Agent setup preview

How agent connection will work

Read how agent setup will work on BoxFetch. Create an agent inside the control center, choose spending controls, review purchase approvals, and let the agent browse or buy boxes — with humans in control. Agent connection is not live on the public site yet.
Setup preview onlyAgent connection coming soon
Purchases are wallet-backed and powered by credits — an in-product unit, not cash, with no buyer cash-out. BoxFetch is not an official Claude or ChatGPT connector; setup is manual, using MCP-style or custom HTTP configuration in your own client.
Developer overviewMarketplaceIntegrations →

Guided journey

Setup flow

Six steps from opening the Control Center to revoking the token when the agent is done.

  1. 1

    Request launch access

    Join first-50 access first. Control Center agent setup is private during launch.

  2. 2

    Create or select an agent

    Pick an existing agent or create a new one. The agent inherits the spending limits and approval threshold the owner sets.

  3. 3

    Generate a scoped access token

    Generate a token scoped to that agent. Tokens are revocable from the same surface at any time.

  4. 4

    Connect an external client manually

    BoxFetch is not a listed Claude or ChatGPT connector. Paste the token into the MCP-style / custom HTTP setup of your client manually.

  5. 5

    Search, preview, quote, buy, or pause for approval

    The agent discovers listings, previews safe metadata, takes a quote, and either captures the purchase inside its limit or pauses for owner approval above the threshold.

  6. 6

    Review activity and revoke if needed

    Review agent activity from the Control Center. Revoke the token whenever the work is done or the agent should stop acting.

Before you start

  • First-50 access to the private Control Center rollout.
  • An MCP-capable or custom HTTP client you configure manually.
  • Clear spending limits and an approval threshold for the agent.
  • Understanding that BoxFetch is not an official Claude or ChatGPT connector.

After connection

  • The agent can search, preview, quote, and purchase inside limits.
  • Larger purchases pause for owner approval before capture.
  • Full content unlocks only after a captured purchase.
  • Tokens are revocable from the Control Center at any time.

Safety model

What the agent can and cannot do

Trust Works routing, preview boundaries, approval thresholds, and revocable tokens.

  • 01Agents only see listed, discoverable assets routed by Trust Works.
  • 02Trust Works public badges are signals, not legal or IP certification.
  • 03Preview never exposes full content.
  • 04Any purchase above the owner threshold pauses for human approval.
  • 05Captured purchase is required before content unlocks.
  • 06Tokens are scoped to a single agent and revocable at any time.
  • 07The owner sets spending limits and the approval threshold; the marketplace enforces them.

Reference

Agent tool flow

The MCP-style tools an agent uses against BoxFetch. Connect them manually through your client’s MCP or custom HTTP setup.

Asset tools

search_assets

List discoverable marketplace assets approved by Trust Works routing.

preview_asset

Read buyer-safe preview text and metadata. Never returns full content.

quote_asset

Take a short-lived price quote in wallet credits.

purchase_asset

Capture inside the agent spending limit, or open an approval request above the threshold.

get_purchase_approval

Poll the status of an approval request. Content stays locked while pending.

get_asset_content

Return full content only after a captured purchase. Otherwise locked.

submit_usage_feedback

Post structured feedback for an asset the agent purchased.

get_agent_budget

Read remaining session and daily spend headroom for this agent.

Media for agents — controlled MCP beta, narrow and explicit-id only

Media on BoxFetch is a separate rail with tighter human controls. A narrow MCP media beta is live: a configured agent can call boxfetch_media_quote, boxfetch_media_purchase, and boxfetch_media_access against explicit media ids only. MCP cannot discover, search, preview, stream, or download media. Human playback stays access-controlled through the BoxFetch watch page.

  • Live MCP-style agent tools cover asset listings (search_assets, preview_asset, quote_asset, purchase_asset, get_purchase_approval, get_asset_content, submit_usage_feedback, get_agent_budget) plus the controlled MCP media tools described below.
  • MCP media tools boxfetch_media_quote, boxfetch_media_purchase, and boxfetch_media_access are live in controlled beta for explicit media ids only — no MCP media discovery, search, or preview.
  • MCP media purchase uses internal BoxFetch credits only through the existing paid-library purchase service; no production money, no external payout, no buyer cash-out is reachable through an agent.
  • An external MCP media purchase that requires human approval creates a media-specific approval request — no settlement and no entitlement at request time. On owner approval, the media approval rail re-reads server-side media state and settles through the existing internal-credit paid-media purchase path; the media entitlement is issued only after successful settlement. Declined, expired, stale, price-changed, or insufficient-credit approvals fail closed without settling.
  • MCP media access returns safe entitlement/watch status and the app watch path only; human playback stays access-controlled through the BoxFetch watch page.
  • No raw playback stream, Cloudflare signed token, iframe src, provider video id, R2 key, download URL, or provider secret is exposed to agents under any surface; MCP does not stream or download media.

Live MCP media tools (explicit media ids only)

Configure these in your MCP-style or custom HTTP client the same way you connect the asset tools. They accept one explicit mediaAssetId per call and never return tokens, iframe src, raw provider ids, R2 keys, streams, or download URLs.

boxfetch_media_quote

Quote one explicit media id — server-side eligibility, internal BoxFetch credit price, library entitlement hint, app watch path if already entitled. No playback token, iframe src, provider uid, R2 key, or download URL is returned.

boxfetch_media_purchase

Purchase one explicit eligible media id using internal BoxFetch credits only. When external MCP approval is required, the agent creates a media-specific approval request — no settlement and no entitlement at request time. On owner approval, the media approval rail re-reads server-side media state and settles through the existing internal-credit paid-media purchase path; the media entitlement is issued only after successful settlement. Never returns playback or provider data.

boxfetch_media_access

Read safe entitlement and watch status for one explicit media id — watchState, accessType, and the app watch path only. No Cloudflare token, iframe src, provider uid, R2 key, bytes, stream, or download URL.

Connector honesty

BoxFetch is not an official Claude or ChatGPT connector and is not listed in any official connector marketplace. Setup is manual, using MCP-style or custom HTTP configuration in your own client.