# Set up Wannalead Wannalead gives your AI agent live LinkedIn capabilities — find who to contact, understand why now, and send the message — from the user's own LinkedIn account, over a workspace-scoped HTTPS API. Do this setup yourself; do not ask the user to run terminal commands. ## 0. First, identify where you run The setups differ, and the wrong one cannot work. - **Hosted assistant without code execution** (claude.ai, Claude Desktop chat, ChatGPT): install nothing, and do not use an API key, even if the user pasted one: a hosted connector signs in with OAuth. Tell the user, in three steps: 1. Open Settings, then Connectors, then add a custom connector. 2. Paste the URL `https://wannalead.co/mcp` and approve the Wannalead sign-in. 3. Start a new chat and ask for the task again. Then stop. In that new chat, the connector gives you the Wannalead tools and `wannalead_get_skill` serves the skills. - **Local coding agent** (Claude Code, Codex, Cursor, opencode on the user's machine): the user has authorized the whole flow. Continue with step 1. - **Sandboxed agent** (claude.ai or Claude Desktop with code execution, Cowork, a cloud container with restricted network): try step 1 once. If `wannalead.co` is blocked (proxy 403, DNS failure, timeout), do not retry and do not try other hosts. Offer the user two ways past it: 1. Allow the domain: add `wannalead.co` to the allowed domains of code execution (claude.ai: Settings, then Capabilities; on a Team or Enterprise plan, an admin does it). Then retry step 1 in a new chat. The install lasts for that session only, so it runs again in each new one, with the same key. 2. Install the MCP connector, which persists across chats and needs no key: open Settings, then Connectors, add a custom connector with the URL `https://wannalead.co/mcp`, approve the Wannalead sign-in, then start a new chat and ask for the task again. ## 1. Install the CLI, agent skills and MCP With the API key the user gave you (it starts with `wlk_`): curl -fsSL https://wannalead.co/launch.sh | bash -s -- --agent --token wlk_... This installs the `wannalead` CLI to `~/.local/bin`, saves the key, copies the Wannalead skills into every detected agent (Claude Code, Codex, Cursor, GitHub Copilot, opencode) and registers the MCP server `https://wannalead.co/mcp` with Claude Code, Codex, Cursor and opencode. If `~/.local/bin` is not on your PATH, add it. Never print the key, put it in a URL, or commit it to a repository. No key yet? Send the user to https://wannalead.co/sign-in to create a workspace and connect LinkedIn, then come back with the key shown at the end of onboarding. Agents load a new MCP server on their next start. Until then, use the CLI in this session. ## 2. Load the skills Invoke the core `wannalead` skill, then the `wannalead-onboarding` skill (Claude Code: the Skill tool or `/wannalead-onboarding`; Codex: `$wannalead-onboarding`). Onboarding identifies the account and loads the saved commercial context. It never blocks the user's task: incomplete context is built up progressively. If those skills are not visible in this session (they were installed just now), fetch them and follow them as if they were installed: wannalead skills get wannalead --full This prints the core skill together with `wannalead-onboarding` and `endpoints`. Before a task, load its skill the same way, for example `wannalead skills get campaign-automation`. ## 3. Run onboarding The first authenticated call must be: wannalead me Then follow `wannalead-onboarding`: load the workspace context it names for the returned LinkedIn identifier and continue with the user's request, even if that context is incomplete. Before any business call, inspect the endpoint catalogue and the exact OpenAPI contract for the endpoint you plan to use. ## 4. Show a result, not an install log Tell the user in one line which LinkedIn account and Wannalead workspace are connected. When the user has not named a task, offer the six use cases (the `wannalead-use-cases` skill carries the menu and the skill to load for each) and propose to run one on their own data: 1. Lookalikes with signals: describe the ICP, get every lookalike profile with the signals that matter (tools in use, posts, job changes). 2. Campaigns: from the user's context and writing style, draft the messages and run sends and follow-ups. 3. Dormant conversations: revisit LinkedIn threads left without a reply and follow up on leads that went quiet. 4. Dormant network: surface first-degree connections never contacted and propose an approach. 5. Saved post to ICP: the user saves a post whose engagers look like the target; pull them and keep those matching the ICP. 6. Messy list to contacts: reconcile a rough CSV of companies with LinkedIn and list the people to contact in each. ## Reference - Core skill: https://wannalead.co/.well-known/skills/wannalead/SKILL.md - Onboarding skill: https://wannalead.co/.well-known/skills/wannalead-onboarding/SKILL.md - Endpoints skill: https://wannalead.co/.well-known/skills/endpoints/SKILL.md - Skills index: https://wannalead.co/.well-known/skills/index.json - Endpoint catalogue: https://wannalead.co/v1/endpoints - OpenAPI contract: https://wannalead.co/openapi.json - Machine index: https://wannalead.co/llms.txt ## Safety Every LinkedIn write runs on the user's own account and is hard-capped well under LinkedIn's own thresholds. Never send invitations or messages without the user's explicit approval. Before a loop of writes or a paid call, tell the user how many actions or how much it costs.