Liam

Liam is an ad manager for LinkedIn (LinkedIn Ad Manager). Create campaigns by talking to Claude (MCP) or from a CLI. Built for go-to-market teams who want to spin up many campaigns from a contact list and a brief, then add creative images themselves. Everything is created as a draft, so nothing spends until you explicitly activate it in Campaign Manager.

Unofficial and not affiliated with or endorsed by LinkedIn.

Covers creation, matched audiences (including building one straight from Salesforce), conversion selection, performance reporting/insights, competitor ad intelligence, and a change journal with before/after lift. See ROADMAP.md for what's shipped and what's planned next.

What you can do with it

Once Liam is connected, you talk to your assistant in plain language. Real requests it handles today, roughly in the order a campaign comes together:

• "How many VPs of demand gen at US SaaS companies can we reach?" (resolves targeting facets, estimates reach before any money is involved)

• "Upload this CSV as a matched audience." (auto-cleans columns, hashes emails, converts company domains to URLs)

• "Build an audience from Salesforce: every contact on an account flagged as a target." (SOQL straight to a matched audience)

• "Launch a draft campaign for that audience, $100/day, tracking the demo-booked conversion." (campaign group + campaign + draft ads in one call)

• "What ads is HubSpot running right now, and what offers are they pushing?" (reads the public Ad Library: copy, formats, and EU targeting metadata for any company)

• "How did retargeting do last month vs the month before?" (trend report with deltas)

• "Which ads should I pause this week?" (account rollup with top/bottom performers and flags)

• "Rewrite the copy on the losing ad." (delete + recreate as a fresh draft; LinkedIn ignores edits to a live ad's post)

• "Did the new headline actually help?" (change journal + before/after lift)

Liam resolves the details and reports back the ids. Every capability is also a liam CLI command for scripted or batch use.

Put it on a loop

There is no scheduler in Liam. Use the one in whatever client you drive it from. Some loops that work well:

• "Every Monday at 9am, pull last week's performance summary, compare it to the week before, and flag anything that moved more than 20%." (a Claude Code scheduled agent)

• "Every Friday, check what new ads our top three competitors shipped this week and summarize the angles."

• "Check the audience match status every few hours and tell me when it clears 300 members." (the minimum for a matched audience to serve)

• Or skip the assistant entirely and put the CLI in cron: 0 9 * * 1 liam report summary -p last_7_days

Related MCP server: linkedin-campaign-manager-mcp

Hierarchy mapping (LinkedIn differs from Google/Meta)

Packages

• @liads/core: LinkedIn REST client, OAuth, resource modules, CSV + SHA256 hashing, Salesforce reader.

• @liads/mcp: MCP server (stdio for local; reused by the hosted app). Primary interface.

• @liads/cli: the liam CLI over the same core, for scripted batch runs.

• @liads/web: Next.js app that hosts the MCP over HTTP on Vercel.

How it works

Both interfaces are thin layers over the same engine. Every MCP tool and CLI command builds a client via createLiads() (core/src/client.ts), which loads config plus an auto-refreshing OAuth token provider, then calls a typed resource module (campaigns.ts, audience.ts, report.ts, ...) that wraps LinkedIn's versioned REST API (LinkedIn-Version pinned in config). Input shapes live as zod schemas in core/src/schemas.ts and are reused verbatim as the MCP tool input schemas, so the CLI and MCP can never drift apart.

The draft-only safety rule is enforced at creation: campaign groups, campaigns, and creatives are all written with DRAFT status, and there is deliberately no "activate" tool or command. Every write also passes through one HTTP chokepoint that appends to a local change journal (~/.liads/changelog.jsonl) for later lift analysis.

AGENTS.md has the full architecture notes and the LinkedIn API gotchas (restli encoding, required fields, DMP audience flow) learned from live testing.

Install

There are two ways to install Liam:

1. As an MCP server, registered with Claude Code, Claude Desktop, Cursor, or any MCP client, so you create campaigns by talking to your assistant. Run it locally, or skip the always-on install entirely and connect to the hosted endpoint with your own credentials.

2. As a terminal CLI, a global liam command for running everything from your terminal, scripts, or cron.

Both run against your own LinkedIn developer app, so your credentials and your ad account stay yours. The first two steps are shared: create a LinkedIn app (step 0) and build + authenticate (step 1). After that, pick your path, or do both; they read the same ~/.liads credentials.

