
liveportfolio
io.github.cliffordnwanna/liveportfolio
Read and edit your LivePortfolio portfolio from your AI assistant. Requires a Pro account.
Documentation
LivePortfolio MCP
A production Model Context Protocol (MCP) server that lets AI agents (Claude, ChatGPT, Claude Code) read and edit a user's LivePortfolio portfolio safely. It is a worked example of agent tool design with security built in: OAuth 2.1 with PKCE and per-user consent, assistant tokens accepted only at the MCP endpoint, versioned edits with compare-and-set and undo, and human-in-the-loop approval for risky changes, so an agent never gets unchecked control.
Engineering highlights
- Human-in-the-loop approval for removals, whole-list replacements, batches and new links (out of band: the agent cannot approve for itself)
- Versioned, undoable edits with optimistic concurrency
- Prompt-injection stance: user content is data, never instructions
- 56 tools covering every editable field, each tested end to end, and a documented threat model
Architecture · Security model · Tool reference · Design decisions
This repository holds the documentation and the registry listing (server.json) for the hosted server. The application itself is not open source.
- Server address:
https://liveportfolio.site/mcp(Streamable HTTP, OAuth 2.1 with PKCE) - Requires a LivePortfolio Pro account
- Docs: https://liveportfolio.site/docs/mcp · Connect guides: https://liveportfolio.site/connect · Agent index: https://liveportfolio.site/llms.txt
How it works
flowchart LR
A[Assistant client<br/>Claude, ChatGPT, Claude Code] -->|HTTPS + OAuth token| M[MCP endpoint<br/>/mcp]
A -. sign-in .-> O[OAuth 2.1 server]
O --> C[Consent page<br/>liveportfolio.site]
M --> G{Checks on every call:<br/>token, app connection,<br/>access, rate limit}
G --> T[Tools]
T --> P[Operation layer<br/>validation + limits]
P --> V[Versions and history<br/>compare-and-set]
V --> D[(Database)]
T -->|needs approval| Q[Approval request]
Q -. link .-> U[You, on the website]
U --> R[Approval page<br/>owner only]
R --> P
D --> S[Public portfolio site]
Documentation
- Architecture: system overview and sequence diagrams for sign-in, a direct edit and an approval-gated edit
- Security: what an assistant can and cannot do, controls, prompt injection, pictures, known limits
- Tools: all 56 tools grouped by section, with limits and what the public page shows
- Design decisions: trade-offs, and the known gap after approval
- Changelog
Assistants
Tested: Claude (web and desktop), ChatGPT and Claude Code. Gemini CLI, Cursor and VS Code may work (any client that speaks MCP over HTTP with OAuth) but may need manual setup.
Connect
- Add
https://liveportfolio.site/mcpas a custom connector. Steps with screenshots: Claude, ChatGPT, others. - Sign in, choose what the assistant may do, and approve. Then start a new chat with the first message shown on the connect page.
Claude Code:
claude mcp add --transport http liveportfolio https://liveportfolio.site/mcp
What it can do (56 tools)
- Read: the portfolio (by section), the editing rules, designs, a completeness checklist, history, a preview of a change, a health check.
- Profile and contact: name, role, headline, About, location, email, phone, WhatsApp, social links; add to a field or replace one phrase.
- Skills: flat list and groups (add, remove, rename, reorder, replace).
- Experience and projects: roles and highlight lines; projects with stack, links, pictures and an optional Power BI dashboard.
- Education, certifications, testimonials: add, fix one, remove one, reorder, replace.
- Photos and design: upload link or a public https picture; template, look, Pro design version.
- Safety net: undo, restore an earlier version, and one batch of up to 60 operations for a whole-CV update.
Security model
- OAuth 2.1 with PKCE. Assistant tokens work only on
/mcp; browser tokens do not work there, and assistant tokens are refused on every other API route. - Access is per person and checked on every call. Disconnect at https://liveportfolio.site/oauth/connections and the next call fails.
- Direct edits are undoable. Removals, whole-list replacements, batches, new links and large cuts need your approval on liveportfolio.site; the assistant cannot approve for itself.
- Your email, phone and WhatsApp are hidden from the assistant unless you allow them. You can connect read-only.
- We do not store your chat messages. See the privacy policy.
Limits
60 operations per batch, 30 flat skills, 1500 characters per project text field, 120 calls per 10 minutes, 60 edits per hour. The public page shows the first 5 stack items, the first testimonial and, when you have skill groups, the groups rather than the flat list; the assistant is warned when something is saved but not shown.
It cannot
Publish your portfolio, change your address or domain, touch billing, change your sign-in, delete anything, or create your first portfolio.
Contact
No installation packages available.