Trusty Squire

Trusty Squire is an MCP server that lets Claude Code, Codex, Cursor, OpenCode, Goose, and other coding agents sign up, provision, and purchase on your behalf. It opens a real browser, works through signup, sign-in, setup, and checkout flows one step at a time, clears the bot-detection and email-verification steps that make operator tools stall, and hands the job back to a person only when one is actually required. That covers wiring up OAuth and API keys for the app you're building as much as it covers paying a checkout, sending a gift, or booking something — the same operator primitives drive all of it.

Provider secrets and payment cards are write-only: the agent's credential tools return references and authenticated results, never stored plaintext. The raw secret never needs to enter the agent's context, so it can't be pasted into a commit, leaked in a log, or read back out over chat. Backend access is a host-scoped, rate-limited, independently revocable grant, so a leaked token is killed without rotating the provider key — and you connect Google or GitHub yourself in a real browser, so the agent never types your password. Full threat model below.

One prompt

Add Google OAuth to this app in one prompt: create the OAuth client, save the client secret, and wire it in without putting the raw key in chat, code, or .env.

Your coding agent plans the job. Trusty Squire operates the website, stores the generated key, and can issue your backend a scoped grant. The backend calls the provider through Trusty Squire, which injects the provider key on the server side.

Other useful asks:

• “Set up Stripe payments for this app and keep the API key out of this conversation.”

• “Create a Render API key for deployment automation and keep it out of this conversation.”

• “Pay this checkout with my saved work card and ask me to approve it on my phone.”

• “Send a gift to my friend without sharing their address with me.”

• “Book this dinner reservation for me.”

• “That app grant leaked. Revoke it without rotating the provider key.”

For supported card checkouts, save a card in the Vault from a passkey-capable device or let your first operate_pay approval link collect one just in time. When no card is specified, Trusty Squire uses the only saved card, starts the add-card ceremony if none exists, or asks you to choose when several exist. The new card is encrypted in your browser with a passkey-derived key and bound to that purchase before approval; if you add it but do not approve in time, it remains saved for a faster retry.

Recognized Visa, Mastercard, Amex, Discover, Diners Club, and JCB cards show their network mark in the Vault while keeping the full bank/network label. Open a card row to see its masked number; reveal runs the passkey ceremony in your browser before showing the number, name, expiry, and billing address. The CVV is never shown, even after reveal. The Activity page also records card additions and removals, payments, and app-grant changes without storing a PAN or CVV.

operate_pay requires a non-empty item and reason (calls that omit either receive a validation error). On a single-page checkout it prefers a machine-read checkout total, sends you a short-lived approval link, and submits only after you approve the purchase. A clean visible labeled total wins; when none is readable, strict schema.org Order/Invoice.totalPaymentDue structured data can supply the amount and currency. If neither source exposes a total, caller-supplied amount_cents and currency become the authoritative approval amount; an omitted merchant name falls back to the checkout URL's hostname. Product and offer prices never qualify as machine-read checkout totals. An unambiguous ISO currency on the page remains authoritative. A notation that cannot identify one ISO currency by itself, such as a shared currency symbol or an FX-preview selector, falls through to the currency already selected or approved for the purchase instead of refusing the checkout. Any live amount or currency drift still fails closed; the authoritative binding contract lives in the security model. Approval, 3-D Secure, Activity, and notification amounts use the currency's minor-unit precision (for example, whole yen for JPY and two decimals for USD). The anonymous approval page shows the merchant, checkout origin, amount and currency, item, and reason directly from the short-lived server record before one passkey ceremony authorizes those canonical payment values. You also see the requesting MCP client (for example, Hermes) and that a saved card will be used before clicking Approve payment to relay the operator-sealed final authorization. When the pre-submission checkout can be machine-read, the payment is refused if its merchant, origin, amount, or currency has changed since approval. If that resume read cannot recover a total, Trusty Squire reuses the original mandate-bound checkout values. Card entry requires the PAN, expiry, and CVV fields; cardholder name and other explicitly labeled billing fields are filled best-effort, so a missing name field does not abort the payment. Sealing and cleanup touch only those selected payment controls; merchant shipping address and country controls remain untouched. A submit is reported as payment_submitted only after the checkout reaches a new merchant order-confirmation URL with a substantive order or receipt identity. The browser completes 3-D Secure natively, including out-of-band bank-app challenges — Trusty Squire never manipulates or intercepts the challenge; it uses read-only checks while polling for that same order-confirmation signal. A bare click that produces no confirmation and no detected challenge returns payment_outcome_unknown instead of guessing that the charge succeeded. A detected challenge that remains unresolved on timeout stays payment_3ds_required with needs_user.wall: "3ds", handing control back for user completion. Neither status is success or permits blind resubmission: manually check the merchant's order state before any retry.