Step 0: create a LinkedIn app (once)

1. Create an app at https://www.linkedin.com/developers/apps. It must be associated with your company's LinkedIn Page.

2. On the Products tab, request access:

   Product	Needed for	Required?
   Advertising API	campaigns, ads, targeting, reporting, conversions	Yes
   Audiences	uploading CSV contact/company lists as matched audiences	Only for audience upload
   LinkedIn Ad Library	competitor ad metadata via the official API	Optional (the browser scraper works without it)

3. On the Auth tab, add http://localhost:53682/callback as an authorized redirect URL, and note your Client ID and Client Secret.

4. Once the Advertising API is approved, map your ad account under Products → Advertising API → View Ad Accounts. Skip this and accounts list comes back empty.

OAuth scopes. liam auth login requests rw_ads (create and edit campaigns), r_ads_reporting (reporting), rw_conversions (conversion tracking), and w_organization_social (image ads create a post owned by your LinkedIn Page). If you add a product or scope later, run auth login again; a token refresh keeps only the scopes you originally granted.

Step 1: build and authenticate (once, both paths)

Requires git, Node.js 20+, and pnpm.

git clone https://github.com/stan-default/liam.git
cd liam
pnpm install && pnpm -r build

# App credentials (alternatively set LIADS_CLIENT_ID / LIADS_CLIENT_SECRET env vars)
mkdir -p ~/.liads
echo '{ "clientId": "...", "clientSecret": "...", "linkedinVersion": "202605" }' > ~/.liads/config.json

node packages/cli/dist/index.js auth login      # opens the browser; tokens land in ~/.liads
node packages/cli/dist/index.js accounts list   # verify, then set defaultAccountId in config.json

Option 1: install as an MCP server

Register the server with your MCP client; it reads the ~/.liads credentials from step 1.

Claude Code

claude mcp add liam -- node /abs/path/liam/packages/mcp/dist/index.js

Claude Desktop, Cursor, or any MCP client, in its MCP config JSON:

{
  "mcpServers": {
    "liam": { "command": "node", "args": ["/abs/path/liam/packages/mcp/dist/index.js"] }
  }
}

Ask for something read-only to confirm it works: "List my ad accounts" or "How did my account do in the last 30 days?"

Hosted MCP: connect with your own credentials, nothing to deploy

The hosted MCP server at

https://liam-mcp.vercel.app/api/mcp

is multi-tenant, bring-your-own-credentials: you pass your LinkedIn developer app's details as request headers, and every call runs against your app and your ad account. The server holds no state for you: credentials are used in memory to call LinkedIn and never logged or persisted.

Do step 0 and step 1 above once (the LinkedIn app and the local auth login; LinkedIn's OAuth consent has to happen in your browser, so the token mint is the one local step). Then print the ready-to-run connect command:

node packages/cli/dist/index.js auth export --mcp
# claude mcp add --transport http liam https://liam-…vercel.app/api/mcp \
#   --header "X-Liads-Client-Id: …" \
#   --header "X-Liads-Client-Secret: …" \
#   --header "X-Liads-Refresh-Token: …" \
#   --header "X-Liads-Account-Id: …"

Any MCP client that supports custom headers works the same way (for clients without native HTTP transport: npx mcp-remote <url> --header "X-Liads-Client-Id: …" …). Verify with a read-only call ("list my ad accounts"), and from then on the connection is just a URL plus headers, usable from machines that never cloned this repo.

Know the trade-off: your client secret and refresh token travel with every request to that server, so only point them at a deployment you trust, or self-host the identical endpoint (next section). The scopes they carry can manage ads but never activate them; the draft-only rule is enforced in the tool layer itself.

Hosted limitations (all by design, they need local resources): upload_audience_csv reads a CSV path on the server, audience_from_salesforce needs your local sf CLI, competitor-ads scraping falls back to API-only metadata (no browser), and the change journal / lift tools need the local ~/.liads/changelog.jsonl. Run those through the local MCP server or CLI.

Put a skill on top. The clean split for teams: your MCP client's config holds the credentials (the headers above), and a small skill or system prompt holds your playbook: default ad account, naming conventions, standing exclusions, house rules like "audience expansion always off". The skills/ folder in this repo is a working example; copy one, swap in your defaults, and your assistant drives the hosted tools with your rules. Credentials never belong in a skill file.

