find-pr-agency
Use whenever the user wants to find, shortlist, vet, or enrich US public-relations and communications agencies — media relations, crisis comms, investor relations (IR), product-launch PR, tech/startup PR, healthcare PR, B2B PR, public affairs, brand reputation, and internal commu
What it does
find-pr-agency
Drive the ServiceGraph API (https://api.servicegraph.co) to find,
shortlist, and enrich US public-relations and communications agencies.
Always pin service_provided:public-relations. Note: the
catalog's nominal industry:pr_comms value returns zero firms in
the live release — PR/comms firms are tagged with
service_provided:public-relations instead, typically under
adjacent industries (marketing_agency, other_pro_services). Pin the
service tag, not the industry. Sub-types (media relations, crisis,
IR, public affairs, healthcare PR, tech PR, B2B PR, internal comms,
brand reputation) are NOT separate tags — sub-type specialization is
a keyword substring search on firm text.
Any HTTP client works (curl, fetch, requests). Examples below use curl.
Sibling skills — defer when scope is broader
If the user wants a multi-service marketing engagement (PR plus
content plus paid plus social), defer to find-marketing-agency —
that skill covers full-service shops where PR is one of several
service lines.
This skill is correct when PR/comms is the primary deliverable — launches, media relations, crisis, IR, public affairs.
When NOT to use this skill
- "Write me a press release / draft these talking points" → DIY work.
- In-house comms/PR hires (Head of Comms, PR Manager).
- PR-software comparisons (Cision, Muck Rack, Prowly).
- Influencer-marketplace asks ("find me an influencer for our DTC product") — that's not PR; PR is earned media.
- "Explain how earned media works" → knowledge question.
- Non-US firms.
- Individual freelance PR people / publicists.
MCP server (preferred for authed calls)
If your agent harness has the ServiceGraph MCP server loaded
(https://mcp.servicegraph.co), prefer its tools for the authed
tier (/search, /get, /stats). The MCP server uses OAuth 2.1 +
PKCE — the host harness handles credentials in its own audited
sandbox, so there's no .env.local, no shell dispatch, and no token
value ever enters the LLM context.
For the anonymous tier (/tags, /check, /explore), MCP is
not preferred — every MCP tool requires OAuth (the server has no
anonymous tier), so plain curl against the REST URL is the simpler
path for discovery calls. Use the REST patterns below for those.
The MCP tools 1:1-map to the public REST endpoints — same backend, same quota, same data:
| MCP tool | REST endpoint | Anon? | Recommended path |
|---|---|---|---|
list_tags | GET /v1/tags | yes | curl |
check_filter | GET /v1/check | yes | curl |
explore_firms | GET /v1/explore | yes | curl |
search_firms | GET /v1/search | no | MCP if loaded, else curl + OTP |
get_firm | GET /v1/get/:id | no | MCP if loaded, else curl + OTP |
catalog_stats | GET /v1/stats | no | MCP if loaded, else curl + OTP |
Detection: if you see any MCP tools with servicegraph in the
name (the harness-specific prefix varies — agents pattern-match the
substring), the ServiceGraph MCP server is loaded. Prefer those
tools for the authed tier; complete any auth flow the harness
initiates if needed. If no servicegraph MCP tools are present,
fall through to the REST + OTP flow below for the authed tier.
The four-tier funnel
| Tier | Auth | Cost | Use it for |
|---|---|---|---|
GET /v1/tags | none | free | First call of every session. Discover legal field names, kinds, operators, values. |
GET /v1/check?filter=... | none | free | Validate a filter before spending an explore/search call. |
GET /v1/explore?filter=... | none | free, IP-throttled | Scope: count + breakdowns. Use to size the candidate pool before quota-spending. |
GET /v1/search?filter=... | bearer | 200 unique firms / month free | Brief firm cards. No url, no contact info. Use for ranking / shortlisting. |
GET /v1/get/:id | bearer | 50 unique firms / month free | Full bundle: url, phone, email, social, legal name, address. Only call for shortlisted firms. |
POST /v1/research | paid | not in MVP | Deferred — skip. |
Quota rule that matters: /search and /get charge per unique
firm viewed per calendar month, not per call. Re-paging the same
query is free. Two different filters that overlap charge once for
the overlap. Re-fetching a firm you already pulled this month is free.
