packages/mcp-server/src/server.ts · aryaamin/openclasp

The AI trust layer between agents: identity, agreements, safeguards, and outcome intelligence for A2A.
Viewing files at commit 9a8fdf2387ffCommit details View latest
packages/mcp-server/src/server.ts 984 lines · 37.7 KB · typescript
import { McpServer } from '@modelcontextprotocol/server';
import { z } from 'zod';
import { createIdentity, FixtureFactCheckProvider, TrustEngine } from '../../core/src/index.js';
import {
  AgentIdentitySchema,
  DEFAULT_EXTENSION_URI,
  DelegationCredentialSchema,
  FeedbackSchema,
  InteractionContractSchema,
  InteractionEventSchema,
  ReceiptSchema,
  TrustEnvelopeSchema,
  canonicalHash,
  type FederatedInteraction,
  type PublicAgentCard,
} from '../../protocol/src/index.js';
import {
  requestAgentSetup,
  resolveInstallation,
  updateAgentProfile,
  type OnboardingStore,
} from '../../persistence/src/onboarding.js';
import { buildPublicAgentCard } from '../../persistence/src/hosted.js';

export const OPENCLASP_TOOL_NAMES = [
  'openclasp_create_identity',
  'openclasp_register_agent',
  'openclasp_get_profile',
  'openclasp_assess_counterparty',
  'openclasp_begin_interaction',
  'openclasp_record_event',
  'openclasp_check_claim',
  'openclasp_validate_commitment',
  'openclasp_suggest_resolution',
  'openclasp_complete_interaction',
  'openclasp_submit_feedback',
  'openclasp_raise_dispute',
  'openclasp_verify_receipt',
  'openclasp_setup',
  'openclasp_get_identity',
  'openclasp_switch_agent',
  'openclasp_update_profile',
  'openclasp_connection_status',
  'openclasp_register_delegation',
  'openclasp_save_contract',
  'openclasp_permit_mediation',
  'openclasp_resolve_dispute',
  'openclasp_find_agent',
  'openclasp_search_agents',
  'openclasp_connect_to_agent',
  'openclasp_list_invitations',
  'openclasp_respond_invitation',
  'openclasp_get_shared_interaction',
] as const;

export const HOSTED_OPENCLASP_TOOL_NAMES = OPENCLASP_TOOL_NAMES.filter(
  (name) => name !== 'openclasp_create_identity',
);

const READ_ONLY_TOOL = {
  readOnlyHint: true,
  destructiveHint: false,
  idempotentHint: true,
} as const;
const WRITE_TOOL = {
  readOnlyHint: false,
  destructiveHint: false,
  idempotentHint: false,
} as const;

const text = (value: unknown) => ({
  content: [{ type: 'text' as const, text: JSON.stringify(value, null, 2) }],
});

export const OPENCLASP_MCP_INSTRUCTIONS =
  'Call openclasp_connection_status first. If unbound, call openclasp_setup with your known public A2A endpoint; one owner approval binds and publishes you. Use openclasp_connect_to_agent with a target and task: safe matching requests activate automatically and return a ready A2A request. Risky requests require approval. Never claim another identity or upload raw conversations. OpenClasp assures A2A; it is not transport.';

export function buildMcpServer(engine = new TrustEngine()) {
  const server = new McpServer(
    { name: 'openclasp', version: '0.1.0' },
    {
      instructions: OPENCLASP_MCP_INSTRUCTIONS,
    },
  );
  return registerOpenClaspTools(server, engine);
}

type HostedRecord = (
  operatorId: string,
  kind:
    | 'agent'
    | 'delegation'
    | 'contract'
    | 'interaction'
    | 'event'
    | 'receipt'
    | 'feedback'
    | 'conflict'
    | 'profile',
  recordId: string,
  value: unknown,
) => Promise<void>;

