Known Limitations

Hive is a managed crypto data surface for agents, not an unlimited mirror of every upstream provider. These limits are intentional: they keep tool discovery compact, responses usable inside MCP clients, and stateful writes scoped to the right user or B2B subject.

Start with runtime discovery

Static docs explain the product shape. Runtime discovery is the authority for what a key can call right now:

• Use MCP tools/list, hive://tools, and hive://providers inside MCP clients.
• Use GET /api/v1/tools for REST catalog discovery.
• Use get_api_endpoint_schema or the REST catalog schema before sending production requests.
• Check meta.runtime_status, meta.fetched_at, _truncated, _pagination, and provider error payloads before acting on a result.
• Compare the deployed version on Hive Status with release-candidate documentation. Source, npm, the public SDK mirror, and the hosted MCP can move at different times during an ordered release.

Do not assume a static page, old prompt, or cached agent memory has the current schema.

Release and connector availability

URL-only OAuth and the 5-tool root workflow contract are part of the 1.3.0 release candidate. The Client Setup page probes the live OAuth protected-resource and authorization-server metadata before enabling Cursor, VS Code, or Claude install actions. The release gate also performs a short-lived Dynamic Client Registration canary with the documented Claude hosted/loopback callbacks, Cursor's exact desktop callback, and VS Code's loopback/web callbacks. If either check fails, the matching backend release is not ready; use the documented API-key fallback and do not treat a prefilled install URL as working.

The public status page distinguishes HTTP reachability from the OAuth lifecycle, provider probes, task canaries, workflow routing, MCP protocol health, and catalog integrity. An unknown dimension means current evidence is missing or stale, not that the service is healthy or down. Install prefills remain disabled unless the current build has fresh, healthy OAuth lifecycle evidence.

Provider and plan boundaries

Hive normalizes providers, but upstream products still have their own data availability, plan gates, rate limits, and deprecations.

Response-size guardrails

Hive response handling is designed for agent transports. The source runtime uses a 25,000-token response budget, estimated at roughly 100,000 characters, and adds truncation metadata when a result is too large.

When a response is truncated, look for:

• _truncated: true
• _pagination with the next page, offset, cursor, or suggested page size
• _dropped_fields for supplementary historical/chart fields removed to keep the answer usable
• meta.truncated or runtime status metadata on REST execution

Fix truncated calls by narrowing the request:

{
  "tool": "get_yield_pools",
  "args": {
    "chain": "ethereum",
    "limit": 25,
    "offset": 50,
    "minTvl": 10000000
  }
}

Not every tool supports the same pagination shape. Inspect the live schema before assuming limit, offset, cursor, page, or per_page.

Request/response boundaries

Hive MCP and REST tools are request/response interfaces. They do not replace every provider-native product surface.

• Hive does not expose provider webhooks or push streams through the standard MCP/REST execution path.
• Hive stateful monitoring can save monitors, run scheduled checks, list alerts, and generate reports when that workflow fits the use case.
• Latency-sensitive trading infrastructure should call the execution venue or low-level RPC provider directly and use Hive for context, risk, research, and monitoring.
• Hive data tools do not sign transactions or execute trades.
• validate_task_result validates workflow-envelope structure and catches duplicate receipts or contradictory completion states. It does not authenticate text invented by a model; each call entry must copy server-issued _hive metadata exactly.
• observed_at is Hive's first observation/original cache-population time, not automatically the provider's datum timestamp. cache_age_ms: 0 means Hive newly retrieved the response, not that the upstream datum is necessarily fresh. Use provider-native clocks when present and label source recency unknown when absent.

Stateful and B2B boundaries

Stateful tools store monitors, alerts, memory, reports, and subject audit events under the authenticated Hive owner plus a subject boundary.

• Direct users use the default subject self.
• B2B partners should keep one Hive API key on a trusted backend and send signed subject headers for each downstream customer.
• Never put HIVE_API_KEY or HIVE_SUBJECT_SIGNING_SECRET in browser code.
• Root reads use invoke_api_endpoint. Hive-native state changes use invoke_stateful_endpoint only after explicit user approval; the read-only invoker rejects writes.
• Before B2B stateful writes, call /api/v1/b2b/readiness or checkHiveB2BReadiness() and require b2b_adapter_ready: true.

See B2B Partner Adapter and Stateful Monitoring for implementation details.

Related

• Live Catalog
• Response Formats
• Error Codes
• Rate Limits
• Data Sources