Every payment response includes its session_id. Pass that same ID to every follow-up operate_pay and canonical operate_payment_status call. Pass wait_seconds (0-15, default 0) to bound-wait for a change instead of an instant peek. Omitting session_id remains compatible only while this MCP process has exactly one session; it never selects a newest or arbitrary checkout.

Some split checkouts collect the card before the final order-confirmation step. On the card-entry page, operate_pay { phase: "fill_card" } first reads the live total. A subtotal qualifies as that payable amount only when the same order summary says shipping is free; recommendation and related-product prices are excluded. If that page exposes no total, caller-supplied amount_cents and currency take precedence as the approval amount. If they are omitted, Trusty Squire may use the most recent real total observed earlier in the same browser session, such as the cart subtotal, only when the checkout origin still matches. One phone approval binds that amount and releases the card; Trusty Squire fills the card without submitting and its role in the purchase ends there. It fills only the merchant's own HTTPS frames or recognized payment-provider frames. The card stays in the page as sealed, observation-masked fields while the agent advances to the review step and places the order. Verify the live final total against the approved amount_cents/currency yourself before placing the order; Trusty Squire no longer re-reads the total or submits anything. For click and js_click, a control whose label looks like pay/place-order may fire only once for that approval. A second recognized attempt is refused and requires a fresh operate_pay approval in a new session. Non-charge-labeled clicks, key presses, and oauth_click remain ungated. After a recognized click dispatches, Trusty Squire best-effort records a secret-free payment_place_order_attempted Activity event bound to the approval, optional mandate, approved amount/currency, merchant, and opaque card reference. This records an attempt, not a verified charge outcome.

operate_pay { phase: "confirm" } just releases the session's pending-fill lock and reports the approved terms back — it makes no browser or provider call, records no audit event itself, and never charges. It can be called any time after the fill — it does not need to happen before you place the order, and it never reads a total or verifies an amount. If a payment gets stuck or a card is declined, recover with operate_finish and start a fresh session; operate_pay does not support refilling a different card mid-session.

Before an initial single-page or fill_card call, Trusty Squire follows the actual visible card-number field and hands the checkout back when that field is hosted by PayPal or Braintree. A separate PayPal express button does not block fillable merchant or Shopify PCI card fields. Trusty Squire does not sign in to PayPal or use vaulted PayPal credentials. After any submit that has not yet reached a confirmed order, Trusty Squire waits 180 seconds by default for native completion, including out-of-band bank-app approval. A linked Telegram chat receives a challenge-specific nudge when 3-D Secure is detected, or cautious bank-app guidance when no on-page challenge appeared. Standard cross-processor 3-D Secure signals and recognized CardinalCommerce or Stripe challenge frames classify the first case only when the containing frame is visibly rendered. Hidden 3-D Secure Method pre-authentication and captcha-hosted frames never count as 3-D Secure, and an ordinary Shopify PCI card-field host alone does not either. It reports a visible decline and hands an unresolved outcome back on timeout, noting whether the Telegram nudge actually went out. three_ds_wait_seconds accepts whole seconds from 0 to 600; set it to 0 on operate_pay to skip the notification and waiting and receive the handoff immediately.

Connect Telegram under Vault Settings to receive secret-free alerts for credential, card, payment, and app-grant lifecycle changes. Routine credential retrieval and proxy access stay in Activity instead of sending a push for every request.

Related MCP server: Daemoon

Install

npx @trusty-squire/mcp connect

connect signs you in with Google or GitHub, detects your coding agent, and merges the squire MCP server into its existing configuration. Restart the agent and ask for the finished website outcome. Trusty Squire is free to start.

