Kedin — a LinkedIn ghost‑writer AI agent (public demo)

Kedin is an AI agent you manage from Telegram. You DM it a topic (or just let its daily job run); it ghost‑writes a LinkedIn post in your voice, generates a clean UML / architecture diagram to go with it, sends both back to your Telegram chat for review, and — only after you approve — publishes to your LinkedIn feed.

It is orchestrated by OpenClaw: OpenClaw hosts the agent, routes your Telegram messages to it, runs the LLM, and enforces the approval/permission model. Kedin itself is the persona (agent-config/) plus a small toolkit (*.js / *.sh) that talks to the LinkedIn API and renders the diagrams.

This is a public, sanitized, dummy‑data version of a private agent. It's a sibling of a digital twin career agent — the same core idea (an agent that represents a person's professional identity online), but with every real credential, path, and personal detail stripped out and replaced with a fictional developer persona ("Sam Carter"). Bring your own credentials and profile and it's yours.

⚠️ Everything in agent-config/USER.md is fake. Replace it with your real profile before posting.

---

What it does

• Drafts posts in your voice from a rotating topic pool, grounded in your profile (USER.md) and an evolving style guide (feedback-dataset.md) — no emojis, no hype, ~70–130 words.
• Illustrates each post with a crisp, legible, colored Mermaid diagram (flowchart / sequence / class / state / ER), rendered to PNG. Falls back to a Google Gemini diffusion image for abstract topics.
• Reviews before publishing: in review mode it sends you the draft + diagram and waits; in autonomous mode it posts directly. You teach it your voice with /feedback.
• Publishes to LinkedIn via the official /rest/posts API (text or text+image), and can reshare, edit, or delete posts.

Architecture

You drive Kedin from Telegram; OpenClaw orchestrates the agent and brokers every external call.

flowchart LR
    U["You<br/>(Telegram)"]:::user

    subgraph OC["OpenClaw gateway (orchestration)"]
        direction TB
        CH["Telegram channel<br/>(route + approvals)"]:::oc
        AG["Kedin agent<br/>(LLM + agent-config/)"]:::oc
        CH <--> AG
    end

    subgraph TK["Kedin toolkit (Node + bash)"]
        direction TB
        POST["post.js / repost.js<br/>(lib.js: OAuth, Images API)"]:::tool
        REND["render-diagram.sh<br/>(mermaid-cli + Chromium)"]:::tool
        GEN["gen-image.sh<br/>(Gemini fallback)"]:::tool
    end

    LI["LinkedIn API<br/>/rest/posts"]:::ext
    GG["Google Gemini<br/>(image model)"]:::ext

    U -- "/draft, /image, /post, /autopost" --> CH
    AG -- "draft + diagram for review" --> CH --> U
    AG -- "publish (on approval)" --> POST --> LI
    AG -- "render UML" --> REND
    AG -- "abstract topic" --> GEN --> GG
    REND -- "outbox-image.png" --> AG

    %% Harmonious soft palette: cool blue → indigo → teal, with one warm amber accent; dark legible text.
    classDef user fill:#dbeafe,stroke:#60a5fa,color:#1e3a8a
    classDef oc   fill:#e0e7ff,stroke:#818cf8,color:#3730a3
    classDef tool fill:#cffafe,stroke:#22d3ee,color:#155e75
    classDef ext  fill:#fef3c7,stroke:#fbbf24,color:#92400e

    style OC fill:#f8fafc,stroke:#e2e8f0,color:#475569
    style TK fill:#f8fafc,stroke:#e2e8f0,color:#475569
    linkStyle default stroke:#94a3b8,stroke-width:1.5px

Loading

Flow: you DM a command → OpenClaw routes it to the Kedin agent → the agent drafts the post and a Mermaid diagram → both come back to your Telegram for review → on /post, the toolkit publishes to LinkedIn. The toolkit is plain Node stdlib + bash and also runs standalone for the LinkedIn and diagram parts; the persona/behavior lives in agent-config/ and runs inside the OpenClaw gateway.

