Use the returned LinkedIn public identifier to inspect ~/.wannalead/workspaces/<linkedin-public-identifier>/ONBOARDING.md. If the file is absent, initialize onboarding. If it is incomplete, resume its next question. Continue to the requested operational workflow only after onboarding records the minimum context as complete. Never treat wannalead endpoints, a task-specific skill, or a remembered workspace as a substitute for this check.

If wannalead me reports missing authentication, run wannalead auth login and retry it. If it reports a missing LinkedIn connection, use wannalead auth connect, then retry the same identity check. Do not run a business endpoint while identity or onboarding state remains unresolved. Resume saved progress instead of restarting the interview.

Workspace not subscribed yet

wannalead me works without a subscription and returns data.billing. When billing.active is false (or any call answers 402 with subscription_required), the workspace has not subscribed to the LinkedIn plan yet. This state applies to LinkedIn endpoints only. Check data.network_access separately before drawing a conclusion about Instagram or X:

Still run the full onboarding interview. It only needs wannalead me and the local workspace files. Do not call LinkedIn data endpoints during it, and do not stop the interview because of billing.
When the user asks for a first LinkedIn result, or right after the onboarding synthesis, say plainly that Wannalead needs the LinkedIn plan subscription and give billing.billing_url (https://wannalead.co/dashboard/billing?checkout=pro). That link opens Stripe checkout directly. Ask the user to tell you once it is done. Do not describe it as a trial that ended, and do not quote dates.
When a LinkedIn CLI or local MCP call answers 402 subscription_required, the CLI has already opened that checkout page in the user's browser (at most once every 10 minutes). Tell the user the payment page just opened, give the same link in case it did not, and wait for them to say it is done. Do not retry the call in a loop.
When the user says it is done, run wannalead me again. If billing.active is true, resume exactly where you stopped (the pending request, or the use-case menu). If it is still false, wait a few seconds and check once more before asking the user to verify the payment page.
For billing_action_required, ask the user to fix the payment at the same URL (it also opens on its own after a 402). For workspace_unavailable, stop and say the workspace cannot be used.

An Instagram or X add-on has its own entitlement. When its add-on is active but network_access.<network>.read_permission is false, ask the user to reconnect Wannalead and grant that network's read scope. A 403 insufficient_scope is a permission problem, not a subscription problem. If an add-on is inactive, describe that network's billing requirement without claiming the other networks are blocked. Continue on any network with an active entitlement and the needed permission.

Choose a use case

The six use cases below cover LinkedIn. Offer this menu, in the user's language, right after onboarding completes and at the start of any session where the user has not named a task. Ask which one to start with instead of guessing; then load the named skill (or, for a use case without a dedicated skill, run endpoint discovery for the listed chain). Every item runs from the user's own LinkedIn account through Wannalead endpoints only.

Lookalikes with signals. Describe the ICP; find every lookalike profile with the signals that matter (tools in use, posts, job changes). Skill: icp-lookalikes, plus job-change-signals for the trigger side.
Campaigns. From the durable context (offer, customers, writing style) draft the messages, then run sends and follow-ups. Skills: campaign-automation, warm-outreach, humanizer.
Dormant conversations. Revisit LinkedIn conversations left without a reply and follow up on leads that went quiet. No dedicated skill: chain messages.conversations → messages.history → warm-outreach for each approved follow-up.
Dormant network. Surface first-degree connections the user never contacted and propose an approach. No dedicated skill: chain me.connections → messages.conversations (to exclude existing threads) → profiles.enrich → warm-outreach.
Saved post → ICP. A post whose reactions or comments look like the target audience: the user saves it on LinkedIn, the agent pulls the engagers and keeps those matching the ICP. Skill: engagement-prospecting from the post URL; read me.saved-posts first to find the saved posts.
Messy list → contacts. A rough CSV (for example family offices): reconcile each row with the company that exists on LinkedIn, then list the people to contact in each. No dedicated skill: chain companies.search / companies.resolve → companies.employees, then account-mapping for the buying committee.

Present the menu as a short numbered list, not as documentation. When the user picks one, restate the concrete outcome in one sentence and start.