Self-host the same endpoint on Vercel

Run your own instance so credentials never touch shared infrastructure, and so you also get a private env-configured tenant:

1. Get your env values locally: node packages/cli/dist/index.js auth export.

2. Deploy apps/web to Vercel (set the project Root Directory to apps/web).

3. In Vercel project settings, add the env vars from .env.example (LIADS_CLIENT_ID, LIADS_CLIENT_SECRET, LIADS_REFRESH_TOKEN, LIADS_LINKEDIN_VERSION, and a strong MCP_AUTH_TOKEN). In Vercel's Deployment Protection settings, disable Vercel Authentication for production (or add a protection-bypass secret); otherwise MCP clients can't reach the endpoint.

4. Your MCP endpoint is https://<your-app>.vercel.app/api/mcp. Connect with the secret in the header:

claude mcp add --transport http liam https://<your-app>.vercel.app/api/mcp \
  --header "Authorization: Bearer <MCP_AUTH_TOKEN>"

Requests carrying MCP_AUTH_TOKEN use the env credentials (your tenant). Requests carrying the X-Liads-* headers above use the caller's credentials instead, so one self-hosted instance can serve your whole team, each member on their own LinkedIn app.

Option 2: install the terminal CLI

Step 1 already left a working CLI at node packages/cli/dist/index.js. To get a global liam command instead of the long node path:

cd packages/cli && pnpm link --global
liam accounts list

(No pnpm link? An alias works too: alias liam="node /abs/path/liam/packages/cli/dist/index.js".)

A sensible first session:

liam targeting search titles "demand generation"   # resolve targeting URNs
liam report summary -p last_30_days                # account rollup + flags
liam launch --brief examples/brief.json            # creates DRAFTS only

The full CLI reference is below.

Or paste this prompt and let your assistant do the setup

Copy the whole block into Claude (or any capable assistant with terminal access) and it will walk you through everything above, one step at a time:

You are helping me set up Liam, an open-source LinkedIn Ad Manager from
https://github.com/stan-default/liam. It runs locally as an MCP server and a CLI.
Everything it creates on LinkedIn is a draft, so nothing spends money until I
activate it myself in Campaign Manager.

Walk me through the setup one step at a time. Run commands for me where you can,
show me exactly what to click where you cannot, and wait for my confirmation
before moving to the next step.

1. Check that git, Node.js 20 or newer, and pnpm are installed. Help me install
   whatever is missing.
2. Clone https://github.com/stan-default/liam and run "pnpm install" then
   "pnpm -r build" in the repo root.
3. Help me create a LinkedIn developer app at
   https://www.linkedin.com/developers/apps (it must be associated with my
   company's LinkedIn Page):
   a. On the Products tab, request access to the "Advertising API" product.
      This is the only required product; approval can take a while.
   b. Optional: also request "Audiences" if I want to upload CSV contact or
      company lists as matched audiences.
   c. Optional: also request "LinkedIn Ad Library" if I want competitor ad
      metadata through the official API.
   d. On the Auth tab, add http://localhost:53682/callback as an authorized
      redirect URL, and show me where the Client ID and Client Secret live.
4. Create ~/.liads/config.json containing my clientId, clientSecret, and
   "linkedinVersion": "202605".
5. Run "node packages/cli/dist/index.js auth login". My browser opens LinkedIn's
   consent screen; after I approve, tokens are stored in ~/.liads/credentials.json.
6. Verify with "node packages/cli/dist/index.js accounts list". If no accounts
   show up, remind me to map my ad account in the Developer Portal under
   Products > Advertising API > View Ad Accounts. Then set my account id as
   "defaultAccountId" in ~/.liads/config.json.
7. Register the MCP server with the assistant I use:
   - Claude Code: claude mcp add liam -- node <abs repo path>/packages/mcp/dist/index.js
   - Claude Desktop or another client: add {"command": "node", "args":
     ["<abs repo path>/packages/mcp/dist/index.js"]} under mcpServers in its config.
8. Confirm the tools load, then show me one read-only call working, for example
   list_ad_accounts or a last_30_days performance summary.

If any step fails, show me the exact error and fix it with me before moving on.

