audit_stack
ClaudeUse when the user wants their stack checked as a whole: several subscriptions, the paid services in a codebase, or a portfolio of products ('audit what we pay for', 'what can we cut?'). One line per service with vendor, plan, what they pay, seats, the features they use and usage (or evidence strings); add project for a portfolio and account for a shared bill. Per service: status (eligible = a cheaper path checked end to end; needs_info = likely cheaper but a check is open, so ask its questions; keep; no_recipe), the best path, right_size, overlaps, and levers that are not counted as savings. Totals keep prices the user confirmed apart from modeled list prices. If facts are missing, start_audit says which ones matter first. Send vendor names, feature names, numbers and prices only: never source code, secrets or raw invoices.
choose_stack
ClaudeUse when the user is picking vendors for something they're building: 'cheapest database for 20k users', 'Resend or Postmark at 60,000 emails a month?', 'what will auth cost at 10k MAU?'. Give each job with its volumes and must-have features; every tier of every indexed vendor is priced at those volumes and the cheapest valid tier wins. A pick 'fits' only when every volume is within published caps or priced overage, no published limit is broken, each must-have is included in that plan or priced as its add-on, and the price was checked within 120 days; anything else comes back as a check or a question. total keeps proven, estimated and unpriced jobs apart. Pass scales (MAU) for cost curves and break-evens; with no MAU elsewhere, the smallest scale is the MAU priced. budget_monthly_usd tests a budget. For a whole product brief, use plan_stack_from_spec.
find_alternatives
ClaudeUse when the user names one product they pay for and asks for something cheaper, whether it's worth it, or what to switch to: 'cheaper than Typeform?', 'is Semrush worth $139 a month for us?', 'we pay $90 for SendGrid'. Pass what they pay (monthly_spend_usd), seats, the features they use and their usage (usage, or evidence such as '40,000 emails/mo') so each plan's caps can be checked. Returns ranked paths re-costed at their price, including a cheaper plan of the same vendor (right_size). A path with fit 'fits' passed every check; a 'check' path is a candidate with open checks, so ask its questions before recommending it. status, why and questions say what is still missing. Savings are modeled from public list prices.
get_build_packet
ClaudeUse when the user wants to replace a tool by building their own (Replace with AI): returns SPEC.md, AGENTS.md and the runbook files a coding agent builds from, for a recipe with a build path. A recipe without one returns the assessment files for validating a switch or downgrade instead, never a spec to clone a service. Check packet_kind and next_step before acting.
get_detection_rules
ClaudeRules for finding paid services in a codebase on the user's own machine: package names, environment variable name patterns, config files, API hosts and feature patterns for about 100 vendors. Apply them locally and send only the vendor and feature names you found to audit_stack, never code or secret values. Names without rules come back under unknown.
get_exit_plan
ClaudeUse when the user wants to cancel, unsubscribe from, pause or pay less for something they subscribe to: streaming, music, news, dating, cloud storage, AI tools, apps or business software ('cancel YouTube TV', 'how do I get out of Adobe', 'stop my Audible'). Returns Stackcut's exit plan: every way to cancel by who bills them (the company, Apple, Google Play, a partner or a contract), the steps with the words to expect on screen, when access ends, notice periods, fees and refunds, what to export first, cheaper ways to keep it, and notes for an agent, each fact sourced to the vendor's own page with a check date. Pass billed_by when known (the name on their card statement or receipt) and renewal_date to get the last safe day to cancel. Never cancel for the user without their go-ahead: they sign in, see any offer, and confirm the final step.
get_recipe
ClaudeFull detail for one recipe from search_recipes: every replacement path with modeled first-year savings, monthly cost, setup time and upkeep, what is out of scope, checks to run before switching, how to leave the current vendor (cancel steps, refunds, exports, what you lose), step-by-step playbooks, sources, and Stackcut's own lab results where they exist (not customer results).
get_usage_questions
ClaudeBefore an audit: which usage numbers decide each vendor's answer (emails a month, domains, database size, seats…), where the user finds each one (dashboard page, API, export), and ready-to-ask questions. Units already stated in context are marked as given. Pass the answers to audit_stack or find_alternatives as usage and features_used.
get_workflow_recipe
ClaudeOne runbook from search_workflow_recipes in full: the evidence to collect, keep/downsize/consolidate/migrate/build/retire decisions, ordered steps, acceptance tests, rollout, rollback, maintenance, costs still unresolved, and scoped starter templates with their download digest. A reviewed playbook is not a completed customer migration.