To choose a target explicitly:

npx @trusty-squire/mcp connect --target=codex

Supported targets: claude-code, cursor, codex, opencode, goose, cline, continue, and hermes.

The isolated operate_* browser runtime currently requires Linux and a local Chrome connection; remote CDP, macOS, and Windows operator sessions are not supported in this migration stage.

What happens

1. Your coding agent names the website and the outcome it needs: an account, authenticated setup, app publishing, a purchase, a gift, or a booking.

2. Trusty Squire works through the service flow one step at a time. Ordinary tasks open an isolated browser profile; tasks that must act as your connected Google identity use the signed-in profile directly and run one at a time.

3. If the flow produces an API key or client secret, Trusty Squire captures it into the vault without returning the raw value through its credential tools.

4. The agent can make an authenticated request, create a host-scoped app grant, or use a saved card for a supported checkout after you approve the purchase.

5. Eligible successful flows can become signed registry skills, so later runs can replay verified steps instead of rediscovering every click.

If a site requires phone verification, a hard CAPTCHA, an unresolved 3-D Secure challenge, an unsupported payment method, or another human decision, the run hands control back and tells you. It does not guess or pretend the task completed.

Supported services

Discovery pages are generated only for services with an active skill in the Trusty Squire registry. The first five detailed pages cover Braintrust, Cerebras, Clerk, DeepInfra, and Zilliz Cloud. Each sample has explicit signup evidence in its active registry record and a provider request checked against official API documentation. The service hub also lists every active registry entry; the remaining detail pages stay unpublished until their workflow and unique content pass review.

Browse the active service catalog. Maintainers can detect registry drift before merging with:

pnpm seo:verify-services

The registry controls which service pages exist. An external list is never used to claim support.

Keep provider keys out of agent context

Ask the agent to create a scoped backend grant:

Grant this backend access to Clerk through Trusty Squire with a limit of 100 requests per hour.

Before minting the grant, make api.clerk.com the credential's primary allowed host in the Vault. The egress proxy refuses every other upstream host. This explicit policy step is required when the signup host and provider API host differ.

The agent calls the real MCP tool with the service and requested limit:

grant_app_access({
  service: "clerk",
  rate_limit_per_hour: 100
})

Rate limits are opt-in. Omitting rate_limit_per_hour creates an unlimited grant; host scoping, auditing, and revocation still apply. When a grant reaches an explicit hourly limit, the proxy returns 429 rate_limited with scope: "grant", Retry-After, and window/reset metadata.

The result contains a host-scoped egress base_url and a token, not the Clerk secret key. The token is returned once through the MCP result and remains valid until revoked. That means the scoped grant token can enter agent context; it is not the provider key. Move it directly into backend-only deployment secret storage, never browser code, logs, or source control. If you need zero grant-token exposure to the model, use use_credential for agent-initiated requests instead. Trusty Squire removes the grant authorization at the boundary and injects the vaulted provider credential into the upstream request.

Security and threat model

• Provider credentials are encrypted in the vault and are write-only to agent credential tools. Those tools return references or authenticated results, not stored plaintext.

• The raw provider key is injected only into the outbound provider request. It does not need to land in chat, generated code, the consuming app, or the project's .env file.

• App grants are host-scoped, auditable, rate-limitable, and independently revocable. A leaked grant can be revoked without rotating the provider key.

• You connect Google or GitHub in a real browser. Trusty Squire does not ask the coding agent to type those passwords.

• Saved cards are encrypted in your browser with a passkey-derived key. For a single-page payment, your phone releases the card only after approving the exact purchase details shown on the approval page. On a split checkout, one amount-bound approval releases the card; Trusty Squire's role ends at the fill and the caller places the order and verifies the final total itself. The API temporarily relays only operator-sealed card ciphertext and its signed mandate. Trusty Squire's API and the coding-agent model never receive plaintext PAN or CVV. See the security model for the signed mandate's binding contract.

• Browser screenshots and diagnostics can contain whatever a website visibly rendered. Treat diagnostic artifacts as sensitive and do not ask an agent to re-observe a page after a secret is shown.