Repository layout

kedin/
├─ lib.js                 # shared: OAuth/refresh, HTTPS, Images API upload, emoji/Unicode sanitize
├─ oauth.js               # one-time LinkedIn OAuth (url / exchange <code>) → stores token + member URN
├─ post.js / post.sh      # publish outbox.md (text or --image); --dry-run builds the request, posts nothing
├─ repost.js / repost.sh  # reshare a post with your commentary
├─ delete.js / editpost.js (+ .sh)  # delete / edit a published post
├─ gen-image.js / gen-image.sh      # Gemini diffusion image (house-styled UML look)
├─ render-diagram.sh      # Mermaid source → crisp PNG (via mermaid-cli + Chromium)
├─ mmd-sanitize.js        # auto-repairs common LLM Mermaid mistakes (unquoted labels, etc.)
├─ draft-diagram.sh       # render Mermaid + deliver to chat (diffusion fallback if it won't render)
├─ draft-image.sh         # diffusion image + deliver to chat
├─ send-image.sh          # push a candidate image to your chat (openclaw message send --media)
├─ mermaid-puppeteer.json # Chromium path + flags for mermaid-cli
├─ agent-config/          # the agent's persona/behavior (IDENTITY, AGENTS, COMMANDS, SKILL, USER, …)
├─ .env.example           # all configuration (copy to .env)
└─ credentials/           # gitignored; *.example.json shows the schema

Prerequisites

• Node.js ≥ 18 (the toolkit uses only the standard library — no npm install for the core).
• For diagrams: @mermaid-js/mermaid-cli and a Chromium binary.

  npm install -g @mermaid-js/mermaid-cli      # provides `mmdc`
  # point mermaid-puppeteer.json at your chromium (default: /usr/bin/chromium)

• A LinkedIn developer app (for client_id / client_secret and the w_member_social product).
• (Optional) a Google Gemini API key — only for the diffusion image fallback.
• (Optional) an OpenClaw gateway — to run Kedin as a chat agent and deliver images to chat.

Setup

git clone <your-fork> kedin && cd kedin
cp .env.example .env                      # then edit .env with your values
cp credentials/linkedin/kedin.example.json credentials/linkedin/kedin.json
#   → fill in client_id + client_secret from your LinkedIn app

set -a && . ./.env && set +a              # load env into the shell
node oauth.js url                         # open the URL, approve, copy the ?code=…
node oauth.js exchange <code>             # stores access_token + your member URN

# try it (no posting):
echo "Hello from Kedin." > workspace/outbox.md
node post.js --dry-run workspace/outbox.md   # prints the request it WOULD send

# render a diagram:
printf 'flowchart LR\n A["API"] --> B["Worker"]\n' > workspace/outbox-diagram.mmd
./render-diagram.sh                          # → workspace/outbox-image.png

Finally, customize agent-config/USER.md (your real profile) and autopost-topics.md (your topics), then run Kedin inside OpenClaw and talk to it: /draft, /image, /post, /autopost, … (see agent-config/COMMANDS.md).

Security notes

• No secrets in this repo. Credentials, tokens, and the member URN load at runtime from credentials/ (gitignored) — never hardcoded. .env and credentials/ are excluded by .gitignore; only *.example.json schemas are tracked.
• Dummy identity. USER.md and kedin.example.json are fictional placeholders. Swap in your own.
• Approval‑gated by default. review mode never publishes without your explicit /post.
• Least‑privilege chat send. Images are staged into an allowed media dir before delivery (OpenClaw blocks sending files straight out of agent workspace dirs).
• Keep your fork private if you put your real USER.md/profile in it.

Credits

Inspired by the "digital twin" career‑agent pattern — an agent that speaks for a person's professional self online. This public edition keeps the engineering and throws away the identity. Diagrams by Mermaid; image fallback by Google Gemini; runs on OpenClaw.