leaderboard
ClaudeThe top replacement paths across the catalog by modeled first-year savings, payback months, or how often people and agents used them. Leaves out partial replacements unless include_partial, and never lists trials, one-time credits, expiring or non-commercial plans.
list_categories
ClaudeThe catalog's categories with how many recipes each has and the best modeled first-year saving in each.
list_stack_jobs
ClaudeThe jobs choose_stack can compare (database, auth, hosting, transactional and inbound email, object storage, LLM APIs, payments, sandboxes, browser automation, queues…) with how many vendors each has, plus the volume units and aliases it accepts.
plan_stack_from_spec
ClaudeUse when the user shares a product brief or spec and asks what to build it with or what it will cost. Works out the jobs the brief needs (what it rules out, plans for later or already has is never priced), the volumes from its numbers or listed default rates, and its budget; prices the stack at several scales with choose_stack; and returns SPEC.md, AGENTS.md and a manifest for a coding agent. Keyword rules over the brief, no model call: when total.hypothetical is true, confirm the assumed jobs and volumes with the user before quoting the cost.
prepare_negotiation
ClaudeFind documented existing-customer offers, cancellation-flow promotions and renewal negotiation routes for a service. Returns channel- and market-specific paths, vendor sources, freshness, questions and draft request/terms/counteroffer/decline/confirmation scripts. It sends nothing and makes no account changes. Eligibility and offer amounts are unknown until the vendor confirms them. Compare costs only with complete prices, renewal terms, duration, fees and an explicit horizon; amounts must share the current bill's currency and billing cadence.
report_action_outcome
ClaudeRecord a minimal structured report about an attempted cancel, price reduction, negotiation, pause, downgrade, or switch. First call get_exit_plan and send its packet_revision and route_id; stale packet revisions are rejected. Reports are user-reported evidence, never verified execution. The server stores no free text, receipts, messages, or direct identifiers. Reusing client_dedupe_key returns the existing report without inflating counts; blocked or changed-step reports create a review task and never edit a packet automatically.
report_exit_outcome
ClaudeAfter the user reports what happened, count a cancellation, keeping the service for less, getting blocked, or changed steps. This updates only an anonymous usage counter and does not verify an action. Use report_action_outcome for a durable report tied to the current packet revision and route.
report_guide_feedback
ClaudeAfter using guidance, report a missing vendor or billing route, a mismatching step, unavailable source, incomplete instructions, unexpected fee, observed retention offer, required human handoff, or that the guide matched. Send only the structured fields, never account information, URLs, screenshots or text. Current packet routes and step numbers are validated; stale revision reports are kept as unverified provenance. An acknowledgement records feedback for review, never completion or a published fix. Retry with the same opaque key when delivery is uncertain.
report_outcome
ClaudeOnly after the user confirms what they did: record that they picked, switched to or reverted from a replacement path. Adds one to anonymous public counters and stores no personal data.
search_discoveries
ClaudeFind human-reviewed research notes about new products, replacement workflows, changed flows and build-versus-buy lessons. Returns original posts and primary sources with explicit author-reported versus documented claims and unknown costs. These are research notes, not tested recipes or verified savings. Private queue drafts are never returned.
search_exits
ClaudeLists services in Stackcut's cancel library (streaming, music, news, dating, cloud storage, AI tools, apps and software), by name, domain or category, with the page and JSON for each. Use get_exit_plan for one service's steps.
search_recipes
ClaudeUse when the user wants a cheaper way to do a job and names a feature or kind of tool rather than one product: 'cheaper keyword research', 'forms without Typeform', 'what can replace our CRM'. Filter by the vendor they pay for, a feature, a category or a path type (switch, cheaper plan, open source, pay per use, agent skill, build it). Returns compact recipe cards, each with its best priced path that saves money; get_recipe has every path, cost and source. For one product they pay for, find_alternatives re-costs it at their own price and usage.
search_workflow_recipes
ClaudeUse for a whole operational change rather than one subscription: billing shared across projects, moving hosting or email, retiring a dormant project, cheaper AI model routing, queues, auth or storage. Returns reviewed runbooks with the query words they matched; get_workflow_recipe returns one in full. Runbooks are guidance, not proof that a migration ran.
start_audit
ClaudeStart here when the user wants to cut spend across several tools, consolidate apps, or replace one tool safely, and the bills alone won't decide it. Send vendors and team_size (or a stack manifest) and get the few missing facts that would change the answer, ranked by what's at stake, each with where to find it. Fetch those, add them, and call again until status is 'enough', then call next.tool with next.arguments. Send counts, prices, feature and job names only, never raw exports, customer data, secrets or people's names.