Connect, then sign in. Add the hosted endpoint to your client; it opens a
browser sign-in (GitHub, Google, or an email code) the first time a tool runs.
The Free plan gives 10,000 monthly credits with no card:
Guided setup · Install guides · Example prompts
Connect
Hive is a hosted MCP server for clients that support remote Streamable HTTP.
Once the hosted OAuth deployment is enabled, interactive clients discover it
from the public endpoint and open browser authorization. Headless agents can
use an API key from secret storage:
Config prefills, URL-only with no key or secret embedded:
[Add to Cursor]
[Install in VS Code]
[Cursor setup]
[VS Code setup]
The prefills carry only the public endpoint URL. The setup pages check the live
protected-resource metadata, server card, release version, and required client
redirect profiles before showing a native install action. Until that gate
passes, use the API-key fallback from a trusted backend/client; never put a key
in an install link. Per-client config blocks are below.
What is Hive Intelligence?
A managed MCP server, REST API, and CLI that give AI agents one evidence-backed
workflow surface over live crypto market data, DeFi, wallets, token security,
DEX flows, NFTs, and on-chain network data. Agents receive
the provider, Hive retrieval time, Hive first-observation/original cache time,
cache age, fallback state, runtime status, and a unique receipt for every
material execution instead of silently mixing provider data.
Connect to your AI client
The hosted endpoint is the same everywhere. After OAuth activation,
interactive clients should start with URL-only OAuth discovery. API-key auth
remains the explicit headless fallback.
Claude Code
Or install the full plugin bundle (the hosted MCP connection plus 16 crypto
workflow skills) from this repository's plugin marketplace:
Cursor
~/.cursor/mcp.json (global) or .cursor/mcp.json (project):
VS Code (GitHub Copilot Chat)
.vscode/mcp.json (note the required type: "http"):
Claude Desktop
Claude Desktop uses the Custom Connectors UI for remote MCP. Open Settings →
Connectors → Add custom connector, set the URL to
https://mcp.hiveintelligence.xyz/mcp, then, after OAuth activation, complete
browser authorization. Do not paste a remote url block into
claude_desktop_config.json; that file is for local stdio servers.
Windsurf / Devin Desktop
Windsurf uses serverUrl (not Cursor's url) in
~/.codeium/windsurf/mcp_config.json:
Gemini CLI
Gemini CLI requires httpUrl in ~/.gemini/settings.json:
Both configs are URL-only and can use native OAuth after hosted OAuth
activation. Full per-client guides:
hiveintelligence.xyz/install.
Headless API-key fallback
For automation that cannot open a browser, keep a Hive API key in secret
storage and send Authorization: Bearer $HIVE_API_KEY. Never embed a real key
in an install link, shared config, screenshot, or repository.
Local stdio (desktop clients / self-host)
For clients that only launch local processes, such as Claude Desktop's
claude_desktop_config.json, run the published CLI as a stdio MCP server. By
default it proxies to the hosted endpoint, so it serves the same production
tools and data, authenticated with the key saved by hive auth login (tool
calls ask you to sign in without one):
Always launch hive-intelligence@latest. npx hive-mcp installs an unrelated
npm package, and npx hive-intelligence without @latest can keep running an
old version. hive doctor --fix repairs configs that do either.
To self-host with your own upstream provider keys, append "serve", "--local"
to args (or set HIVE_STDIO_MODE=local).
Local stdio does not use a hosted HIVE_API_KEY in that mode; configure only
the upstream provider keys you want that local process to call. Keyless providers remain available and providers without
credentials stay discoverable with a classified missing_key runtime status.
hive-intelligence is a stdio transport (a JSON-RPC channel, not an
interactive command). For terminal use, see the CLI.
Authentication
- Every tool call needs sign-in: OAuth in clients that support it, or an API key. Connecting and listing tools works before you sign in.
- Get a key at hiveintelligence.xyz/dashboard/keys; the Free plan needs no card.
- Authenticate the hosted endpoint with
Authorization: Bearer hive_live_... (legacy alias x-api-key also works).
- One key, least privilege: keys are scoped to your plan's rate limit and credits. Rotate or revoke from the dashboard; never commit a key or paste it into client-side code.
Powered by
Ten upstream provider integrations plus Open Data Fetch, normalized into one tool surface:
Tools & discovery
Hive's default contract is a compact six-tool root: three hero tools for the
dominant intents, plus a three-tool discovery/execution loop over the full
catalog.
Agents start with compact, paginated search_tools results, inspect
get_api_endpoint_schema, then call invoke_api_endpoint for reads.
The long-tail catalog remains discoverable behind that workflow surface: 435
callable tools across 10 categories.
Public catalog overview: www.hiveintelligence.xyz/tools/live-catalog. Authenticated REST catalog: https://mcp.hiveintelligence.xyz/api/v1/tools.
Security & trust
Crypto answers are only useful if they're trustworthy. Hive is built for that:
- Provenance on every response. Tool results carry a server-minted receipt with provider, Hive retrieval/observation time, cache age, source state, runtime status, server/build version, and SHA-256 input/result self-checks. The digests are not signatures or a retained lookup service.
observed_at is Hive's first-observation/original cache time, not necessarily the upstream event time; cache_age_ms: 0 only means newly retrieved by Hive. source reports the delivery state, while origin_source preserves whether cached data originally came from the live or fallback tier. Use provider timestamps, blocks, slots, transactions, or candle closes for source recency, and mark it unknown when absent. Hive never silently mixes provider data: fallback, cached, or degraded data is labeled as such.
