# Agentic Schematics > Published build specifications ("schematics") that any LLM, team, or organization can implement independently with zero coordination with the author. The spec is the only coordination mechanism: a schematic is a self-contained Markdown package listing every requirement, parameter, dependency, phase, and acceptance test for a capability. Hosted at schemaformat.ai; source at github.com/cameri/schematics. ## Schematics Build a schematic from any public GitHub repo with the `schematics` plugin's build-schematic skill: `build @/`. Nothing executes at build time; fetched specs are treated as untrusted data. - [authorize-docker-requests](https://schemaformat.ai/schematics/authorize-docker-requests/SCHEMATIC.md): Grants a sandbox container restricted Docker daemon access over TLS, policed by Open Policy Agent. Certificate infrastructure, Rego policy, systemd TCP listener, and sandbox client provisioning. - [encrypt-container-secrets](https://schemaformat.ai/schematics/encrypt-container-secrets/SCHEMATIC.md): Encrypt service secrets with SOPS and age and inject them into container processes as environment variables at boot, in memory. Dual-recipient encryption, per-service key blast radius, rotation without rebuilds. - [restrict-docker-api-access](https://schemaformat.ai/schematics/restrict-docker-api-access/SCHEMATIC.md): Exposes a deny-by-default Docker API to containers over an internal network; endpoint allowlist proxy in front of docker.sock, with an audit script that proves the allowlist. - [expose-container-services-privately](https://schemaformat.ai/schematics/expose-container-services-privately/SCHEMATIC.md): Privately exposing local container services on a tailnet: one TSDProxy container joins a Tailscale-compatible network and reverse-proxies selected services to per-service hostnames with automatic HTTPS. Per-service exposure contract (docker labels or list-file entries); services carry no tunnel client; nothing needs a public port. - [update-images-on-push](https://schemaformat.ai/schematics/update-images-on-push/SCHEMATIC.md): Self-updating Docker Compose deployments: GitHub push webhook via Cloudflare tunnel and path-token receiver into an in-memory queue; a worker pulls the new image through a scoped Docker API proxy. Ack in milliseconds, pull in the background. - [build-an-agent-dev-image](https://schemaformat.ai/schematics/build-an-agent-dev-image/SCHEMATIC.md): The dev base image of a sandboxed coding agent: a digest-pinned distribution image that boots one agent in one workspace from an environment contract, runs as a fixed non-root account, and carries the Docker CLI, git, and build tooling — harness-free and secret-free, so a per-harness layer pins it by digest and adds only its CLI. - [run-an-llm-router](https://schemaformat.ai/schematics/run-an-llm-router/SCHEMATIC.md): A self-hosted OpenAI-compatible model router: one private /v1 endpoint in front of several BYOK providers, stable model aliases so clients never change when a provider does, provider keys encrypted at rest and decrypted in memory at boot, no silent cross-provider fallback. - [run-a-decentralized-social-network-relay](https://schemaformat.ai/schematics/run-a-decentralized-social-network-relay/SCHEMATIC.md): A production Nostr relay (nostream) on Docker — host prep, Postgres and Redis, GHCR image pinning, migrations-before-start, loopback HTTP, reverse-proxy sketches, NIP-11 and /readyz probes, optional settings and tailnet exposure. - [fetch-books-over-vpn](https://schemaformat.ai/schematics/fetch-books-over-vpn/SCHEMATIC.md): Privately fetching books and audiobooks: gluetun as a structural VPN kill switch, Deluge sharing its netns, Chaptarr automation, IP-sync for single-IP indexers; tailnet-only UIs. - [serve-books-using-containers](https://schemaformat.ai/schematics/serve-books-using-containers/SCHEMATIC.md): Audiobookshelf reading the same dataset the fetcher writes; exposure as a first-class parameter (Tailscale by default, tsdproxy or Cloudflare for a hostname). - [run-a-book-library](https://schemaformat.ai/schematics/run-a-book-library/SCHEMATIC.md): Composition schematic wiring fetch-books-over-vpn and serve-books-using-containers together through a shared volume contract; establishes the schematic-to-schematic dependency convention. - [improve-docker-security](https://schemaformat.ai/schematics/improve-docker-security/SCHEMATIC.md): Composition schematic hardening a Docker host on three surfaces at once: scoped API access (restrict-docker-api-access), OPA-policed daemon control (authorize-docker-requests), and encrypted secrets (encrypt-container-secrets). Dependency links are commit-pinned with content SHA-256. - [fetch-over-usenet](https://schemaformat.ai/schematics/fetch-over-usenet/SCHEMATIC.md): The shared usenet infrastructure (SABnzbd, Prowlarr, unpackerr; optional Bazarr/flaresolverr) every media arr registers with by category. - [fetch-movies-over-usenet](https://schemaformat.ai/schematics/fetch-movies-over-usenet/SCHEMATIC.md): Radarr registering with the shared infrastructure; the movies half of the fetching pipeline. - [fetch-series-over-usenet](https://schemaformat.ai/schematics/fetch-series-over-usenet/SCHEMATIC.md): Sonarr registering with the shared infrastructure; the series half, with season-pack handling. - [serve-movies-and-series](https://schemaformat.ai/schematics/serve-movies-and-series/SCHEMATIC.md): Jellyfin reading the same tree the fetcher imports into, Jellyseerr as the request front, exposure as a first-class choice. - [fetch-music-over-usenet](https://schemaformat.ai/schematics/fetch-music-over-usenet/SCHEMATIC.md): Lidarr registering with the