Prepare MCP Integration Agent Skill

Prepare MCP Integration

Take an input link or vendor brief through two separate phases: a reviewable

research PR in paperclip-content, then connector implementation in Paperclip

App only after a human accepts the exact research revision and connection set.

Preserve These Boundaries

• Treat paperclip-content/integrations/README.md as the research contract and

paperclip-content/integrations/skills/integration-harness/SKILL.md as its

entrypoint. Reference and run them; do not copy their schemas, templates,

state machines, reconciliation rules, or internal gates into this skill.

• Treat paperclip/doc/connections/CONNECTOR-PLAYBOOK.md on the implementation

target branch as the connector contract. Follow it end to end; do not

substitute remembered behavior or the examples in this skill.

• Finish Phase A with a research-only PR. Do not create an App implementation

branch, task, or code change before the research gate is accepted.

• Bind acceptance to one research PR head SHA and an explicit connection set.

Acceptance does not cover later commits or additional connections.

• Create one Paperclip App PR per approved connection. Shared prerequisite

infrastructure or broad playbook corrections may use separate prerequisite

PRs; never combine distinct connections into one connector PR.

• Keep vendor credentials in approved secret storage. Never put secrets in

briefs, catalog files, issue text, plans, fixtures, screenshots, logs, branch

names, commits, or PRs.

Default Connection Experience

• Request the broadest vendor permissions and scopes the connection can

support by default. Operators should not have to predict every future tool

they may need during setup. Enforce safe use after connection through

Paperclip's action catalog, resource boundaries, ask-first policies,

quarantine, and audit controls.

• Keep the default wizard limited to the minimum information needed to create

a working connection: connection identity, authentication, and any

unavoidable tenant or resource boundary. Put optional scope reduction,

feature groups, individual tool filters, response modes, transport tuning,

and other expert controls behind one collapsed Advanced disclosure.

• Give advanced controls working broad defaults so an operator can finish

setup without opening them. When a provider truly requires an explicit

advanced choice, document the exception and explain it in plain language

instead of exposing protocol details by default.

• Treat vendor permission breadth and Paperclip execution governance as

separate layers. Do not reduce requested vendor permissions merely to stand

in for missing action review, approval, quarantine, or audit policy.

Preflight

• Load the current Paperclip skill for checkout, comments, interactions,

durable state, and final disposition. Load the standard PR-preparation skill

(prepare-paperclip-pr for Paperclip agents) before opening any PR.

• Resolve the input URL(s), vendor/platform, intended MCP endpoint or API,

target repositories and branches, and the Paperclip issue that owns the

work. Ask only when these cannot be determined safely from the brief.

• Fetch both repositories and record the target commit hashes. Read the

canonical files from those target commits, not from a possibly stale working

tree. Refresh and reread them again immediately before Phase B.

• Inventory existing integration catalog entities, open PRs, branches, and

current AppDefinitions/connectors before creating anything. Match by

meaning, not title or slug.

• Determine whether the input describes one connection or several. A

connection has one coherent credential owner, endpoint/transport, resource

boundary, and independently reviewable action catalog. Record the proposed

split early and refine it as evidence arrives.

• Prefer official vendor documentation, protocol/RFC sources, safe live

probes, and current Paperclip code. Use third-party sources only to find or

qualify primary evidence. Record every factual claim with URL and access

date; mark unresolved facts explicitly instead of guessing.

Do not mutate a vendor account, register a client, grant consent, or invoke a

write tool merely to research it. A safe unauthenticated metadata probe is

allowed when it does not change vendor state. Route credential- or

browser-dependent validation through an explicit QA task when it becomes

necessary.

Keep The Run Resumable

Maintain one concise checkpoint in the Paperclip issue or an issue document:

• current phase and owning next action;
• input links and target repository commits;
• research PR URL and exact head SHA;
• proposed and accepted connection sets;
• research-gate interaction and accepted target revision;
• one branch, PR URL/head, and verification summary per connection;
• prerequisite/playbook PRs and remaining blockers.

On every restart, reconcile this state with files, branches, PRs, reviews, and

interactions before creating anything. Reuse semantic matches. Never duplicate

catalog entities, regress terminal pipeline phases, reuse stale acceptance, or

open a second PR for the same connection accidentally.

Phase A: Research In paperclip-content

• Create an isolated worktree and branch from the refreshed content target.

Preserve unrelated local changes.

• Intake the supplied links or brief through the integration harness. Let the

harness load its sibling skills for discovery, feature research, proposal

reconciliation, user stories, UI planning, implementation planning,

examples, and docs briefing. Obey every internal human gate in

integrations/README.md; the final research gate below does not replace

them.