The list is not capped at six. Load the wannalead-use-cases skill for the full routing, including scraping a competitor's own audience, registering for a LinkedIn event and mining its attendees, reconciling any roster (CSV / Luma / Slack) to LinkedIn, mining best customers' enrichment and recommendations for ICP patterns, and batch-replying to the inbox through a shared review artifact that then feeds OUTREACH-STYLE.md.

Mandatory live endpoint discovery

After the first-use onboarding gate is satisfied, run wannalead endpoints and scan the complete current catalog before any business API call. Search the relevant domain with wannalead routes <filter>. For every candidate endpoint, run both:

wannalead endpoints <name> --examples --json-schema
wannalead openapi path <exact-/v1-route>

Read the complete returned contract before building a request: method, path, parameters, body schema, examples, responses, permissions, write guards, idempotency, pagination, deprecation, and replacement metadata. The live OpenAPI contract overrides examples or remembered endpoint shapes in any skill. Never claim that a capability is unavailable or impossible until the full catalog has been searched using the user's terms and relevant synonyms, and every plausible endpoint and replacement has been inspected. If nothing matches, state which catalog searches and candidate endpoints were checked.

Setup

Start with wannalead me. Authenticate only if that command reports missing authentication, then retry it. Run endpoint discovery only after onboarding is checked.

wannalead me
# Only when the previous command reports missing authentication:
wannalead auth login
wannalead me
# Only after onboarding is checked:
wannalead endpoints

If the API returns linkedin_connection_required, run:

wannalead auth connect

Wait until wannalead me reports the workspace connection as ready, then retry the exact read once. A linkedin_connection_required response creates no job and consumes no product unit. For a guarded write, preserve the original body and idempotency key instead of constructing a replacement request.

For a new account, or when durable commercial context is missing, use the wannalead-onboarding skill before prospecting. It builds the user's persistent local ME.md, COMPANY.md, ICPS.md, and OUTREACH-STYLE.md files through a resumable interview tracked in ONBOARDING.md.

For later prospecting, audience, account-mapping, campaign, or outreach tasks, use the progressive context rules in wannalead-onboarding: load only the relevant durable files, read current operational state from its authoritative system, and persist a durable change only after the user confirms it. Do not store CRM records, prospects, connections, conversations, campaigns, or run logs in the local Markdown workspace.

MCP clients can launch command wannalead with args ["mcp"]. The MCP and CLI expose the same public API capabilities.

Core workflow

wannalead endpoints companies.search
wannalead routes events
wannalead openapi path /v1/linkedin/events/register
wannalead call companies.search -body '{"keywords":"legaltech","company_size":["51-200"]}'

Prefer dedicated commands when available:

wannalead profiles enrich -profiles alice,bob --poll
wannalead profile-resolve -public-identifier alice

Use wannalead call <endpoint-name-or-path> for catalogued product capabilities. Use wannalead create -endpoint-key ... only as a last-resort escape hatch when no product alias exists.

For stored LinkedIn conversation history, use the named mirror reads:

wannalead call mirror.conversations -param limit=100
wannalead call mirror.messages \
  -param conversation_id=<conversation-id> \
  -param limit=100

The API key selects the associated LinkedIn account. Do not ask the user for an account identifier or expose local configuration, raw HTTP commands, refresh cycles, polling, or synchronization internals. Return the stored business result directly. Inspect mirror.status only when missing or incomplete data would materially change the answer, and describe that limitation without narrating the backend workflow.

For a multi-step LinkedIn campaign with profile visits, delayed invitations, acceptance-aware messages, per-recipient copy approval, and lifecycle tracking, use the installed campaign-automation skill.

For agent chaining, prefer the stable identifiers already projected by the API:

messages.conversations returns exact conversation URNs plus participants with profileAcoaa/profileUrn/publicIdentifier when exposed.
messages.thread and messages.seen-receipts retain sender/actor ACoAA.
events.attendees retains profileAcoaa for direct profiles.enrich calls.
me.notifications returns recent inbound cards with post/profile/company identifiers for follow-up reads.