Session-start ritual
Before constructing any filter, call:
GET https://api.servicegraph.co/v1/tags?include_values=1
Cache the response for the conversation. Confirm public-relations
is in the service_provided value list — that's the pin this skill
relies on. The nominal pr_comms industry value is empty in the
current release; don't pin it.
Field kinds you'll use most:
- tag_set_with_evidence:
service_provided(always includepublic-relations) — op:with optional@evidence - categorical:
state,pricing_model,company_size_signal,geography_served— op: - numeric:
rating,review_count_total,founded_year— ops= >= <= > < - presence:
has:phone,has:clutch,has:rating,has:linkedin_company, … - keyword: free-text substring across firm name / brand / title / meta / legal_name. This is how you specialize on sub-type and vertical (tech, healthcare, IR, crisis, public affairs, B2B, etc.).
Auth
/tags, /check, and /explore are anonymous. /search and /get
require a bearer token.
Security model — keep the token out of the LLM context.
- Never read
.env,.env.local, or any other credential file into your context. The token's literal value should never appear in the conversation. - Use shell dispatch for every authed request so the token flows
directly from the user's environment / dotenv file into the
Authorizationheader without round-tripping through the LLM. - Always ask the user once per session before using a detected
token, even if it's already in their shell or
.env.local.
Resolution rule:
-
Detect whether a token is available — without reading its value. Run a shell check that only inspects exit codes:
( [ -n "${SERVICEGRAPH_TOKEN:-}" ] \ || grep -qs '^SERVICEGRAPH_TOKEN=' .env.local \ || grep -qs '^SERVICEGRAPH_TOKEN=' .env )Exit code
0= token is available somewhere; non-zero = no token. -
Confirm with the user before the first authed call this session:
"I found a
SERVICEGRAPH_TOKENin your environment /.env.local. OK to use it for ServiceGraph API requests this session?"If the user says no, stay on the anonymous tiers (
/tags,/check,/explore) and skip authed calls. Don't re-ask later unless the user asks for authed work. -
Dispatch via shell — every authed call goes through a shell wrapper so the literal token never enters the conversation:
# If exported in the shell environment: curl -H "Authorization: Bearer $SERVICEGRAPH_TOKEN" \ 'https://api.servicegraph.co/v1/search?filter=...' # If in .env.local — source it inside a subshell so it doesn't # leak into the parent shell either: ( set -a; . ./.env.local; set +a; curl -H "Authorization: Bearer $SERVICEGRAPH_TOKEN" \ 'https://api.servicegraph.co/v1/search?filter=...' )Capture the response body to a tmp file or jq-process it, but do NOT echo the request command with the token expanded.
-
OTP flow if no token is detected — capture the new token directly into
.env.localwithout surfacing its value to the LLM:# 1. trigger the email — agent prompts the user for $EMAIL curl -fsS -X POST 'https://api.servicegraph.co/v1/auth/request-otp' \ -H 'Content-Type: application/json' \ -d "{\"email\":\"$EMAIL\"}" # 2. exchange the code — agent prompts the user for $CODE. # The ?format=env query param returns SERVICEGRAPH_TOKEN=<token> # as plain text appended to .env.local — no jq needed. The -f # flag makes curl exit non-zero on 4xx so a wrong code doesn't # pollute the file (the error mirror is also a `# comment` line, # safe to ignore even if it lands). curl -fsS -X POST 'https://api.servicegraph.co/v1/auth/verify-otp?format=env' \ -H 'Content-Type: application/json' \ -d "{\"email\":\"$EMAIL\",\"code\":\"$CODE\",\"name\":\"claude-cli\"}" \ >> .env.local # 3. confirm capture without revealing the value grep -q '^SERVICEGRAPH_TOKEN=' .env.local && echo "OTP token captured."After a successful capture, the user has implicitly consented (they just completed the flow), so proceed to dispatch (step 3). The token is now persistent in
.env.localfor future sessions. -
If a
/searchor/getreturns401 unauthorizedmid-session, the token expired or was revoked — re-run the OTP flow.
