title: Lara Marketing Connector emoji: 🚀 colorFrom: blue colorTo: indigo sdk: docker app_port: 7860 pinned: false license: mit

Lara Marketing Connector

A remote MCP server ("connector", AgentFactory style) that carries Lara's full marketing brain — 45 marketing skills + the "which-skill-when" orchestration logic + persona + per-client context — so any MCP host (Claude Code, Codex, OpenCode, OpenClaw, claude.ai) connects to one URL and gets the complete digital-marketing employee.

Built specs-first — see specs/. Status: v0.1 (public/demo).

Live MCP endpoint (when the Space is running): https://rrizwan98-lara-marketing-connector.hf.space/mcp

Related MCP server: Skills MCP Server

Layout

lara-connector/
├── specs/            # 00-overview … 06-auth  (read these first)
├── server.py         # FastMCP gateway + all 13 tools
├── auth.py           # verified identity (demo now, OAuth seam for prod)
├── session.py        # HMAC session token (seals sub + tier)
├── db.py             # SQLite state (users, clients, deliverables, usage)
├── config_store.py   # Lara's brain: persona, rules, router map, tiers
├── skills_repo.py    # serve skills/<name>/SKILL.md (path-safe)
├── skills/           # the 45 marketing skills
├── Dockerfile        # Hugging Face Space (Docker SDK), port 7860
├── .github/workflows/deploy.yml   # test -> (pass) -> deploy to HF
└── tests/            # test_smoke.py (in-memory) + test_http.py (live)

CI/CD

Push to main → GitHub Actions runs tests/test_smoke.py → only if tests pass, the repo is synced to the Hugging Face Space, which rebuilds the Docker image and serves the connector.

Run locally (demo — no sign-in, all access)

pip install -r requirements.txt
python server.py            # serves http://127.0.0.1:8000/mcp
python tests/test_smoke.py  # 16 checks

Defaults: AUTH_DISABLED=1, GATING_ENABLED=false, PUBLIC_TIER=max, SQLite. Set SESSION_SIGNING_SECRET before any public deployment.

The tools (13)

• Open: health, begin_session (call first — returns persona, rules, router_map, skills, token)

• config_*: config_get_persona, config_get_rules

• domain_*: domain_list_skills, domain_get_skill, domain_route_task

• user_*: user_get_profile, user_list_clients, user_get_client_context, user_save_client_context, user_get_history, user_log_deliverable

Every tool except health/begin_session needs the session_token from begin_session.

ChatGPT App (Apps SDK widget) — progressive enhancement

The server ships a widget template (ui://widget/lara-v1.html, MIME text/html;profile=mcp-app, source web/lara-widget.html). In ChatGPT, begin_session, domain_list_skills, domain_route_task, and domain_get_skill render an interactive Lara panel (session dashboard, 45-skill catalog with search, route results, skill playbooks). Tool _meta carries both the MCP Apps standard key (ui.resourceUri) and the ChatGPT alias (openai/outputTemplate). Hosts without widget support (Claude, Claude Code, Codex, CLIs) ignore the template and get the exact same text/structured responses as before — no behavior change.

Connect from a host

• ChatGPT (as an App): Settings → Apps & Connectors → enable Developer mode → Create app → paste the /mcp URL. Widget-enabled tools render the Lara panel.

• claude.ai: Settings → Connectors → Add custom connector → paste the /mcp URL.

• Claude Code / Codex: add an MCP server pointing at the /mcp URL.

• OpenClaw: add it as an MCP server in openclaw.json.

Tiers (free / pro / max)

Built in now, enforced later (see specs/05-tiers.md): wire OAuth (AUTH_DISABLED=0), add billing → db.set_tier(sub, tier), set GATING_ENABLED=true.

The 4 invariants (the book)

One gateway · tools only · identity from verified sign-in (never model args) · fail closed.