MCP 服务器

Stela

Self-hosted MCP server for publishing, sharing, and reviewing AI agent artifacts with version history and pinned comments.

README

        ███████╗ ████████╗ ███████╗ ██╗       █████╗
        ██╔════╝ ╚══██╔══╝ ██╔════╝ ██║      ██╔══██╗
        ███████╗    ██║    █████╗   ██║      ███████║
        ╚════██║    ██║    ██╔══╝   ██║      ██╔══██║
        ███████║    ██║    ███████╗ ███████╗ ██║  ██║
        ╚══════╝    ╚═╝    ╚══════╝ ╚══════╝ ╚═╝  ╚═╝
           self-contained · versioned · pinned in place

Stela is a self-hosted home for agent-built artifacts — the self-contained HTML that AI agents produce (dashboards, prototypes, one-pagers). Any MCP-speaking agent can publish to it; anyone your gate admits can view, and collect pinned visual comments on a stable URL with full revision history. Inspired by Claude's artifacts — built so no vendor decides who gets to see your work.

A stela is an inscribed stone slab erected for public display. Same idea, less quarrying.

Stela home

Why

Every AI vendor now hosts artifacts in its own silo, shared on its own terms — usually a fully public link for the whole internet, or a seat inside the same walled garden. Neither one is "show this to my team, behind our own login." And each vendor's gallery only holds its own agent's work.

Stela is that missing middle, and stays small being it. One Node process, one SQLite file, no accounts to manage, no cloud dependencies. Your agents — whichever vendor, however many — publish to one place; anyone your gate admits can view, comment, and follow along as you revise.

  any MCP agent ──"Stela this"──▶  publish_artifact ─▶  ┌──────────────────────────────┐
  (Claude Code / ChatGPT /                              │  Stela                       │
   Grok / Copilot / your own)                           │  · versions, immutable       │
                                                        │  · sharing, three tiers      │
                                                        │  · pinned comments 📌        │
                                                        └──────────────┬───────────────┘
                                                                       │ stable share URL
                                          your auth proxy ◀────────────┘
                                                  │
  teammate (no AI seat at all) ──opens link──▶ sandboxed iframe (no-network CSP)
                                                  │
                                       drops pinned comments ──▶ read_comments ──▶ the agent

What it does

📤 Publish from any agent — "publish this to Stela" / "Stela this" in Claude Code, claude.ai, ChatGPT, Grok, Copilot Studio, or anything else that speaks MCP turns the current artifact into a hosted, shareable page.
🔗 Stable URLs + revision history — republishing creates a new immutable version at the same URL; every version stays viewable.
🔒 Three-tier sharing — private (you), everyone (anyone signed in to your server), or restricted (named people, by user id or email).
📌 Pinned visual comments — reviewers pin feedback to a spot on the artifact; threads resolve, scoped to the version they were made on.
♻️ Close the loop — read_comments pulls the feedback back into whatever agent published the thing, so it can address it and republish.
🛡️ Safe rendering — artifacts run in an opaque-origin sandboxed iframe under a strict default-src 'none' CSP: no network egress, no reach into the portal.
🪶 Zero heavy dependencies — storage is Node's built-in node:sqlite writing one file in DATA_DIR. Backup = copy the file. (An Azure Table/Blob driver ships too, if that's your thing.)

Quick start

Local, zero config

Requires Node ≥ 22.5 (and pnpm via npx pnpm or corepack).

git clone https://github.com/kdowding/Stela
cd Stela
npx pnpm install
npx pnpm dev

Open http://localhost:5173 — you're signed in as a dev user, storage lives in packages/app/.data/. Then connect an agent — the Connect button in the app walks through each client. For Claude Code it's one command:

claude mcp add --transport http stela http://localhost:5173/mcp

then /mcp to authenticate (the consent screen opens in your browser). For ChatGPT, Grok, claude.ai, or Copilot Studio, add Stela as a custom connector / remote MCP server pointed at http://localhost:5173/mcp — those hosts are allowlisted out of the box, and OAUTH_ALLOWED_CLIENT_HOSTS admits any other. Then say "Stela this" in any chat. Even purely local, you get what a file on disk can't give you: stable URLs, revision history, and pinned comments the agent can read back.

Connect an agent

Docker demo

docker compose --profile demo up

Open http://localhost:8080 — a bundled Caddy stamps a static demo identity on every request. Single-user, for kicking the tires; not an auth system.