Filter DSL
One query parameter, GitHub-search-style.
filter := orExpr
orExpr := andExpr ("OR" andExpr)*
andExpr := notExpr (("AND")? notExpr)* # whitespace = implicit AND
notExpr := ("NOT" | "-") notExpr | atom
atom := "(" filter ")" | predicate
predicate:= IDENT op valueOrList | bareword
op := ":" | "=" | ">=" | "<=" | ">" | "<"
valueOrList := value ("," value)*
value := IDENT | NUMBER | tagAtEvidence
tagAtEvidence := IDENT "@" ("low"|"medium"|"high")
bareword := IDENT | NUMBER # → keyword:<bareword>
Four rules that bite:
- AND binds tighter than OR.
a OR b cparses asa OR (b AND c). Use parens. - Comma list = OR within one predicate.
state:CA,NY,TXmatches any of the three. - Negation is
-xorNOT x. Negative literals inside a comma list are not allowed:state:CA,-NYis rejected. Usestate:CA -state:NY. - Bareword = keyword search. Any IDENT or NUMBER not followed by an operator becomes a free-text substring across name / brand / title / meta / legal_name. Multiple barewords AND.
PR-flavored examples (validate yours with /v1/check):
service_provided:public-relations tech state:NY
service_provided:public-relations healthcare
service_provided:public-relations b2b saas
service_provided:public-relations crisis
service_provided:public-relations investor relations
service_provided:public-relations ipo state:NY,CA
service_provided:public-relations public affairs state:DC
service_provided:public-relations rating>=4 has:clutch
When in doubt, hit /v1/check?filter=... first.
Sub-type / vertical → keyword mapping:
| User asks for | Add as keyword(s) |
|---|---|
| Tech / startup PR | tech, startup, saas |
| Healthcare / pharma PR | healthcare, pharma, biotech |
| Crisis comms | crisis |
| Investor relations / IR | investor relations, ir, ipo |
| B2B PR | b2b |
| Public affairs | public affairs |
| Brand reputation | brand reputation, reputation |
| Internal communications | internal communications, internal comms |
| Earned media / media relations | earned media, media relations |
firm_id contract
firm_id is a stable 12-hex-char handle:
firm_id = sha256(apex.lower().rstrip(".")).hexdigest()[:12]
import hashlib
def firm_id(apex):
return hashlib.sha256(apex.lower().rstrip(".").encode()).hexdigest()[:12]
echo -n "edelman.com" | tr 'A-Z' 'a-z' \
| openssl dgst -sha256 -hex | awk '{print substr($2,1,12)}'
Recipes
A. Tech PR for a Series-B announcement
User: "Tech PR agency in NY for our Series-B announcement."
GET /v1/explore?filter=service_provided:public-relations+tech+state:NY
GET /v1/search?filter=service_provided:public-relations+tech+state:NY&limit=10
GET /v1/get/<firm_id> # ×3
B. IR firms for an IPO
User: "Three IR firms for our upcoming IPO roadshow."
GET /v1/search?filter=service_provided:public-relations+(investor relations OR ir)+ipo
C. Crisis comms (urgent)
User: "Crisis comms help — brand reputation issue blowing up online."
GET /v1/search?filter=service_provided:public-relations+crisis&limit=10&order_by=relevance
Surface as urgent: skip /v1/explore, jump straight to /v1/search,
present briefs immediately.