Reuse profile identifiers exactly as returned. Prefer profileAcoaa or a profile URN for enrichment; otherwise preserve the exact Unicode or percent-encoded /in/{publicIdentifier} value. Never construct a slug from a display name or strip its accents: LinkedIn can return 403 for the ASCII form while the canonical Unicode profile remains accessible.

Response handling

Job creation can return completed data or job_id / job_ids.
Use --poll for local polling (five-minute default deadline) or wannalead job -id <job_id>.
Use -wait -timeout-ms <n> for bounded API waits, at most 25000 ms.
Keep -client-timeout-ms above the bounded API wait.
A poll deadline with a queued/running job is a pending result, not a failed LinkedIn call. Keep the same job_id; do not create a duplicate job.
Re-check a pending job at most once per advised delay (Retry-After, poll_after_ms, or the MCP retry_after_seconds). Never re-check a job that is already terminal, and never script or schedule a loop over job ids: a finished job never changes.
A confirmed invitations.send can return 202 with status invitation_profile_resolve_pending or relationship_preflight_pending before any invitation job exists. Poll preflight_job_id, then retry the exact request with the same idempotency key after it succeeds. The dedicated CLI and wannalead call invitations.send do this automatically when --poll is set. The key is reserved for that canonical target and message while pending. Stale relationship evidence is refreshed under that same key. Internal prerequisite reads do not consume product or LinkedIn account quota; quota starts only if the invitation job is created.
Invitations are capped per LinkedIn account at 50 a UTC day and 200 a week, shared between direct calls and campaigns. A newly connected account ramps up first: 20 on the UTC day it was connected to Wannalead, then 30, 35, 40, and 50 from its fifth day. The ramp follows the calendar, so idle days still count. The workspace dailyInvitationLimit is the 50 ceiling, not the account's budget for today. An invitation over the budget is deferred to the next UTC day (account_budget), not failed: tell the user, do not retry.
An invitation job with terminal status failed never changes. Reusing its Idempotency-Key returns 409 invitation_attempt_failed with the original job, not a new send. Inspect data.error for the cause. For a confirmed LINKEDIN_RATE_LIMITED response, stop; LinkedIn does not give a reliable reset time. After any Wannalead account cooldown, a separately approved new attempt needs a new key and a fresh relationship check. Never treat HTTP 200 alone as proof an invitation was sent; require data.status=succeeded.
LINKEDIN_TARGET_UNAVAILABLE is scoped to that read target. If the input was manually constructed or ASCII-normalized, retry once with the exact source profileAcoaa, profile URN, or profile URL. Otherwise skip or report it without treating the connected account as unavailable. Never brute-force possible accents.
Browser-backed read success is trusted for 24 hours while the same account binding and healthy extension session remain active.
Stop on rate-limit, restriction, authentication, or connection errors; do not retry-loop.
A CLI command using --poll exits non-zero when the terminal job is failed, queue_failed, parse_failed, timed_out, or canceled, after printing the final job response. Inspect both the process exit code and the returned status.
Stdout is agent-safe, not identifier-free: URNs and ACoAA values may remain when they are useful for chaining calls. Provider-domain URLs, credentials, browser routing, and unprocessed response bodies must not appear.

Daily agent journals

Create one stable run, then append stable events. Replays return the same run/event; reusing a key with a different body returns 409.

wannalead agent-runs start \
  -client-run-id "daily-outreach:2026-09-01" \
  -channel daily-outreach \
  -objective "Process the daily prospect queue" \
  -idempotency-key "agent-run:daily-outreach:2026-09-01"

wannalead agent-runs append \
  -run-id <run-id> \
  -client-event-id "target-12:send" \
  -event-type invitation.succeeded \
  -status succeeded \
  -job-id <job-id> \
  -payload '{"processed":1,"succeeded":1}' \
  -idempotency-key "agent-event:<run-id>:target-12:send"

Journal writes do not consume LinkedIn product units. The linked Wannalead jobs remain the source of truth for metering, quotas, execution, and polling.

Invitation campaigns

Campaigns remain extension-backed. Preparing a campaign does not authorize its LinkedIn writes; require explicit approval of the exact target set before launching, approving, or scheduling it.

