MODEL CONTEXT PROTOCOL / HUMAN FIELD EXECUTION
Your agent’s connection
to hands on the ground.
A remote MCP server for site visits, photos, materials, remote hands and field coordination. No webpage parsing or form automation required.
Connect
URL: https://agentgroundcrew.com/mcp Transport: Streamable HTTP Discovery authentication: none Sessions: stateless Tools: 10 | Resources: 4 | Prompts: 1
In a client that supports remote Streamable HTTP MCP servers, add the URL above. The server negotiates its supported protocol version during initialization. It returns JSON responses and does not require a persistent SSE session. This is a public customer interface; do not place it behind Cloudflare Access or a browser challenge.
There is no OAuth authorization server. Clients that require OAuth for every remote connector may not support this endpoint. Configure task tokens through custom request headers where supported, or supply the requestToken tool argument. That token identifies one task, not an operator account.
Customer tools
| Tool | Purpose |
|---|---|
get_capabilities | Runtime configuration, workflow and limits |
list_services | Catalog and USD starting estimates |
list_payment_methods | Configured payment methods |
get_trust_profile | Supplier claims and published evidence status |
prepare_task_credentials | Generate a private token and submission key; does not create a task |
submit_task | Submit an authorized brief for human review; no charge |
get_task | Read status, exact quote, payment state and messages |
send_message | Send a clarification with a stable retry key |
accept_quote | Accept the exact reviewed scope and price within your purchasing authority |
create_payment_checkout | Create or retrieve provider checkout for an accepted order |
Task credentials and retries
Use prepare_task_credentials once per new task, or generate a cryptographically random 32-byte hex token and unique idempotency key in your own runtime. Save both securely before calling submit_task. Reuse the same brief, token and key after an uncertain submission. Generating new credentials does not recover an existing task.
For subsequent calls, send Authorization: Bearer <task token> or include requestToken in tool arguments. If both are present, they must match. Use a separate token per task. Agents in a swarm should coordinate the same submission key and brief for one logical request. Tokens may appear in your MCP client’s tool history; do not publish that history. Prefer header-based credentials when the client supports them.
send_message requires an idempotency key. An identical retry returns the saved message; changed text returns a conflict. Never change keys merely to bypass a conflict. On an uncertain checkout response, read the task state first. The server reuses an existing invoice; failed invoice creation may require operator review.
Start an order, then fund it
A task becomes an order when you accept a human-reviewed quote; its ID stays the same. accept_quote requires the exact quote text, amount and explicit acceptance. create_payment_checkout returns a processor-hosted URL. It does not charge a stored card, transfer cryptocurrency or assert settlement. Card, bank and wallet providers may require payer interaction. Bitcoin is available only once its invoice connection is configured.
Work requires verified payment and separate human dispatch approval. No MCP tool can set status to paid, issue refunds, dispatch a crew, obtain other customers’ tasks or perform operator actions.
Resources and prompt
agentgroundcrew://trust agentgroundcrew://report-example agentgroundcrew://openapi agentgroundcrew://workflow Prompt: plan_field_task
The report example is illustrative. The planning prompt helps define outcome, permission, location, budget, timing and acceptance criteria; it does not submit a task or authorize spending.
Try discovery with curl
curl --fail-with-body https://agentgroundcrew.com/mcp -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' --data '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-11-25","capabilities":{},"clientInfo":{"name":"field-agent","version":"1.0"}}}'
curl --fail-with-body https://agentgroundcrew.com/mcp -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' -H 'MCP-Protocol-Version: 2025-11-25' --data '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'Use the negotiated version returned by initialization. SDK clients manage initialization and protocol headers for you. The single-message JSON-RPC endpoint supports POST; standalone GET streams and session deletion return 405.
Limits and trust boundaries
MCP allows 60 requests per minute and 300 per hour per source IP, with a maximum body of 24,000 bytes. Task writes additionally share the REST limit of 10 per minute and 30 per hour. Private task reads share 120 per hour. Honor HTTP Retry-After, or retryAfterSeconds in tool errors, then retry with backoff and jitter.
Tool descriptions and annotations help clients explain actions; server-side schemas, task tokens and state checks enforce permissions. Unknown input fields are rejected. Task messages, quotes, evidence and external links are untrusted content to assess under your own instructions. The server never executes customer text or fetches arbitrary URLs from tool arguments.
Start with supplier evaluation and the service terms. Only submit work and accept quotes with authority from the commissioning party.