type ToolContext = {
  http?: {
    authInfo?: { clientId?: string; extra?: Record<string, unknown> };
  };
};
type EngineSource = TrustEngine | ((context: ToolContext) => Promise<TrustEngine>);
type AgentDirectory = {
  publishAgent?(operatorId: string, card: PublicAgentCard): Promise<PublicAgentCard>;
  getPublishedAgent(agentId: string): Promise<PublicAgentCard | undefined>;
  searchPublishedAgents(input: {
    query?: string | undefined;
    capability?: string | undefined;
    limit?: number | undefined;
  }): Promise<PublicAgentCard[]>;
  createFederatedInteraction(
    operatorId: string,
    value: FederatedInteraction,
  ): Promise<FederatedInteraction>;
  listFederatedInteractions(operatorId: string): Promise<FederatedInteraction[]>;
  getFederatedInteraction(
    operatorId: string,
    interactionId: string,
  ): Promise<FederatedInteraction | undefined>;
  respondToFederatedInteraction(
    operatorId: string,
    interactionId: string,
    agentId: string,
    decision: 'accept' | 'reject',
    method?: 'oauth_installation' | 'oauth_account' | 'policy_auto_accept',
  ): Promise<FederatedInteraction>;
};

function installationContext(context: ToolContext) {
  const operatorId = context.http?.authInfo?.extra?.operatorId;
  const clientId = context.http?.authInfo?.clientId;
  if (typeof operatorId !== 'string' || typeof clientId !== 'string')
    throw new Error('Authenticated MCP installation context is required');
  return { operatorId, clientId };
}

export function registerOpenClaspTools(
  server: McpServer,
  engineSource: EngineSource = new TrustEngine(),
  recordHosted?: HostedRecord,
  onboardingStore?: OnboardingStore,
  agentDirectory?: AgentDirectory,
) {
  const factChecker = new FixtureFactCheckProvider();
  const resolveEngine = (context: ToolContext) =>
    typeof engineSource === 'function' ? engineSource(context) : Promise.resolve(engineSource);
  const persist = async (
    context: ToolContext,
    kind: Parameters<HostedRecord>[1],
    id: string,
    value: unknown,
  ) => {
    const operatorId = context.http?.authInfo?.extra?.operatorId;
    if (recordHosted && typeof operatorId === 'string')
      await recordHosted(operatorId, kind, id, value);
  };

  const requireBoundAgent = async (context: ToolContext, claimedAgentId?: string) => {
    if (!onboardingStore) return undefined;
    const connection = installationContext(context);
    const binding = await resolveInstallation(
      onboardingStore,
      connection.operatorId,
      connection.clientId,
    );
    if (binding.status !== 'connected')
      throw new Error('Call openclasp_setup and obtain owner confirmation before using this tool');
    if (claimedAgentId && binding.agent.agentId !== claimedAgentId)
      throw new Error('The claimed agent does not match this MCP installation');
    return binding;
  };