• Trusty Squire does not bypass phone verification, hard CAPTCHAs, 3-D Secure, payment authorization, or decisions that belong to a person. It stops for human input.

See the security model for the card and credential trust boundaries, and architecture for the system and data flows.

MCP tools

The default MCP registry exposes 16 tools. The essential operator surface is operate_start, operate_observe, operate_act, operate_pay, operate_payment_status, operate_finish, operate_recipe_run, and operate_recipe_save — every former standalone workflow/lifecycle/login tool name was dropped and its behavior folded into operate_act as a kind (or into operate_finish's outcome); no delegating aliases remain. Poll payment status via operate_payment_status(wait_seconds). The maintainer-only list_extract_failures → get_extract_failure DOM-diagnostics pair is excluded from that surface; set TRUSTY_SQUIRE_DIAGNOSTICS=1 in the MCP server environment to opt into the 18-tool diagnostics profile.

• Rejected tool calls return a JSON error envelope with a stable code and message. Malformed and unknown calls fail only that request; they do not stop the shared stdio process or discard its active in-memory operator session. server_unavailable includes retry.max_attempts: 1: retry once, and never kill or restart the shared operator process.

• operate_start, operate_observe, and operate_act open a website, inspect the current state, and perform one browser action at a time. Ordinary controls inside same- and cross-origin frames are included in observations with a frame_origin; known captcha challenge frames stay behind the dedicated captcha flow. Frame refs and live text=…/css=… locators preserve that boundary: same-registrable-domain frames are reachable, cross-domain frames must pass the same domain scope as goto/allow_host, opaque frames are refused, and type_secret never targets any cross-domain frame. Frame refs currently support click, js_click, type, type_secret, and select; upload, oauth_click, and oauth_login fail closed. If a visible control has no observed ref, the four locator-capable actions (click, js_click, type, and type_secret) can use a live locator; that one-off fallback is not replayable. In a live operator session, in-page XHR/fetch calls to merchant API sibling subdomains are automatically in scope only when they share the registrable domain of a host trusted at session start. Calls outside the session scope fail promptly instead of hanging; page-load resources continue normally, and a mid-session allow_host does not seed sibling-domain widening. For a task gated by the user's connected Google account, pass require_live_identity: true to operate_start. The start fails closed with a connect handoff if that Google session is unavailable; otherwise it uses the signed-in profile directly, with one such session active at a time. When DOM churn invalidates an @e: ref, operate_act returns target_stale with the last observation generation, reobserve_required: true, best-effort label-keyed replacement_candidates, and retry_policy: "do_not_retry_old_ref"; re-observe and choose a current ref. Malformed operate_act calls return error.code: "invalid_arguments" and an error.guidance repair object with the allowed kinds, missing fields, a valid example, and a safe alternative instead of only a validation string. For a provider login, pass the observed provider-button ref to the atomic oauth_login action. It retains the product tab across provider-owned popup redirects and closes, then returns the post-login product observation even if detail is none. oauth_click and oauth_settle remain for legacy replay compatibility. If an observation races that legacy transition, the response reports oauth.state: "in_progress" and directs the host to observe again.

• operate_act also owns eight consolidated workflow/lifecycle kinds — the entire operator surface beyond navigation, payment, finish, and recipe replay is reached through operate_act's kind:

  • select_many accepts an ordered label/ref-to-option map for coupled variant, shipping, or similar selectors. It applies selections sequentially, re-observes after every success, tolerates partial failure, and returns each field's selected or failed outcome plus a current observation.

  • cart_add is the retry-safe add-to-cart path. Give it the canonical product identity, selected-variant options hash, and a stable idempotency key; it post-verifies the exact cart line and returns added or already_in_cart, cart_delta (+1, 0, or unknown), and the canonical cart URL when observable, without clicking again for the same product and variant. Cart and checkout observations expose an informational, best-effort checkout_state with stage, product and variant identity, quantity, separately observed subtotal and shipping, payable total when known, canonical cart URL, and one next_action. The single and fill_card payment phases derive their authoritative approval amount independently of this state, preferring live checkout data according to the payment guide above.

