LinkedCamp

LinkedCamp drives LinkedIn and email outreach: campaigns, prospect sourcing, multi-step sequences, a unified inbox and AI reply agents. Everything goes through the hosted LinkedCamp MCP server, https://mcp.linkedcamp.com/mcp. Treat it as the only source of truth for account data.

0. Make sure the tools are connected

OpenClaw exposes the server's tools as linkedcamp__<tool> — for example linkedcamp__whoami, linkedcamp__campaign_list. This skill names tools without the prefix.

If no linkedcamp__* tools are available, LinkedCamp is not connected yet. Do not improvise with web requests or curl. Read {baseDir}/references/setup.md and walk the user through it. Short version, run on the Gateway host:

openclaw mcp set linkedcamp '{"url":"https://mcp.linkedcamp.com/mcp","transport":"streamable-http","auth":"oauth"}'
openclaw mcp login linkedcamp

The login prints a URL; the user opens it and pastes their API key from LinkedCamp → Settings → Integrations into LinkedCamp's own page.

Never ask for the API key in chat, and never write it into a command line, config value or file yourself. If the user pastes one into the conversation anyway, do not repeat it back; tell them to rotate it in Settings → Integrations, because chat transcripts are stored.

1. Orient first

1. Call whoami once at the start of a conversation. _derived reports the plan, isPro, permissions, and which Pro-only steps are locked.
2. Call reference before authoring a sequence or filtering by a status or enum. The vocabulary is not guessable — campaign daily limits live under campaginLimits, spelled exactly that way.
3. Read before you write. Use ids returned by tools; never fabricate an id, metric, prospect, message or account state.
4. Say plainly whether something is a proposal, a saved draft, or a live change.

The server also publishes mcp://guide/campaign-playbook, the long-form build flow, if resources are available to you.

2. Building a campaign

1. campaign_create — creates a DRAFTED campaign.
2. search_url_build — turns the described audience into a search URL. It generates job titles and keywords for basic LinkedIn and Sales Navigator searches. It cannot generate geography, industry, company size, seniority or years-in-role: Sales Navigator takes those only as its own numeric ids, which cannot be derived from a name. Ask the user to paste a URL with those filters set. Recruiter (/talent/) URLs always come from the user.
3. campaign_set_source — show the user the URL first; it decides who gets messaged.
4. campaign_update — sending account, schedule and daily limits. LinkedIn steps need linkedAccountId; EMAIL steps need emailAccountIds.
5. flow_set — writes the sequence from an ordered list of steps. This is the only tool that authors a sequence. flow_from_template applies a prebuilt one instead; flow_get reads back what is saved.
6. campaign_get — _readiness.blockers must be empty.
7. campaign_launch — only after the confirmation in section 5.

Account and source must match

A campaign sends from exactly ONE LinkedIn account and may rotate across SEVERAL mailboxes. Never choose either for the user: if linkedin_account_list or email_account_list returns more than one, ask. Only attach mailboxes when the user wants EMAIL steps.

Then ask whether that LinkedIn account has Sales Navigator, Recruiter, both or neither — the API does not expose it.

• Neither: basic LinkedIn search URL. A /sales/ URL will not work and INMAILS steps cannot run.
• Sales Navigator: /sales/ search URL; INMAILS available.
• Recruiter: a /talent/ URL from the user and an INMAILS-only sequence (no connect or ordinary message step), or a normal campaign from a basic search — ask which.

InMails also need credits to reach anyone who is not an Open Profile member. Credits are not visible here. Without them roughly one InMail in seven lands and the rest fall through to the next branch — raise this before launch.

Do not copy from other campaigns

• Never start by duplicating. campaign_duplicate is only for when the user names a specific campaign to duplicate.
• Never reuse another campaign's source URL. uniqueProfiles skips anyone already held by another campaign, so two campaigns on one search split it and the second finds almost nothing.
• Never reuse another campaign's tag. Mint one with tag_create (interest: "NEUTRAL", a short label from the campaign name). The reusable tags — Interested, Not Interested, Replied, Meeting Booked, Deal, Maybe — classify replies and do not belong in campaign settings.
• Never infer an agent. See section 3.

Editing a sequence

To change a draft's copy, call flow_set again. Do not use sequence_update on a draft: launching rebuilds every step from the saved flow graph and silently reverts the edit. sequence_update is for ACTIVE or PAUSED campaigns. flow_set on an ACTIVE campaign re-evaluates every lead and can cancel or reschedule pending actions; flow_audit reports what moved.

3. Replies and AI agents

An agent attached to a campaign answers real prospects by itself, so never attach one unprompted. Ask, and offer both routes:

• existing — agent_list, then agent_get to show its personality, intent and instructions so the user can judge the fit;
• new — agent_upsert with isActive: false, kb_upsert + kb_attach for source material, agent_train, then test with agent_query before it touches a real conversation.

No agent is a normal outcome: replies land in the inbox for a human.

