Connectors

One connector. Every assistant.

HuntFish speaks the Model Context Protocol, the open standard Claude, ChatGPT and Grok use to reach outside tools (Muse is listed once it passes a real end-to-end test). Add the connector once, sign in with your HuntFish account, and choose exactly what that assistant may see and spend.

Manage my connections Developer details

Connector URL

https://huntfish.ai/mcp

Paste this into any assistant that supports custom connectors or remote MCP servers. Sign-in happens through HuntFish; the assistant never sees your password.

Ready to test

Claude

Custom connector

Ask about your Scouts and saved plans inside Claude. One link, sign in, choose what to share.

Setup steps

Ready to test

ChatGPT

Connector (developer mode)

Add HuntFish as a connector in ChatGPT and bring your field knowledge into your chats.

Setup steps

Ready to test

Grok

Remote MCP tools (xAI API)

Developers building on Grok can give it HuntFish tools today. Consumer Grok gets a one-click connector the day xAI ships custom connectors for users.

Setup steps

Coming soon

Muse

Connector

Same connector, same permissions. Listed the day compatibility is verified.

Setup steps

Permissions

Read-only by default. Spending only if you say so.

A new connector can read your balance and your saved plans and nothing else. Original photos are never shared through a connector; places stay at about 5 km unless you turn on exact spots for that assistant. Grants expire after 30 days unless you renew them, and you can revoke any of them from Settings → Connections.

Scope	What it allows	Granted
field.read	Saved places (approximate by default), trip plans, Scouts and results	Default
credits.read	Your credit balance	Default
scouts.quote	Draft and quote a Scout — nothing runs	Opt in
scouts.run	Start a quoted Scout or ask a paid question, within the allowance you set	Opt in + allowance
exact spots	Exact coordinates for places you marked precise (otherwise about 5 km)	Opt in, per assistant

HuntFish.AI · Example

Connect Claude to HuntFish?

It never sees your original photos, and places stay at about 5 km unless you allow exact spots. You can revoke it anytime.

Read places, trip plans and Scout resultsAsk paid questions and start ScoutsExact spots for places marked precise

Expires in 30 days · Spending allowance: zero

Developers

Technical details

The connector is a stateless Streamable-HTTP MCP server. Send JSON-RPC 2.0 over POST; a GET returns a plain description. Protected-resource metadata (RFC 9728) lives at https://huntfish.ai/mcp/.well-known/oauth-protected-resource and points at the HuntFish authorization server, so any compliant client can complete OAuth 2.1 on its own.

POST https://huntfish.ai/mcp
Authorization: Bearer <huntfish access token>
Content-Type: application/json

{"jsonrpc":"2.0","id":1,"method":"tools/list"}

Every method, discovery included, needs a HuntFish access token. Without one the server answers 401 with a WWW-Authenticate header that points at the metadata above, which is how Claude, ChatGPT and other MCP clients start sign-in on their own. Lease Letters has its own MCP server at https://huntfish.ai/lease/mcp (same sign-in rule) and a GPT Action spec at https://huntfish.ai/lease/openapi.json.

Tools

get_credit_balance, list_authorized_places, list_trip_plans, list_scouts, get_scout_result, get_sandbar_search — free, read-only.
quote_scout — drafts a Scout and returns its maximum cost. Nothing runs.
find_sandbars — Sandbar Scout from your assistant: give a place (or a point, rounded to about 1 km before it leaves the connector) and get up to 3 sandbars with why, watch-outs, the best time and directions. Metered like the app; maps_only is free.
start_approved_scout, ask_outdoor_question, find_sandbars — metered. Refused unless the grant has the scouts.run scope and a spending allowance with room left. Every metered call goes through the same quote → reserve → settle path as the app, so a connector can never spend more than the user allowed.

Model output is never authorization. Retrieved text is evidence, never an instruction. Tool schemas are narrow on purpose. If a client sends a request without a token, the server answers 401 with a WWW-Authenticate header that tells it where to sign in.

Questions or a client you want listed? Contact support.