For an invitation-only campaign, use invitations.send directly as the sequence root. The server reuses cached identity, resolves a missing identity as an internal prerequisite, and reruns profile.actions immediately before dispatch. First-degree, pending, or unavailable relationships terminate as non-billable skips.

Add visible profile.full_unary or profile.actions steps only when the user explicitly wants a profile visit, a delay anchored to that visit, or separate routing for existing first-degree connections. They are not invitation safety prerequisites.

To message only existing first-degree connections without creating invitation jobs, publish relationship.check as the sequence root and route relationship.first_degree to messages.send_profile. Leave relationship.not_connected without a branch so those targets complete without an invitation or message. Inspect the full request with wannalead endpoints automation.campaigns.sequences.create --examples --json-schema or use the campaign-automation skill.

The activation order is validate, launch, review, then schedule. launch selects and binds the workspace's healthy LinkedIn account; clients must not require a campaign binding before calling it. Reviews require the campaign to be running. More than 100 targets require the review confirmation token returned by the first call before scheduling.

Once scheduled, keep campaign and job ids durable. Poll existing state without recreating writes. The server owns extension-offline deferral, account-binding changes, pacing, quotas, cooldowns, retries, and target-level idempotency. After an invitation obtains its dispatch slot, the automation worker reruns profile.actions through the extension immediately before the write. Only a fresh relationship.not_connected result proceeds; first-degree, pending, and unavailable results terminate as non-billable skips. When the target has already sent the account an invitation (relationship.invitation_received), the worker accepts that invitation instead of inviting and the step finishes as relationship.first_degree, so the message branch continues. The same fallback runs inside every relationship.check step and before a messages.send_profile send.

Product feedback

When Wannalead itself returns an incorrect or incomplete result, has unclear endpoint guidance, or blocks a workflow reproducibly, preserve the endpoint name plus the response request_id and job_id when available. Explain the observed behavior, the expected behavior, and the impact without copying credentials, cookies, raw provider payloads, or unrelated private content.

Submitting feedback is an external write. Show the exact report to the user and obtain approval before sending it. Then use the catalogued meta endpoint:

wannalead endpoints feedback.send --examples --json-schema
wannalead call feedback.send \
  -body '{"message":"<observed behavior and expected behavior>","context":"endpoint=<name> request_id=<id> job_id=<id>"}'

feedback.send does not create a LinkedIn job or consume LinkedIn product quota. Do not send automatically, include secrets, or retry a rejected report in a loop. Reporting the issue does not replace returning the usable partial result or explaining the current blocker to the user.

Writes

Request write scope and validate with dry-run before sending:

wannalead auth login --write
wannalead invitations send \
  --dry-run \
  --profile-url "<profile-url>" \
  --message "Hello"
wannalead invitations send \
  --idempotency-key invite-001 \
  --profile-url "<profile-url>" \
  --message "Hello" \
  --poll \
  --yes

Post engagement uses the same guarded contract. Wannalead resolves canonical thread/root identifiers server-side; do not derive them or call the provider directly:

wannalead posts comment --activity-id <id> --text '<final-text>' --dry-run
wannalead posts reply --parent-comment-urn '<urn>' --text '<final-text>' --dry-run
wannalead posts repost --activity-id <id> --dry-run

After explicit approval, repeat the chosen command with a stable --idempotency-key and --yes. The write result may retain comment/repost URNs for chaining, but never echoes the submitted text.

Use the warm-outreach workflow for every invitation, DM, follow-up, comment, or reply. It requires the installed humanizer skill before approval, dry-run, and send. Send only the exact humanized text the user approved.

Every write requires explicit user intent. Preserve idempotency and never turn a guarded one-recipient workflow into bulk automation.

Boundaries

Never read LinkedIn cookies, browser profiles, local LinkedIn state, or gateway WebSocket internals. Never assemble private LinkedIn identifiers or call LinkedIn endpoints directly. Pass user-facing product inputs to Wannalead and let the API resolve execution details.

wannalead

Wannalead

Non-bypassable first-use gate