π OpenHands: AI-Driven Development
import { INTEGRATION_CATALOG as MCP_MARKETPLACE } from "@openhands/extensions/integrations";
import {
findCatalogEntryForServer,
getMcpMarketplaceCatalog,
} from "#/utils/mcp-marketplace-utils";
import type {
MCPServerConfig,
MCPTestToolCall,
MCPTestToolResult,
} from "#/types/mcp-server";
/**
* Credential validation specs for marketplace MCP servers whose credentials
* are only exercised on tool invocation (listing tools succeeds with any
* credentials). The test endpoint runs `toolCall` after listing and reports
* the outcome verbatim; `interpret` decides whether that outcome proves the
* credentials are invalid.
*
* `toolCall` MUST be read-only β it runs on every "Test connection" /
* pre-save validation.
*/
export interface CredentialValidation {
toolCall: MCPTestToolCall;
/** Returns the provider's error code/message, or null when creds pass. */
interpret: (toolResult: MCPTestToolResult) => string | null;
}
/**
* Slack error codes that mean the credentials themselves are bad. Anything
* else (e.g. `missing_scope`) proves the token *authenticated* β a valid
* token lacking a scope must not be reported as invalid credentials.
*/
const SLACK_AUTH_FAILURES = new Set([
"invalid_auth",
"not_authed",
"account_inactive",
"token_revoked",
"token_expired",
]);
/**
* Interpreter for hosted HTTP API servers (GitHub, Linear). The service only
* calls `interpret` when the probe tool was advertised by the server AND
* invoked, so `is_error` here means the read-only call itself failed β for a
* "who am I" / "list teams" probe that is an auth/permission failure.
*/
const interpretReadOnlyToolResult = (
toolResult: MCPTestToolResult,
): string | null =>
toolResult.is_error ? toolResult.text || "tool call failed" : null;
const VALIDATION_BY_ENTRY_ID: Record<string, CredentialValidation> = {
github: {
// Official GitHub MCP server (api.githubcopilot.com): read-only
// "get details of the authenticated user".
toolCall: { name: "get_me", arguments: {} },
interpret: interpretReadOnlyToolResult,
},
linear: {
// Linear hosted MCP (mcp.linear.app): read-only teams listing with no
// required arguments.
toolCall: { name: "list_teams", arguments: {} },
interpret: interpretReadOnlyToolResult,
},
slack: {
// Read-only: lists at most one channel. The Slack MCP server returns
// the raw Slack API JSON ({ok: boolean, error?: string}) as text.
toolCall: { name: "slack_list_channels", arguments: { limit: 1 } },
interpret: (toolResult) => {
if (toolResult.is_error) {
return toolResult.text || "tool call failed";
}
try {
const parsed = JSON.parse(toolResult.text);
if (parsed?.ok === false && SLACK_AUTH_FAILURES.has(parsed.error)) {
return String(parsed.error);
}
} catch {
// Non-JSON payload β make no claim about the credentials.
}
return null;
},
},
};
/**
* Look up the credential validation for a server by matching it to its
* marketplace catalog entry (same matching the MCP page uses for icons).
* Returns undefined for custom servers β their test behaves as before.
*/
export function getCredentialValidationForServer(
server: MCPServerConfig,
): CredentialValidation | undefined {
const entry = findCatalogEntryForServer(
server,
getMcpMarketplaceCatalog(MCP_MARKETPLACE),
);
return entry ? VALIDATION_BY_ENTRY_ID[entry.id] : undefined;
}