Skills: playbooks on top of Liam

The skills/ directory ships ten portable agent skills that turn Liam's raw tools into opinionated workflows. Each encodes a methodology (what to pull, how to judge it, significance floors, caveats, a fixed report format) and works over MCP or the CLI. Analysis and monitoring skills are read-only; the operating skills create drafts only, after confirmation.

Analyze:

• liam-spend, "analyze my spend": budget concentration, wasted spend, reallocation moves

• liam-performance, "how are my ads doing": winners, losers, fatigue, actions

• liam-leads, "which ads drive leads, which don't": drivers, pause list, angle patterns

• liam-competitors, "what is <company> running": Ad Library teardown and gaps

Monitor:

• liam-weekly, "weekly snapshot": fixed-format week-over-week digest, schedule-friendly

• liam-health, "anything on fire": silent-unless-fire daily guardrails, one line when all clear

Operate:

• liam-launch, "launch a campaign": brief to confirmed draft with house guardrails applied

• liam-experiments, "did that change help": hypothesis logging plus before/after lift

• liam-audiences, "audit my audiences": health, dry-run uploads, Salesforce refresh loop

• liam-account-audit, "is my account set up right": hygiene scorecard and fix list

./skills/install.sh   # symlinks them into ~/.claude/skills (Claude Code personal skills)

See skills/README.md for details and conventions.

Tools (MCP)

• Accounts: list_ad_accounts

• Targeting: list_targeting_facets, search_targeting (typeahead a facet for entity URNs), list_facet_entities, estimate_audience (structured spec). Talk in plain language ("VPs of demand gen at SaaS companies in the US") and Liam resolves the facets and estimates reach before building the campaign.

• Audiences: upload_audience_csv (auto-cleans the CSV: normalizes column names, drops non-matcher columns, hashes emails, converts company domains to website URLs; supports both contact and company lists), audience_from_salesforce (SOQL to matched audience), get_audience_status

• Conversions: list_conversions (select an existing insight-tag conversion to track). create_campaign and launch_from_brief accept conversionIds or conversionName, and fall back to defaultConversionName from config.

• Campaigns: create_campaign_group, create_campaign, create_text_ad, create_image_ad; list_campaigns (the account structure: campaign groups and their ad groups, drafts included, unlike reporting which only shows entities with spend), list_ads, delete_ad (removes the creative and best-effort its Direct Sponsored Content post). LinkedIn's editor ignores post edits on a live ad, so to change copy you recreate: delete_ad + create_image_ad.

• Orchestrator: launch_from_brief (audience + group + campaign + draft creatives in one call)

• Reporting: performance_summary (account rollup + top/bottom + flags), get_performance (per-entity KPIs at any level), performance_trend (weekly/monthly with deltas). KPIs: CTR, CPC, CPM, CPL, conversion rate, cost per conversion. Levels: campaign_group, campaign, creative.

• Competitor intel: inspect_competitor_ads reads any company's ads from the LinkedIn Ad Library (no ad-account access needed). The official Ad Library API (GET /rest/adLibrary, requires the "LinkedIn Ad Library" product grant) returns structured metadata (advertiser, payer, format, and for EU-served ads run dates, impression ranges, per-country split, and targeting facets) but no creative. A Playwright scraper of the public library supplies the ad copy/image. Engines (engine): api (metadata only, fast, works hosted), scraper (copy via browser, local, supports company-id), and auto (default: API metadata plus copy layered from each ad's detail page, falling back to the scraper if the API isn't provisioned). Search by advertiser name or keyword (the API has no company-id or date filter). CLI: liam competitor ads.

• Change journal & lift: log_ad_change (record a change), list_ad_changes, compute_lift (before-vs-after performance for each recorded change). Liam auto-journals every change it makes; see Change journal & lift below.

Targeting spec

Structured targeting uses short facet names mapped to entity URNs (resolve URNs with search_targeting). URNs within a facet are ORed; facets are ANDed; excluded facets are ORed.