  if (!onboardingStore)
    server.registerTool(
      OPENCLASP_TOOL_NAMES[0],
      {
        title: 'Create local cryptographic identity',
        description:
          'Create an Ed25519 identity for local development. This returns private key material and is intentionally unavailable on hosted MCP.',
        inputSchema: z.object({
          agentId: z.string(),
          operatorRef: z.string(),
          capabilities: z.array(z.string()),
        }),
        annotations: { readOnlyHint: false, destructiveHint: false, idempotentHint: false },
      },
      async (input, context) => {
        const authenticatedOperator = context.http?.authInfo?.extra?.operatorId;
        const created = createIdentity({
          ...input,
          operatorRef:
            typeof authenticatedOperator === 'string'
              ? `auth0:${authenticatedOperator}`
              : input.operatorRef,
        });
        return text(created);
      },
    );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[1],
    {
      title: 'Register cryptographic identity',
      description: 'Register and verify a signed agent identity',
      inputSchema: AgentIdentitySchema,
      annotations: { ...WRITE_TOOL, idempotentHint: true },
    },
    async (input, context) => {
      const authenticatedOperator = context.http?.authInfo?.extra?.operatorId;
      if (
        typeof authenticatedOperator === 'string' &&
        input.operatorRef !== `auth0:${authenticatedOperator}`
      ) {
        throw new Error('Agent identity is not owned by the authenticated operator');
      }
      await requireBoundAgent(context, input.agentId);
      const engine = await resolveEngine(context);
      const registered = engine.registerAgent(input);
      await persist(context, 'agent', registered.agentId, registered);
      return text(registered);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[2],
    {
      title: 'Get contextual reliability',
      description: 'Get a task-specific behavioural profile',
      inputSchema: z.object({ agentId: z.string(), version: z.string(), taskCategory: z.string() }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      await requireBoundAgent(context);
      const engine = await resolveEngine(context);
      return text(engine.getRisk(input.agentId, input.version, input.taskCategory));
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[3],
    {
      title: 'Assess counterparty',
      description: 'Assess a counterparty in context',
      inputSchema: z.object({
        envelope: TrustEnvelopeSchema,
        action: z.string(),
        dataClasses: z.array(z.string()).optional(),
        humanApproved: z.boolean().optional(),
      }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      await requireBoundAgent(context);
      const engine = await resolveEngine(context);
      return text(engine.assess(input as any));
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[4],
    {
      title: 'Begin assured interaction',
      description: 'Begin an interaction and return its identifier',
      inputSchema: z.object({
        purpose: z.string(),
        parties: z.array(z.string()),
        taskCategory: z.string(),
      }),
      annotations: WRITE_TOOL,
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      if (binding && !input.parties.includes(binding.agent.agentId))
        throw new Error('Interaction parties must include the agent bound to this installation');
      const interaction = {
        interactionId: crypto.randomUUID(),
        status: 'contract_required',
        createdAt: new Date().toISOString(),
        ...input,
      };
      await persist(context, 'interaction', interaction.interactionId, interaction);
      return text(interaction);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[5],
    {
      title: 'Record signed event',
      description: 'Validate and record a signed structured event',
      inputSchema: InteractionEventSchema,
      annotations: { ...WRITE_TOOL, idempotentHint: true },
    },
    async (input, context) => {
      await requireBoundAgent(context, input.agentId);
      const engine = await resolveEngine(context);
      const event = engine.recordEvent(input);
      await persist(context, 'event', event.eventId, event);
      return text(event);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[6],
    {
      title: 'Check factual claim',
      description: 'Check an objective factual claim',
      inputSchema: z.object({ claim: z.string(), permission: z.boolean().optional() }),
      annotations: { ...READ_ONLY_TOOL, openWorldHint: true },
    },
    async (input) => text(await factChecker.check(input.claim, input.permission)),
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[7],
    {
      title: 'Validate commitment',
      description: 'Check whether a commitment is represented by a signed event',
      inputSchema: z.object({ interactionId: z.string().uuid(), commitment: z.string() }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      await requireBoundAgent(context);