For the inbox: conversation_list / conversation_get to read, message_draft_ai_reply to draft. Present drafts as drafts.

4. Plan and permission gates

The server enforces these; a refusal is real and has no workaround. Pro and Agency plans carry isPro: true; Turbo does not.

• Pro only: persona_*, connecting an email account, the EMAIL, EVENT, GPT, FINDER and QUALIFIER steps, the If_Interested and CUSTOM triggers, and ENGAGEMENT.commentsToAdd.
• Permission-scoped: agent_* and kb_* need Agents; post_* needs Publish; QUALIFIER also needs Personas.

Check whoami._derived before proposing any of these. If a tool refuses on plan grounds, say so and stop.

5. Confirm before anything reaches people, changes live outreach, or destroys data

When the LinkedCamp plugin is installed, it pauses the run and asks the user to approve every call listed in this section through OpenClaw's plugin approvals: every tool the LinkedCamp server flags destructive, plus the unflagged calls marked † below. That prompt is a backstop, not the conversation: ask first anyway. With the skill alone there is no prompt at all and you are the only safeguard — especially on a chat channel where the user is typing on a phone. For each call: say exactly what will happen, show the content and recipients, get an explicit "yes" in the conversation, and never put several behind one yes. If the approval is denied or times out, stop and say so; do not retry or look for another tool that does the same thing.

Reaches real people:

• campaign_launch — starts real outreach on the campaign's schedule. Show the audience URL and the full sequence first. Never bulk-launch.
• campaign_set_status with ACTIVE † — resumes a paused campaign; it may have been paused for a reason. Pausing needs no ceremony and is not gated.
• message_send — delivers immediately on LinkedIn or email and cannot be recalled. Show the exact text and recipient.
• campaign_send_test_email — a real email. Show recipient, subject and body.
• post_upsert — unless status is DRAFTED (private) or REQUIRE_APPROVAL (goes to a human reviewer), the post is queued to publish publicly. Show the exact post and when it goes out.
• post_schedule_set — replaces a weekday's publishing slots.
• automation_upsert with isActive: true † — runs against live events and can message people. Save it inactive while drafting.

Changes who is contacted or what live campaigns and agents say:

• campaign_set_source, campaign_add_prospects — decide who gets messaged. Show the URL.
• flow_set — replaces the sequence; on an ACTIVE campaign it re-evaluates every lead and can cancel pending actions. The prompt appears for drafts too, because the gate cannot see the campaign status — say so when the user is iterating on a draft.
• sequence_update — changes what a running campaign sends.
• campaign_update with agentId † — the agent will answer prospects by itself.
• agent_upsert unless isActive: false, kb_upsert, persona_update.
• template_upsert, saved_reply_upsert, lead_note_update — overwrite shared copy or notes. activity_skip — the scheduled action will not run.

Destroys or hides data — read {baseDir}/references/safety.md first: campaign_delete, lead_delete, lead_blacklist, message_delete, agent_delete, kb_delete, persona_delete, tag_delete, blacklist_delete, placeholder_delete, reminder_delete, holiday_delete.

Not gated, because they only build drafts or read: creating and configuring a campaign, adding leads, post_upsert with DRAFTED, agents and automations saved inactive, and every read.

6. Agency sub-accounts

An agency owner can work on sub-accounts from their own connection.

1. Get ids from subaccount_list — never construct or guess a userId; a wrong id does not fail, it answers for a different account. An empty result means either no active sub-accounts or a non-owner key; say so.
2. Pass one as userId to campaign_list, linkedin_account_list, email_account_list, analytics_campaign, analytics_reporting, tag_list, or agent_duplicate (a write — name the client and confirm).

Record ids (campaignId, leadId, conversationId, linkedAccountId) from a sub-account work on that sub-account, including writes. Say which account you are about to change. When comparing accounts, label every number with its account; nothing returns a combined roll-up.

7. Never invent a real-world value

Anything identifying something real — a profile or company URL (such as followCompanyUrl), a search URL, an account id, a tag, a sender address, a webhook endpoint — comes from the user or a read tool. Never from memory, another campaign, or a plausible guess. These do not fail loudly: a wrong URL sends real messages to the wrong people. If it is missing, stop and ask.

8. Out of scope

Not available, and rephrasing will not make them available: connecting a LinkedIn account (it takes the LinkedIn password and a PIN — web app only), CRM integrations, billing, team management, rotating the API key, reading stored credentials, deleting saved replies or lead notes, archiving a campaign, saving a raw flow graph, and deprecated webhooks (use automations). Never ask for a LinkedIn password, session cookie or verification code. Point the user to the LinkedCamp web app for these.

9. Responses

Ground summaries in tool results and keep observed data separate from generated copy. When data is missing, stale or plan-restricted, say so and give the next useful step. On chat channels keep answers short; offer detail rather than dumping whole sequences unprompted — except before a launch or send, where the user must see everything that will go out.

For a map of all toolsets, see {baseDir}/references/tools.md.