- Point-in-time, not drift. Time-series tools accept
at / block_number so agents answer historical questions without quietly defaulting to "latest."
- Security-first tools.
get_token_security, detect_rugpull, approval-risk, and Tenderly transaction simulation return structured risk flags so an agent can check before a user signs.
- Least privilege & prompt-injection awareness. Use a scoped key per environment and rotate from the dashboard. As with any tool-using agent, treat on-chain text (token names, memos) as untrusted input. Hive returns structured fields rather than free-form instructions to reduce injection surface.
- No client-side keys. Keep your Hive key server-side; browser UIs should call your own backend, which uses the key.
Example prompts
Once Hive is connected, ask in plain English. Each prompt maps to a real tool call you don't have to write:
More workflow guides: hiveintelligence.xyz/use-cases.
Agent skills
Installable agent skills teach Claude Code, Cursor, Codex, and other agents the Hive workflow: MCP setup, tool discovery, and live crypto research:
Claude Code, OpenAI / Codex, and Cursor plugin bundle
This repository root is also a distribution-ready plugin bundle. It combines
public, reviewable pieces:
.claude-plugin/: Claude Code plugin and marketplace
manifests over the same MCP connection and workflow skills
(claude plugin marketplace add hive-intel/hive-sdk).
.codex-plugin/plugin.json: product and
interface metadata, starter prompts, and component declarations.
.cursor-plugin/plugin.json: the
Cursor-native manifest over the same MCP connection and workflow skills.
.mcp.json: the URL-only hosted Hive connection. It contains
no API key, token, header, or credential placeholder.
skills/: bundled setup, discovery, and crypto-research
workflows that teach the agent to inspect schemas, stay within workflow call
budgets, preserve provenance, and require explicit approval before a
Hive-native state change.
marketplace-review.json: release-bound listing
copy, starter prompts, and the exact five positive plus three negative review
cases. It is a public preparation fixture, not a Cursor or OpenAI manifest;
reviewer credentials, domain challenges, availability choices, and legal
attestations remain portal-only.
A Codex catalog or Cursor plugin review can ingest the repository root so the
remote MCP connection and its workflow guidance arrive together. This bundle
does not claim that Hive is already listed in either public marketplace; use the current
Hive install guide until a catalog
listing is live.
CLI
The hive CLI is a thin terminal client over the same API. Set HIVE_API_KEY (or run hive auth login once):
Global flags include --json, --pretty, --jq <expr>, --csv, --fields, --timeout, -q/--quiet. Auth: hive auth login | whoami | profiles | switch. Diagnostics: hive doctor, hive status. Shell completion: hive completion <bash|zsh|fish> --install. Aliases: hive alias set btc 'market price --ids bitcoin --vs usd'. Full reference: hiveintelligence.xyz/cli.
Configuration
Pricing
One credit = one material endpoint execution, regardless of provider or response size. search_tools, get_api_endpoint_schema, MCP tools/list, MCP resource reads, and authenticated GET /api/v1/tools are free. Full pricing: hiveintelligence.xyz/pricing · machine-readable: hiveintelligence.xyz/pricing.md.
Why Hive over a single-provider MCP?
Single-provider MCPs win on niche depth. Hive wins when the agent needs broad crypto context in one request: prices + DeFi + wallet + security + DEX in a single conversation, without anyone figuring out which tool lives in which provider.
FAQ
What does an API key cost?
Nothing to start: the Free plan gives 10,000 monthly credits with no card required. Paid plans start at $49/month (Starter, 100K credits); Pro is $149/month for 500K. Get a key.
Hosted vs local stdio: which should I use?
Hosted (https://mcp.hiveintelligence.xyz/mcp) is recommended for most integrations: no local server, Hive runs auth, rate limits, and provider infrastructure. Use local stdio for desktop setups, self-hosting, or your own upstream provider keys.
Which AI clients support MCP?
Claude Desktop, Claude Code, Cursor, Windsurf, VS Code (Copilot Chat), Codex CLI, Gemini CLI, OpenAI Responses API, and clients that support Streamable HTTP MCP. Native OAuth connectors become available after hosted OAuth activation; API-key headers remain the trusted headless fallback.
What chains are supported?
EVM (Ethereum, Arbitrum, Optimism, Base, Polygon, BNB Chain, Avalanche, and 90+ more), Solana with full Helius DAS coverage including compressed NFTs, Bitcoin, and others depending on the provider mix per category.
Is the source code open?
The Hive agent skills are MIT-licensed and open source. The Hive MCP server that powers mcp.hiveintelligence.xyz is a managed, proprietary service.
How does Hive keep answers trustworthy?
Every material result carries provider attribution, source/cache state, runtime status, and a server-minted receipt. Hive separates its own observation time from provider source recency, never silently mixes provider data, and exposes point-in-time parameters where providers support them. See Security & trust.
Documentation
Support
License
MIT © Hive Intelligence