      const engine = await resolveEngine(context);
      return text({
        valid: [...engine.events.values()].some(
          (event) =>
            event.interactionId === input.interactionId &&
            event.eventType === 'commitment' &&
            JSON.stringify(event.payload).includes(input.commitment),
        ),
      });
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[8],
    {
      title: 'Suggest dispute resolution',
      description: 'Suggest attributable resolutions for a consented conflict',
      inputSchema: z.object({ conflictId: z.string() }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      const engine = await resolveEngine(context);
      const conflict = engine.conflicts.get(input.conflictId);
      if (binding && (!conflict || !(binding.agent.agentId in conflict.permissions)))
        throw new Error('The bound agent is not a participant in this dispute');
      return text(engine.conflicts.get(input.conflictId)?.possibleResolutions ?? []);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[9],
    {
      title: 'Complete interaction',
      description: 'Submit a signed completion receipt',
      inputSchema: ReceiptSchema,
      annotations: { ...WRITE_TOOL, idempotentHint: true },
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      if (binding && !input.participants.includes(binding.agent.agentId))
        throw new Error('Receipt participants must include the bound agent');
      const engine = await resolveEngine(context);
      const receipt = engine.submitReceipt(input);
      await persist(context, 'receipt', receipt.receiptId, receipt);
      return text(receipt);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[10],
    {
      title: 'Submit bilateral feedback',
      description: 'Submit receipt-backed structured feedback',
      inputSchema: FeedbackSchema,
      annotations: { ...WRITE_TOOL, idempotentHint: true },
    },
    async (input, context) => {
      await requireBoundAgent(context, input.reviewerAgentId);
      const engine = await resolveEngine(context);
      const result = engine.submitFeedback(input);
      await persist(context, 'feedback', input.feedbackId, input);
      for (const profile of engine.profiles.values())
        await persist(
          context,
          'profile',
          `${profile.agentId}|${profile.agentVersion}|${profile.taskCategory}`,
          profile,
        );
      return text(result);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[11],
    {
      title: 'Raise dispute',
      description: 'Create a mutually consented dispute',
      inputSchema: z.object({
        interactionId: z.string().uuid(),
        issue: z.string(),
        participants: z.array(z.string()).min(2),
      }),
      annotations: WRITE_TOOL,
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      if (binding && !input.participants.includes(binding.agent.agentId))
        throw new Error('Dispute participants must include the bound agent');
      const engine = await resolveEngine(context);
      const conflict = engine.createConflict({
        interactionId: input.interactionId,
        issue: input.issue,
        participants: input.participants,
        positions: {},
        evidence: [],
        contractClauses: [],
        missingInformation: [],
        possibleResolutions: [],
      });
      await persist(context, 'conflict', conflict.conflictId, conflict);
      return text(conflict);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[12],
    {
      title: 'Verify receipt',
      description: 'Verify a signed interaction receipt',
      inputSchema: ReceiptSchema,
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      await requireBoundAgent(context);
      const engine = await resolveEngine(context);
      try {
        engine.verifyReceipt(input);
        return text({ valid: true });
      } catch {
        return text({ valid: false });
      }
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[13],
    {
      title: 'Set up this agent',
      description:
        'Propose this installation, public endpoint, and safe automation policy. One owner approval binds and optionally publishes the agent.',
      inputSchema: z
        .object({
          agentName: z.string().trim().min(1).max(100),
          projectName: z.string().trim().min(1).max(100).optional(),
          projectId: z.string().optional(),
          framework: z.string().trim().max(100).optional(),
          description: z.string().trim().max(500).optional(),
          agentVersion: z.string().trim().min(1).max(100).optional(),
          a2aEndpoint: z.string().url().optional(),
          autoPublish: z.boolean().optional(),
          autoAcceptPolicy: z.enum(['off', 'safe_matching']).optional(),
          autoAcceptTaskCategories: z.array(z.string().trim().min(1).max(100)).max(100).optional(),
          capabilities: z.array(z.string().trim().min(1).max(100)).max(100).optional(),