• Build the complete reviewable OKF/planning package required by the current

pipeline. For MCP work, make the evidence sufficient to decide:

• official endpoint and transport;
• auth mode, credential ownership, scopes, discovery and DCR behavior;
• endpoint-precedence and redirect-origin constraints;
• token lifetime, rotation, refresh, revocation, and re-auth behavior;
• tool inventory, action risk, resource filters, account/tier/pricing

constraints, administrator setup, and validation needs;

• exact service involvement and system boundaries in Paperclip.
• Ground every Paperclip-surface claim in the maintained surface map and

current App code. Reconcile before creating, update required indexes/logs,

and preserve lineage, timestamps, immutable slugs, and absorbing phases as

required by the research contract.

• Open one research-only PR to paperclip-content. Include the full planning

package and any tightly coupled content-playbook correction, but no

Paperclip App implementation.

• Run focused validation and the required PR workflow. Do not present the gate

until checks are green, Greptile is 5/5, all actionable review comments are

resolved, and the recorded PR head still matches the reviewed head.

Gate Research Before Building

• Create or update a dedicated issue document that names:
• the research PR URL and exact head SHA;
• the content and App source commits used for research;
• the proposed connection set and why each item is independent;
• known limitations, prerequisites, and deferred questions.
• Create a Paperclip request_confirmation interaction targeted at that issue

document's latest revision. Use a revision-specific idempotency key and a

wake_assignee continuation policy so either acceptance or rejection wakes

the assignee. Ask the reviewer to include revision notes when rejecting.

• Put the issue in in_review and stop. Do not prepare App worktrees or code

while the interaction is pending.

• On rejection, use the interaction response and any revision notes to revise

Phase A only and present a new revision. If the research PR head, gate

document, or connection set changes, withdraw/supersede the old confirmation

and request a fresh one.

• On acceptance, verify that the response still targets the latest gate

revision and recorded PR head. Implement only the accepted connections.

Phase B: Implement In Paperclip App

Refresh the App target branch, reread the current Connector Playbook, and record

its commit before writing code. For each accepted connection:

• Create one isolated worktree, branch, and PR. Reconcile against current

AppDefinitions and connector code first.

• Follow the Connector Playbook's current decisions for catalog entry versus

plugin, reuse path, AppDefinition, transport, credential refs, resource

filters, action catalog, governance, wizard behavior, health/catalog,

availability, revocation, audit, and validation.

• Derive OAuth behavior from evidence. In particular, do not ship a complete

authorization/token endpoint pair for a discovery-capable MCP vendor unless

it is intentionally authoritative: current broker precedence can make that

pair bypass stored endpoints, challenge hints, and RFC discovery. Probe DCR

and redirect constraints safely, reuse registered clients, and cover token

rotation/terminal refresh failures when the vendor requires them.

• Add the connection documentation mandated by the current playbook,

including the service-involvement statement, a sequence diagram with exact

auth/discovery/registration/callback endpoints, and step-by-step

administrator setup.

• Add focused automated tests and the production-like validation hook for

connect, catalog discovery, an allowed read, an ask-first write, a

denied/quarantined action, revoke, and audit. Include negative company,

actor, resource, and changed-schema cases where applicable.

• Use a first-class QA child issue only when real credentials, vendor consent,

or browser evidence cannot be completed safely by the implementing agent.

Link it as a blocker and give QA exact, secret-safe steps and expected

evidence.

• Run the standard PR-preparation workflow for this connector PR. Do not hand

it back for merge until focused verification passes, all required checks are

green, Greptile is 5/5, and every actionable comment is resolved.

Complete and report each connection independently. Failure or review delay on

one connection must not cause another connection to be bundled into its PR.

Feed New Rules Upstream

At both phases, compare new evidence with the two canonical playbooks.

• Record vendor-specific facts in that vendor's catalog artifacts, manifest,

docs, and tests.

• When evidence changes a reusable schema, gate, template, auth rule, risk

policy, documentation standard, or validation rule, update the owning

upstream playbook/skill/template and add regression coverage where possible.

• Include a narrow, tightly coupled correction in the relevant research or

connector PR. Use a separate prerequisite PR when the correction affects

several connectors, changes shared infrastructure, or would obscure the

one-connection review.

• Rebase/reconcile dependent work and rerun affected planning and verification

after the correction. If the accepted research PR or connection set changes,

repeat the research gate.

• Never knowingly leave implementation, tests, and the owning playbook

inconsistent.

Finish

Before marking the issue done, report the exact research PR and head, accepted

gate revision, every connector/prerequisite PR and head, focused verification,

CI/review state, and any upstream playbook changes. Leave the issue in_review

only for a real pending interaction/reviewer/monitor path, blocked only for a

named owner and concrete unblock action, and done only when every approved

connection has a merge-ready PR and no required follow-up remains on the issue.