For real

Stela does not manage accounts — see Security model. Put an identity-aware proxy in front of the container and tell Stela which headers to trust:

ORIGIN=https://stela.example.com        # the public URL users hit
AUTH_MODE=header
AUTH_HEADER_ID=cf-access-authenticated-user-email   # whatever YOUR proxy injects
AUTH_HEADER_NAME=...                    # optional
AUTH_HEADER_EMAIL=...                   # optional
STELA_API_KEY=<long random value>       # admin/CI publishing
STORAGE_DRIVER=sqlite
DATA_DIR=/data

Good gates, in rough order of effort:

Cloudflare Tunnel + Access — no open ports at all; free logins (Google/GitHub/email one-time-code) for up to 50 users; set AUTH_HEADER_ID=cf-access-authenticated-user-email.
oauth2-proxy — federate to any OIDC provider; headers x-auth-request-user / -email.
Authelia — fully local accounts (its own user database + TOTP), no external IdP.
Azure App Service Easy Auth — supported as a preset: AUTH_PRESET=easyauth.

The compose.yaml in this repo is the deployment skeleton — it deliberately does not publish Stela's port; traffic must come through your proxy.

Serverless hosts (Vercel, etc.): not yet. Stela wants a long-lived process and a disk — SQLite needs a filesystem, live comment notifications use SSE, and header-trust needs a proxy actually in front. A network storage driver + verified-JWT auth mode are the roadmap items that would change this.

Security model

Stela has no account system, on purpose. Identity is delegated to whatever sits in front of it: if a request arrives with the trusted identity header, that identity is a user — first sight is enrollment. The gate decides who can access; the header decides who they are; Stela only does authorization (who owns what, who a thing is shared with).

The consequences you should actually know:

Never expose Stela's port directly. With header-trust on, anyone who can reach the port can claim any identity. Stela refuses to boot in production without AUTH_MODE configured, and logs precisely which headers it trusts — but the network posture is on you: only the proxy should reach the app.
User ids should be stable. Ownership keys off the id string in the header, forever. Prefer an immutable subject id over email if your proxy offers one.
API access is token-only — the admin key you configure, or per-user Bearer tokens minted through the OAuth/pairing consent flows (issued only to someone who first got through your gate). Secrets are stored as SHA-256 hashes.
Artifacts are hostile input, handled accordingly — arbitrary agent-authored HTML/JS renders inside an opaque-origin sandbox with a no-network CSP. The one outbound request Stela ever makes (fileUrl ingest) is SSRF-hardened and allowlisted.

Architecture

pnpm monorepo, TypeScript strict end-to-end, Svelte 5 runes:

packages/
├─ shared/   Zod schemas + types (Artifact, Version, Comment, Anchor, sharing DTOs)
├─ app/      SvelteKit — the deployable: UI + API + remote MCP endpoint (/mcp)
│  └─ src/lib/server/
│     ├─ storage/   Store interface + two drivers: sqlite (default) & azure
│     ├─ auth/      trusted-header identity + Easy Auth preset + dev shim
│     └─ oauth/     Stela as its own OAuth 2.1 AS (PKCE, DCR) for MCP clients
└─ mcp/      stdio MCP server for CLI agents (publish by file path, auto-versioning)

Two details worth reading the code for:

The Store interface (storage/types.ts) has two full implementations — sqlite.ts (node:sqlite, STRICT tables, WAL, transactions) and azure.ts (Tables + Blobs, ETag concurrency) — and one conformance suite that runs every contract test against both. The abstraction is load-bearing, not decorative.
Artifact immutability: a version is never edited; republishing appends. Identical bytes dedup to the existing version. Comments anchor to versions, so feedback never drifts under revision.

Development

npx pnpm install
npx pnpm dev                # app on :5173, dev identity, sqlite in .data/
npx pnpm -r check           # svelte-check + tsc --strict (0 errors expected)
npx pnpm -r test            # vitest; sqlite tests need nothing extra
npx pnpm dev:storage        # optional: Azurite, only for the azure-driver test leg

The storage conformance suite (packages/app/src/lib/server/storage/*.conformance.test.ts) runs against SQLite always, and against Azurite when it's listening.

Design

The interface is its own little design system — cut basalt, ochre pigment, lapis seal — documented in DESIGN.md. The mark is a boustrophedon: one continuous incision snaking through three rows, the way the oldest stelae were actually written. The OAuth consent screen recognizes whoever comes knocking:

Consent screen

Status

Stela is complete and stable, shared as-is. It does what it set out to do; expect slow-to-no maintenance rather than a roadmap.

License

MIT