          limitations: z.array(z.string().trim().min(1).max(300)).max(100).optional(),
        })
        .refine((input) => input.projectName || input.projectId, {
          message: 'projectName or projectId is required',
        }),
      annotations: WRITE_TOOL,
    },
    async (input, context) => {
      if (!onboardingStore) throw new Error('Hosted agent onboarding is not configured');
      const connection = installationContext(context);
      const current = await resolveInstallation(
        onboardingStore,
        connection.operatorId,
        connection.clientId,
      );
      if (current.status === 'connected') return text(current);
      const request = await requestAgentSetup(onboardingStore, connection.operatorId, {
        clientId: connection.clientId,
        ...input,
      });
      return text({
        status: 'pending_confirmation',
        confirmationRequired: true,
        request,
        confirmationUrl: `${process.env.OPENCLASP_DASHBOARD_URL ?? 'https://openclasp.vercel.app'}/connect`,
        next: 'Ask the owner to confirm this setup request in the OpenClasp dashboard, then call openclasp_get_identity.',
      });
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[14],
    {
      title: 'Get my agent identity',
      description:
        'Return the agent identity automatically bound to this authenticated MCP installation.',
      inputSchema: z.object({}),
      annotations: READ_ONLY_TOOL,
    },
    async (_input, context) => {
      if (!onboardingStore) throw new Error('Hosted agent onboarding is not configured');
      const connection = installationContext(context);
      return text(
        await resolveInstallation(onboardingStore, connection.operatorId, connection.clientId),
      );
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[15],
    {
      title: 'Switch agent identity',
      description:
        'Request that this MCP installation switch to another existing agent. Dashboard confirmation is required.',
      inputSchema: z.object({ existingAgentId: z.string().min(1) }),
      annotations: WRITE_TOOL,
    },
    async (input, context) => {
      if (!onboardingStore) throw new Error('Hosted agent onboarding is not configured');
      const connection = installationContext(context);
      const request = await requestAgentSetup(onboardingStore, connection.operatorId, {
        clientId: connection.clientId,
        action: 'switch',
        existingAgentId: input.existingAgentId,
      });
      return text({
        status: 'pending_confirmation',
        confirmationRequired: true,
        request,
        confirmationUrl: `${process.env.OPENCLASP_DASHBOARD_URL ?? 'https://openclasp.vercel.app'}/connect`,
      });
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[16],
    {
      title: 'Update my agent profile',
      description: 'Update the profile of the agent bound to this MCP installation.',
      inputSchema: z.object({
        name: z.string().trim().min(1).max(100).optional(),
        framework: z.string().trim().min(1).max(100).optional(),
        description: z.string().trim().max(500).optional(),
        agentVersion: z.string().trim().min(1).max(100).optional(),
        a2aEndpoint: z.union([z.string().url(), z.literal('')]).optional(),
        autoPublish: z.boolean().optional(),
        autoAcceptPolicy: z.enum(['off', 'safe_matching']).optional(),
        autoAcceptTaskCategories: z.array(z.string().trim().min(1).max(100)).max(100).optional(),
        capabilities: z.array(z.string().trim().min(1).max(100)).max(100).optional(),
        limitations: z.array(z.string().trim().min(1).max(300)).max(100).optional(),
      }),
      annotations: WRITE_TOOL,
    },
    async (input, context) => {
      if (!onboardingStore) throw new Error('Hosted agent onboarding is not configured');
      const connection = installationContext(context);
      const agent = await updateAgentProfile(
        onboardingStore,
        connection.operatorId,
        connection.clientId,
        input,
      );
      const published = await agentDirectory?.getPublishedAgent(agent.agentId);
      if (published && agentDirectory?.publishAgent)
        await agentDirectory.publishAgent(
          connection.operatorId,
          buildPublicAgentCard(
            agent,
            process.env.OPENCLASP_PUBLIC_URL ??
              process.env.OPENCLASP_DASHBOARD_URL ??
              'https://openclasp.vercel.app',
            published,
          ),
        );
      return text(agent);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[17],
    {
      title: 'Check OpenClasp connection',
      description:
        'Check whether this MCP installation is unbound, awaiting confirmation, or connected.',
      inputSchema: z.object({}),
      annotations: READ_ONLY_TOOL,
    },
    async (_input, context) => {
      if (!onboardingStore) throw new Error('Hosted agent onboarding is not configured');
      const connection = installationContext(context);
      const binding = await resolveInstallation(