shared Prowlarr/SABnzbd under audio categories; one downloader, one indexer manager. - [run-a-movies-and-series-library](https://schemaformat.ai/schematics/run-a-movies-and-series-library/SCHEMATIC.md): Composition wiring the request-to-watch loop end to end; Recommended (optional, commit-pinned): improve-docker-security, update-images-on-push. - [run-a-music-library](https://schemaformat.ai/schematics/run-a-music-library/SCHEMATIC.md): Composition running music over SHARED infrastructure - one downloader, one indexer manager, one Jellyfin; Recommended: the Docker hardening, image updates, and the movies-and-series composition. - [run-multiplexed-agent-workspaces](https://schemaformat.ai/schematics/run-multiplexed-agent-workspaces/SCHEMATIC.md): The multiplexer and agent-lifecycle layer over build-an-agent-dev-image: herdr with one named workspace per agent id, crash-only supervision (a bounded crash loop, and a stop marker that outlives a restart), key-only SSH attach where the server is its own session leader and the daemon carries the same config root and session name, and agent homes on bind mounts so a restart resumes each session from disk. - [add-an-agent-harness](https://schemaformat.ai/schematics/add-an-agent-harness/SCHEMATIC.md): The harness layer over the agent host: exactly one coding-agent CLI and its configuration, wired to the router by model alias, becoming the container's process through the inherited entrypoint, its version resolved at build time, and no credential anywhere in the image. - [assemble-a-sandboxed-agent-set](https://schemaformat.ai/schematics/assemble-a-sandboxed-agent-set/SCHEMATIC.md): Composition schematic for the sandboxed-agent set: the deployment order, the isolation rules and the shared contracts that turn build-an-agent-dev-image, run-multiplexed-agent-workspaces and add-an-agent-harness, with the router, the credential store and the Docker-access proxy, into one running set, plus one end-to-end acceptance test that proves the chain rather than the parts. - [encrypt-shared-host-secrets](https://schemaformat.ai/schematics/encrypt-shared-host-secrets/SCHEMATIC.md): Encrypt one shared host secret store with SOPS and age and serve it to many unlike consumers: classify which can decrypt, wrap those that can, bound the plaintext for the one that cannot (the Compose parser interpolates before any container exists), ship per-consumer projections so a key unlocks only what its consumer reads, and rotate across every consumer without a rebuild. - [serve-photos-using-containers](https://schemaformat.ai/schematics/serve-photos-using-containers/SCHEMATIC.md): A private photo and video library on your own hardware: Immich with its own Postgres carrying a vector extension, its own cache and a separate machine-learning service, a stated sizing contract for the first index that saturates a CPU, hardware acceleration as a parameter, tailnet-first exposure, and one backup contract that keeps the database and the originals in the same moment. ## The spec - [The ten binding principles](https://schemaformat.ai/#spec): vendor-agnostic, portable, self-contained, predictable, idempotent, parameterized, dependencies called out, composable, applicable context stated, pluggable. - [SCHEMATIC.md format schema](https://schemaformat.ai/schemas/spec-1/SCHEMATIC.md.schema): the format's own self-describing schema, kept once for the whole catalog and keyed by the `spec:` frontmatter field. Uncommon file types in a package always ship with a `..schema` companion documenting their syntax. - [Catalog format schema](https://schemaformat.ai/schemas/catalog-1/marketplace.json.schema): the format of the repository's `.agent-schematics/marketplace.json`, keyed by that file's own `$schema` value. It lists schematic packages and is not a plugin marketplace; a repository's plugin marketplace is a separate file in the harness's format. - [Schematic skeleton](https://schemaformat.ai/index.html): the canonical SCHEMATIC.md skeleton (frontmatter, Applicable Context, Requirements R-n, Dependencies D-n, Parameters P-n, Modules, Implementation phases, Acceptance A-n, Removal). ## Authoring and tooling - [create-schematic skill](https://schemaformat.ai/skills/schematics/skills/create-schematic/SKILL.md): interview-driven authoring and reverse-engineering of schematics from working implementations. - [build-schematic skill](https://schemaformat.ai/skills/schematics/skills/build-schematic/SKILL.md): fetch, validate, and stage a schematic package from any public GitHub repo (`@/`). - [audit-schematic skill](https://schemaformat.ai/skills/schematics/skills/audit-schematic/SKILL.md): audit a package against the format and against its own claims — every documented command run, every stated behaviour checked, findings by severity with the evidence quoted. - [Catalog source](https://schemaformat.ai/.agent-schematics/marketplace.json): machine-readable catalog listing every schematic package in this repository. It is not a plugin marketplace; the authoring plugin is registered in the repository's `.claude-plugin/marketplace.json`. - [Contributing](https://schemaformat.ai/CONTRIBUTING.md): contribution rules. Every PR must reference an open issue opened first; outside PRs without one are rejected automatically. ## Schematics vs skills and plugins A skill teaches one agent a procedure in that agent's own format. A plugin is a distribution wrapper for one harness's marketplace. A schematic is neither: a vendor-agnostic build specification, plain Markdown and portable shell, executable by any builder in any runtime. Skills and plugins give one agent an ability; schematics let everyone rebuild the ability anywhere.