D. Healthcare / regulatory PR
User: "Healthcare PR agency familiar with FDA regulatory comms."
GET /v1/search?filter=service_provided:public-relations+healthcare+(fda OR regulatory)
E. Indirect intent — "we need press"
User: "We need press for our Series-B — get us into TechCrunch, WSJ, and the trade press."
That's product-launch / tech PR. Translate:
GET /v1/search?filter=service_provided:public-relations+(tech OR startup)+(launch OR series-b)&limit=10
If thin, drop the launch/series-b keyword — most tech PR firms do launches as a default.
F. Public affairs (state government)
User: "Public affairs firms with state-government experience in California."
GET /v1/search?filter=service_provided:public-relations+public affairs+state:CA
G. BYO apex list — enrich domains
User pastes 8–20 PR-firm domains. For each:
- Compute
firm_idlocally. GET /v1/get/<firm_id>— full bundle if in catalog, 404 if not.- Aggregate, present, flag the not-found ones.
Gotchas
- Always pin
service_provided:public-relations. Don't pinindustry:pr_comms— it's empty in the live catalog. PR firms are tagged with the service tag across adjacent industries (marketing_agency, other_pro_services). Without the service pin, "tech PR" / "crisis" / "investor relations" as keywords leak into general marketing. - Defer to
find-marketing-agencyfor full-service marketing engagements. When PR is one of several service lines (PR + content + paid + social), the marketing-agency skill is the right fire. - Sub-types are keyword-only. Multi-word sub-types split into ANDed barewords (
investor relations→investorANDrelations). - "Write me a press release" / "draft talking points" is do-the-work. Refuse and offer to find a firm if the user wants to engage one.
- Influencer marketing is NOT PR. Influencer marketplaces (Aspire, Grin, etc.) and influencer outreach are paid-media work; defer to
find-marketing-agencyor refuse for marketplace questions. - PR-software comparisons (Cision, Muck Rack, Prowly) are NOT procurement.
looks_not_pro_services404 is not a bug. Afirm_idmay exist in/searchbut 404 on/getif it's been flagged. Skip and continue; not charged./v1/explorek=20 suppression. When fewer than 20 firms match, the response is{"count": "<20", "suppressed": true, "breakdowns": {}}. Drilling further makes the count smaller. Broaden or escalate to/v1/search.- Briefs from
/searchdo NOT includeapex,url,phone_primary,email_primary,legal_name, or address. If the user asks for contact info, you must/get/:id. - Quota is per-user-per-month, deduped on first view. Re-views are free; re-pagination is free.
Errors
All errors return JSON: {"error": {"code": "...", "message": "..."}}.
| Status | Code | What to do |
|---|---|---|
| 400 | filter_parse_error | Payload includes position. Fix the filter, re-validate with /v1/check. |
| 400 | filter_required | Empty filter where one is required. |
| 400 | invalid_firm_id | firm_id must be 12 lowercase hex chars. Re-derive. |
| 401 | unauthorized | Token missing/expired. Re-run OTP. |
| 404 | not_found | Firm not in catalog or flagged. Not charged. Skip and continue. |
| 429 | rate_limited | Honor Retry-After header / retry_after field. |
| 429 | monthly_quota_exhausted | Switch to /v1/explore-only mode for the rest of the month. Tell the user. |
End-to-end example
User: "Three tech PR agencies in NY for a Series-B announcement, ideally with 4-star ratings and Fortune 500 client experience."
GET /v1/tags?include_values=1
GET /v1/check?filter=service_provided:public-relations+tech+state:NY+rating>=4
GET /v1/explore?filter=service_provided:public-relations+tech+state:NY+rating>=4
GET /v1/search?filter=...&limit=10
GET /v1/get/<firm_id> # ×3
End of session: report X-Quota-Remaining-Month.
Capabilities
Install
Quality
deterministic score 0.53 from registry signals: · indexed on github topic:agent-skills · 160 github stars · SKILL.md body (16,159 chars)