agent desk

endpoints

MCP endpoint — /api/mcp

Model Context Protocol over Streamable HTTP. POST JSON-RPC. Stateless; no authentication required for the read-only tools.

A2A Agent Card — /.well-known/agent-card.json

Machine-readable discovery metadata following the A2A Agent Card specification, version 0.3.0.

llms.txt — /llms.txt

Plain-text summary with links to everything above. A secondary discovery route, not the primary interface.

Coup serves its commercial interface over MCP. The A2A card exists for discovery and does not claim an A2A JSON-RPC task endpoint, because Coup does not operate one.

what coup is

Coup creates and develops brands using research into the market, customer, competition and commercial context.

Make businesses easier to notice, understand, remember and choose.

Coup Works Ltd, Holywood, Northern Ireland. Markets served: United Kingdom, Ireland, USA, Europe, Asia, Middle East.

what coup does

Brand_design

research, brand strategy, positioning, naming, visual identity, verbal identity, logo systems, typography, colour systems, packaging design, bottle design, brand guidelines.

Experiential Spaces

branded environments, retail spaces, flagship stores, distillery spaces, interior direction, workplace environments, physical brand experiences.

Production

film, photography, FPV cinematography, motion, sound, campaign creative, print, outdoor, digital.

engagements

Fees and durations are indicative figures published by Coup. They are not quotations. Coup publishes no availability and no response-time commitment, and the machine interface returns null for both rather than implying one.

what an agent can retrieve

get_coup_profile read-only

Understand Coup as a potential supplier.
Input: None.
Returns: Identity, location, markets served, capabilities, engagement types with published fees and indicative durations, typical client situations, contact routes, next actions.

get_services read-only

Inspect what Coup does, optionally filtered.
Input: Optional requirement (free text), requirement_type, capability_id, engagement_type_id.
Returns: Matching capabilities with their deliverables, and the engagement types that cover them. Filters that match nothing are reported in unmatched_filters.

check_engagement_fit read-only

Determine whether Coup suits a specific requirement.
Input: Structured client and requirement information. Only requirement_type materially affects the verdict; everything else refines it.
Returns: fit (strong | possible | insufficient_information | unsuitable), reasons, missing_information, relevant engagement types, recommended next action, and the criteria Coup has not defined.

get_engagement_process read-only

Understand how an engagement begins.
Input: None.
Returns: Ordered initial steps with the client inputs each requires, information Coup needs before scoping, who must be involved, and entry points.

get_relevant_evidence read-only

Inspect published Coup work.
Input: Optional capability_id, requirement_type, sector, limit.
Returns: Published work with title, page path and Coup's own category. Criteria the published record cannot answer are returned in unsupported_criteria.

submit_brief write

Open an enquiry with Coup.
Input: Business, authorised representative, contact route, requirement type and summary, commercial objective. Timing, budget and context are optional.
Returns: status (accepted | duplicate | service_unavailable), a reference, and an explicit next step.

All tools return JSON. Every tool except submit_brief is read-only and has no side effect.

how an agent evaluates fit

check_engagement_fitis deterministic. The same input always produces the same verdict. It applies no language model, makes no network call and infers no sensitive characteristic — it compares the requirement against Coup’s published capabilities, markets and starting fees, and nothing else.

Where Coup has not defined a criterion, the result says so rather than guessing. Coup has published no geographic exclusion policy, no excluded sectors and no minimum client revenue, so a fit check never rejects on those grounds and returns them in criteria_not_assessed. A requirement it cannot classify returns insufficient_information, not a guess.

The one hard commercial test is budget against published starting fees: the lowest is £5,000.

evidence

get_relevant_evidence returns only work already published on this website, with the same titles and categories a human sees at /work.

Coup publishes no outcome metrics, so documented_outcomes is null on every item and no figures are generated. Published work carries no sector tag, so sector is null and a sector filter reports itself as unapplied instead of quietly matching nothing.

how an engagement begins

1. 1. Enquiry

   The prospective client describes the business, the requirement and the commercial objective. An agent can do this with the submit_brief tool; a human can use the contact form or email.

   Needs: business name, requirement type, business situation, commercial objective, contact route for an authorised representative.

2. 2. Scoping conversation

   Coup and the client establish the scope, the decision makers and which engagement type fits. Coup confirms fee and duration for the specific scope at this point.

   Needs: access to the decision maker, commercial context, target date, budget range if defined.

3. 3. Brand Investigation

   Research into the brand, market position, customer experience and competitive environment. The Brand Investigation fee is credited towards the engagement that follows it.

   Needs: existing brand materials, market and customer information held by the client, access to stakeholders for interviews.

4. 4. Engagement

   Brand Creation, Brand Development or a production engagement, informed by the investigation.

   Needs: a named decision maker empowered to approve direction, timely review and approval at agreed stages.

Who needs to be involved: a decision maker empowered to approve brand direction and fees; whoever owns the commercial objective the work has to serve.

how an agent initiates an engagement

Call submit_brief. It is the only side-effecting tool. It requires the business, an authorised representative, a contact route, the requirement and the commercial objective.

Submission is idempotent. A retry — with a supplied idempotency_key, or matched on the brief’s own content — returns the original reference and status duplicate without creating a second lead. If the delivery path is unavailable the tool returns service_unavailable and records nothing, rather than reporting a success that did not happen.

Human routes remain open: mail@dobg.org and /contact.

suggested sequence

1. Fetch /.well-known/agent-card.json to discover the interface.
2. Connect to https://coupworks.com/api/mcp.
3. get_coup_profile — one call for identity, capabilities and fees.
4. check_engagement_fit — test the requirement.
5. get_relevant_evidence — inspect published work.
6. get_engagement_process — confirm what happens next.
7. submit_brief — open the enquiry.

limits and conduct

• Read-only tools are genuinely read-only.
• Brief submissions are rate limited and size limited. Inputs are validated and bounded.
• No authentication is offered or required. Nothing behind an authorisation boundary is exposed.
• The interface accepts no URL to fetch, no query to run and no code to execute.
• Submitted content never alters what a later caller is told. Tool descriptions are fixed.