  • extract captures a generated credential into a sealed slot or the vault.

  • solve_captcha drives the in-session captcha gate and returns the fail-fast needs_user handoff when it cannot be cleared.

  • await_verification reads the user's own inbox for an email verification code/link, with sender-scoped search, explicit inbox consent, and sealed-OTP transfer through into_slot.

  • login_prepare_signup, login_store_signup, and login_load_saved own the sealed username/password lifecycle. login_prepare_signup seals the user's captured email and a generated password, login_store_signup vaults those slots with explicit login-host policy, and login_load_saved retrieves an allowed saved login through encrypted browser-fill into sealed session slots. Raw values never enter the tool result.

• Observed card controls are marked payment_field and interaction: "vaulted_card_only", with operate_pay { phase: "fill_card" } as the recommended action. Typing a Luhn-valid, card-number-shaped value manually through operate_act is refused with safe_alternative: "operate_pay" and the missing prerequisite verified_cart_total.

• operate_finish closes the session and optionally accepts a nested outcome. none only closes; credentials requires store and preserves credential extraction, vault storage, and auto-promotion; result requires summary or data and can run verify_recipe before closing. Finish first stops new calls and drains calls already using that session. Payment state never blocks teardown: finish clears any remaining payment state and destroys, rather than pools, a payment-sensitive browser profile.

• operate_recipe_save saves a postcondition-verified local recipe under a closed task verb plus the service's registrable domain. It records stable target attributes and exact provenance for Squire-supplied values, not observed refs or plaintext secrets. Recipes that pass a share-eligibility check (no personal or secret-shaped literals) and a registrable-domain lock are also written live to the shared registry, making them immediately reusable by other installs without a promotion step. operate_recipe_run binds the replaying user's own values and replays those steps, preferring the local recipe and falling back to the shared one. A recipe cannot navigate outside the site it was recorded for; normal keyed replay refuses a violation before navigation and continues with cold driving. On one ordinary missed step, replay returns a local repair point and can continue in the same session. Older name-only recipes remain planning hints.

• list_payment_cards returns saved-card labels and opaque references; operate_pay accepts an explicit session_id and phase of "single" (the default, also implied by omitting phase), "fill_card", or "confirm". It can use a selected card, the only card on file, or a just-in-time add-card approval. The single-page flow fills the checkout and applies the post-submit outcome wait described above before handing back unresolved outcomes. Split checkouts use the fill_card then confirm flow described above. operate_payment_status follows the payment guide polling contract. It returns the session ID and includes it in every follow-up tool hint, so an approval is always resumed in its originating browser. Malformed calls return the same error.guidance repair fields as operate_act, including a safe resolution when card_ref and card_label conflict.

• list_credentials and use_credential find saved credentials and make authenticated API calls without returning raw values.

• grant_app_access and revoke_app_access create and remove scoped backend access.

• audit_log reports credential activity without exposing credential values.

One README for GitHub and npm

This root file is the canonical README. The npm pack lifecycle copies it into @trusty-squire/mcp byte-for-byte, then removes the generated package-local copy after packing. GitHub and npm therefore publish the same product explanation.

Development

git clone https://github.com/Trusty-Squire/trusty-squire.git
cd trusty-squire
./scripts/bootstrap.sh

After bootstrap, pnpm typecheck and pnpm test should pass. Stop local services with docker compose -f docker-compose.dev.yml down; add -v to reset their data.

Requirements: Node 20.11.0 (.nvmrc), pnpm 8.15+, Docker, and Docker Compose.

Repository map:

trusty-squire/
├── apps/
│   ├── api/        Accounts, OAuth, machine tokens, proxy, inbox, vault, and billing
│   ├── mcp/        MCP server, browser operation tools, and credential tools
│   ├── registry/   Signed website skills and verification service
│   └── web/        Marketing site and vault UI
└── packages/
    ├── vault/         Encrypted credential storage and audit log
    ├── skill-schema/  Shared schema for replayable website skills
    └── recipe-schema/ Shared wire schema for operator replay recipes

Product and public-web changes should follow PRODUCT.md and DESIGN.md.

License

MIT © Trusty Squire