{ "include": { "locations": ["urn:li:geo:103644278"], "seniorities": ["urn:li:seniority:7"], "titles": ["urn:li:title:26587"] },
  "exclude": { "industries": ["urn:li:industry:47"] } }

CLI reference

All commands also work via node packages/cli/dist/index.js <cmd>.

liam auth login                         # OAuth, stores tokens in ~/.liads
liam auth export                        # print env vars for a self-hosted (Vercel) server
liam auth export --mcp                  # print the hosted-MCP connect command with your headers
liam accounts list                      # list accessible ad accounts
liam targeting search <facet> <query>   # typeahead a facet for entity URNs
liam targeting estimate <facet> <urns…> # audience size for one facet's URNs
liam audience upload -n <name> -f <csv> # clean + upload a CSV as a matched audience
liam audience upload -n <name> -f <csv> --dry-run   # preview the cleaned CSV, no upload
liam audience from-salesforce -n <name> -q "<SOQL>"   # Salesforce query -> matched audience
liam audience status <segmentId>        # matching status + resolved size
liam conversions list                   # account conversions (pick one to track)
liam campaigns list [--group <id>] [--all]   # campaign groups + ad groups, drafts included
liam ad list <campaignId>               # ads in an ad group (id, status, name), drafts included
liam ad delete <creativeId>             # delete an ad (creative + its DSC post)
liam report summary [-p <period>]       # account rollup: totals, top performers, flags
liam report perf <level> [--parent <id>] # per-entity KPI rows
liam report trend <level> <id> [-b weekly|monthly]  # trend with deltas
liam launch --brief <brief.json>        # audience + group + campaign + draft creatives
liam competitor ads <advertiser>        # any company's ads from the public Ad Library
                                        #   <advertiser> = name, company id, or company URL
                                        #   -k <keyword> -c <countries> -e auto|api|scraper -m <max> --json
liam changelog list [-t <type>] [-i <id>]           # recorded ad changes, newest first
liam changelog add -t <type> -i <id> -f <field> --after <v> [-l <label>]   # log a change made elsewhere
liam lift <level> <id> [-w <days>]      # before-vs-after performance for each recorded change

Periods: last_7_days, last_30_days, last_90_days, month_to_date, last_month.

--account defaults to defaultAccountId from config where applicable.

Change journal & lift

Liam keeps an append-only journal of every change made to an ad entity so you can measure the lift of a change: how performance differed in the window before it versus after.

• Where it lives: ~/.liads/changelog.jsonl (one JSON object per line). A plain local file with nothing to set up. Override the path with LIADS_CHANGELOG_PATH.

• Auto-capture: every create/update Liam makes to a campaign group, campaign, or creative is journaled automatically (captured at the one HTTP chokepoint, so it can't be forgotten).

• Manual entries: log changes made outside Liam (e.g. in Campaign Manager), or attach a hypothesis, with liam changelog add / the log_ad_change tool. Use -l/--label to name the test (e.g. "outcome-led headline test") and --tags to group related changes.

• Lift: liam lift <level> <entityId> (or compute_lift) reads the journal for that entity and, for each change, compares the --window days before (default 14) against the same window after, reporting before/after KPIs and per-metric deltas (CTR, CPC, conversion rate, cost-per-conversion, …). Recent changes get a clamped, partial after-window.

This is a directional pre/post comparison, not a controlled experiment. It is confounded by seasonality, the LinkedIn learning phase after an edit, and any concurrent budget change, so treat the deltas as a signal. Set LIADS_NO_CHANGELOG=1 to disable auto-capture; journaling is also off on hosted deploys (read-only filesystem).

Configuration

Local config lives in ~/.liads/config.json (mode 0600, never in the repo):

Hosted equivalents are the LIADS_* env vars plus MCP_AUTH_TOKEN (see .env.example). OAuth tokens are stored separately in ~/.liads/credentials.json.

Salesforce integration

Liam reads Salesforce by shelling out to the authenticated sf CLI (sf data query), so it reuses your existing login and needs no new credentials. audience_from_salesforce (and liam audience from-salesforce) take a SOQL query that selects an email column and turn the result into a matched audience, closing the loop from "accounts flagged in Salesforce" to "LinkedIn targeting." Example:

liam audience from-salesforce -n "Q3 target accounts" \
  -q "SELECT Email FROM Contact WHERE Account.Target_List__c = true AND Email != null"

Safety

• Every campaign/creative is created DRAFT/PAUSED. Activation is a separate, explicit step.

• Matched-audience matching takes up to 48h, and a campaign needs ~300 matched members to serve.

• LinkedIn does not expose person-level ad-view data; cross-referencing is account/segment-level.

• Secrets never enter the repo. Local: ~/.liads. Hosted: Vercel env vars.

License

MIT