed in the UI under Settings > Authentication; changes apply without a restart.
• MCP API keys: created in the UI under Settings > MCP (scope Read / Write / Admin per key). A fresh install has none. The key is shown once at creation and stored only as a SHA-256 hash — copy it then. Keys from older versions are converted to hashes on the first start and keep working. Grant Write/Admin only to keys that need it.
• Data persistence: the metrics database (SQLite by default, or PostgreSQL), JSON stores and certificates live under the data directory (WHISKERS_DATA_DIR, default /app/data); never in the image.
• Health probes: anonymous /healthz (liveness) and /readyz (readiness) endpoints; the image ships a Docker HEALTHCHECK and Kubernetes can use them as probes. (/health without the z is the in-app status page.)

Tech stack

Screenshots

The screens below trace the governance model end to end — controlled access, enforced governance, complete accountability. A live walkthrough of the hero workflow (unhealthy workload → agent analysis → guardrail decision → approval → execution → audit trail) lives on the project website: whiskers.app.lupusmalus.dev, and the in-app handbook (Help) carries screenshots for every area.

Guardrails — enforced governance. A code-enforced, non-bypassable policy checked at the tool boundary (not in the prompt): a read-only kill switch, the highest level allowed without confirmation, protected resources, and forbidden argument patterns.

[Guardrails: a code-enforced policy with a read-only kill switch, maximum autonomous level, protected resources and forbidden argument patterns]

Agent History — controlled access. Every agent/MCP tool call, with actor, privilege level (read / write / admin), the guardrail verdict (Allow / Confirm / Deny), duration and outcome.

[Agent History: a table of tool calls showing actor, level, verdict and status, including an admin-level call denied by policy]

Audit Log — complete accountability. The approval and execution trail: who approved what, which action a guardrail blocked, and the resulting change.

[Audit Log: approval-granted and guardrail-denied entries with actor, action and target]

Secure AI Operations — a governed starting point. Connect a client, mint a least-privilege read-only key, and activate a starter guardrail preset (Observe only / Safe operations / Approval required).

[Secure AI Operations onboarding: connect a client, create a read-only key, choose a starter guardrail preset]

Dashboard — the reach. Real-time status across Docker hosts and Kubernetes workloads, with live host metrics — the surface the governance model sits in front of.

[Container overview dashboard: total/running/stopped/unhealthy counts and a connected Docker host with live CPU and memory charts]

Screenshots are captured against an anonymized demo dataset. See docs/product/demo-script.md for the scripted walkthrough and docs/product/screenshots.md for how each image is reproduced.

Roadmap & beta

Beta is feature-rich but not finished. Planned / not-yet-implemented:

Accounts & roles

• 2FA (TOTP) and passkeys for the local login; finer-grained per-server roles / teams.

Monitoring & triggers

• Proper traffic / anomaly detection and a dedicated "extreme traffic" trigger (today: sustained CPU/RAM/disk thresholds + a simple rolling-z-score outlier).

Fleet & deployment

• Live Docker events (vs. polling); editing resource limits in place; richer Compose templates.
• Kubernetes track B.3: pod exec terminal, MCP tools for pods, metrics-server stats (managing k3s clusters + running on Kubernetes via Helm shipped in 0.12).

Agent governance (building on the shipped Agent-History, approvals/human-in-the-loop and rich chat widgets)

• Extend approvals to block external/direct MCP calls too — today the human-in-the-loop gate covers the in-process agent; direct tools/call requests are recorded but not held for approval.
• Real per-tool diffs in the approval card (show exactly what a write would change).
• A fuller rich-widget / MCP-Apps catalog beyond the curated chart + status card.

Hardening & resilience

• Break-glass / disaster recovery: short-lived SSH certificates issued via step-ca as an auditable emergency-access path when the normal control plane is unavailable.
• Distroless/chiseled "remote" image for the hardened profile (a locked-down container profile and a pluggable Tailscale/NetBird VPN already ship — see docs/container-hardening.md).

Have a request? Open an issue.

Project structure

Each source folder carries its own README.md describing the files within. High-level map:

Whiskers/├── src/│   ├── Whiskers/            # the application│   │   ├── Components/         # Blazor UI (Pages, Layout, Shared)│   │   ├── Configuration/      # strongly-typed settings classes│   │   ├── Hubs/               # SignalR hubs (container + terminal streams)│   │   ├── Mcp/                # MCP server tools + permission layer│   │   ├── Models/             # data models (Agent, Cloud, Cve, Hetzner, Hostinger)│   │   ├── Services/           # all business logic (see Services/README.md)│   │   ├── Utils/              # small helpers (secret redaction, shell quoting)│   │   ├── wwwroot/            # static assets│   │   └── Program.cs          # composition root (DI, middleware, MCP, auth)│   └── Whiskers.Tests/      # xUnit test suite├── deploy/telemetry/           # mesh/mTLS deploy templates (node_exporter, VictoriaMetrics, Tailscale ACL)├── docs/ARCHITECTURE.md        # architecture, security-critical areas & change control├── docs/product/               # product positioning (see docs/product/POSITIONING.md)├── Dockerfile├── docker-compose.yml└── README.md

The Services/ tree is the heart of the app — see src/Whiskers/Services/README.md for a guided tour.

Code conventions

• Interface-first: services are defined behind an IFoo interface and registered in DI; consumers depend on the interface.
• English in-code comments and XML docs throughout; the user-facing UI is localized (English default, German available).

Development

Requirements: the .NET 10 SDK.

# builddotnet build src/Whiskers/Whiskers.csproj
# run testsdotnet test src/Whiskers.Tests/Whiskers.Tests.csproj
# run locally (listens on :8080)dotnet run --project src/Whiskers/Whiskers.csproj

Contributing

Issues and pull requests are welcome. Please:

• keep changes interface-first and add/extend tests under src/Whiskers.Tests/;
• run dotnet build (0 warnings) and dotnet test before opening a PR;
• write in-code comments and docs in English.

License

Apache License 2.0, see LICENSE.

Copyright © 2026 Whiskers Contributors