        onboardingStore,
        connection.operatorId,
        connection.clientId,
      );
      if (binding.status === 'connected') return text(binding);
      const state = await onboardingStore.list(connection.operatorId);
      const pending = state
        .filter((row) => row.kind === 'setup_request')
        .map((row) => row.payload as { clientId?: string; status?: string })
        .find(
          (request) => request.clientId === connection.clientId && request.status === 'pending',
        );
      return text(
        pending
          ? {
              status: 'pending_confirmation',
              confirmationUrl: `${process.env.OPENCLASP_DASHBOARD_URL ?? 'https://openclasp.vercel.app'}/connect`,
            }
          : binding,
      );
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[18],
    {
      title: 'Register delegation',
      description: 'Register and verify a signed, scoped delegation between two known agents.',
      inputSchema: DelegationCredentialSchema,
      annotations: { readOnlyHint: false, destructiveHint: false, idempotentHint: true },
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      if (
        binding &&
        input.parentAgentId !== binding.agent.agentId &&
        input.childAgentId !== binding.agent.agentId
      )
        throw new Error('The bound agent is not a party to this delegation');
      const engine = await resolveEngine(context);
      engine.delegations.set(input.delegationId, input);
      if (!engine.verifyDelegation(input.delegationId)) {
        engine.delegations.delete(input.delegationId);
        throw new Error('Delegation verification failed');
      }
      await persist(context, 'delegation', input.delegationId, input);
      return text({ valid: true, delegation: input });
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[19],
    {
      title: 'Save signed agreement',
      description:
        'Verify and save a complete interaction contract signed by every participating agent.',
      inputSchema: InteractionContractSchema,
      annotations: { readOnlyHint: false, destructiveHint: false, idempotentHint: true },
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      if (binding && !input.parties.includes(binding.agent.agentId))
        throw new Error('Contract parties must include the bound agent');
      const contract = (await resolveEngine(context)).saveContract(input);
      await persist(context, 'contract', contract.interactionId, contract);
      return text(contract);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[20],
    {
      title: 'Consent to mediation',
      description:
        "Record the bound agent's explicit consent to mediation for a dispute it participates in.",
      inputSchema: z.object({ conflictId: z.string(), agentId: z.string().optional() }),
      annotations: { readOnlyHint: false, destructiveHint: false, idempotentHint: true },
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      const agentId = binding?.agent.agentId ?? input.agentId;
      if (!agentId) throw new Error('agentId is required for local MCP');
      const engine = await resolveEngine(context);
      const conflict = engine.permitMediation(input.conflictId, agentId);
      await persist(context, 'conflict', conflict.conflictId, conflict);
      return text(conflict);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[21],
    {
      title: 'Resolve mediated dispute',
      description:
        'Save an agreed resolution only after every dispute participant has consented to mediation.',
      inputSchema: z.object({ conflictId: z.string(), resolution: z.string().trim().min(1) }),
      annotations: { readOnlyHint: false, destructiveHint: false, idempotentHint: false },
    },
    async (input, context) => {
      const binding = await requireBoundAgent(context);
      const engine = await resolveEngine(context);
      const existing = engine.conflicts.get(input.conflictId);
      if (binding && (!existing || !(binding.agent.agentId in existing.permissions)))
        throw new Error('The bound agent is not a participant in this dispute');
      const conflict = engine.resolveConflict(input.conflictId, input.resolution);
      await persist(context, 'conflict', conflict.conflictId, conflict);
      return text(conflict);
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[22],
    {
      title: 'Find published agent',
      description:
        'Find an exact owner-published agent card. Returns capabilities and limitations, never owner or conversation data.',
      inputSchema: z.object({ agentId: z.string().min(1) }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      await requireBoundAgent(context);
      if (!agentDirectory) throw new Error('The shared agent directory is not configured');
      return text((await agentDirectory.getPublishedAgent(input.agentId)) ?? { found: false });
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[23],
    {
      title: 'Search published agents',
      description:
        'Search owner-published agent cards by name, framework, or capability without exposing owner identities or projects.',
      inputSchema: z.object({
        query: z.string().trim().max(100).optional(),
        capability: z.string().trim().max(100).optional(),
        limit: z.number().int().min(1).max(50).optional(),
      }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      await requireBoundAgent(context);
      if (!agentDirectory) throw new Error('The shared agent directory is not configured');
      return text(await agentDirectory.searchPublishedAgents(input));
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[24],
    {
      title: 'Automatically connect to another agent',
      description:
        'Infer safe contract defaults, auto-activate when the responder policy permits, and return a ready-to-send A2A request.',
      inputSchema: z
        .object({
          targetAgentId: z.string().min(1).optional(),
          targetAgentCardUrl: z.string().url().optional(),
          task: z.string().trim().min(1).max(2000).optional(),
          purpose: z.string().trim().min(1).max(500).optional(),
          taskCategory: z.string().trim().min(1).max(100).optional(),
          requestedOutcome: z.string().trim().min(1).max(1000).optional(),
          successCriteria: z.array(z.string().trim().min(1)).min(1).max(50).optional(),
          allowedActions: z.array(z.string().trim().min(1)).max(100).default([]),
          prohibitedActions: z.array(z.string().trim().min(1)).max(100).default([]),
          allowedData: z.array(z.string().trim().min(1)).max(100).default([]),
          prohibitedData: z.array(z.string().trim().min(1)).max(100).default([]),
          evidenceRequirements: z.array(z.string().trim().min(1)).max(100).default([]),
          deadline: z.string().datetime().optional(),
          expiresInMinutes: z.number().int().min(5).max(10080).default(60),
        })
        .superRefine((value, context) => {
          if (!value.targetAgentId && !value.targetAgentCardUrl)
            context.addIssue({
              code: 'custom',
              message: 'targetAgentId or targetAgentCardUrl is required',
            });
          if (!value.task && !value.purpose)
            context.addIssue({ code: 'custom', message: 'task or purpose is required' });
        }),
      annotations: { ...WRITE_TOOL, openWorldHint: true },
    },
    async (input, context) => {
      if (!agentDirectory) throw new Error('Federated agent connections are not configured');
      const binding = await requireBoundAgent(context);
      if (!binding) throw new Error('A hosted, bound MCP installation is required');
      const connection = installationContext(context);
      let initiatorCard = await agentDirectory.getPublishedAgent(binding.agent.agentId);
      if (
        !initiatorCard &&
        binding.agent.autoPublish &&
        binding.agent.a2aEndpoint &&
        agentDirectory.publishAgent
      )
        initiatorCard = await agentDirectory.publishAgent(
          connection.operatorId,
          buildPublicAgentCard(
            binding.agent,
            process.env.OPENCLASP_PUBLIC_URL ??
              process.env.OPENCLASP_DASHBOARD_URL ??
              'https://openclasp.vercel.app',
          ),
        );
      if (!initiatorCard)
        throw new Error(
          'This agent is private. Enable automatic publishing once in its dashboard settings.',
        );
      let targetAgentId = input.targetAgentId;
      if (!targetAgentId && input.targetAgentCardUrl) {
        const cardUrl = new URL(input.targetAgentCardUrl);
        const trustedOrigin = new URL(
          process.env.OPENCLASP_DASHBOARD_URL ?? 'https://openclasp.vercel.app',
        ).origin;
        const match = cardUrl.pathname.match(/^\/agents\/([^/]+)\/card\.json$/);
        if (cardUrl.origin !== trustedOrigin || !match?.[1])
          throw new Error('Only OpenClasp-hosted Agent Card URLs are accepted');
        targetAgentId = decodeURIComponent(match[1]);
      }
      if (!targetAgentId) throw new Error('Target agent is required');
      const responderCard = await agentDirectory.getPublishedAgent(targetAgentId);
      if (!responderCard) throw new Error('Target agent is not published on OpenClasp');
      const responderTransport = responderCard.transports[0];
      if (!responderTransport) throw new Error('Target agent has not published an A2A endpoint');
      const task = input.task ?? input.purpose!;
      const purpose = input.purpose ?? task.slice(0, 500);
      const taskCategory = input.taskCategory ?? responderCard.capabilities[0] ?? 'general';
      const requestedOutcome = input.requestedOutcome ?? task.slice(0, 1000);
      const successCriteria = input.successCriteria ?? [
        'Return a clear result that directly addresses the requested task',
      ];
      const existing = (await agentDirectory.listFederatedInteractions(connection.operatorId)).find(
        (candidate) =>
          candidate.initiatorAgentId === binding.agent.agentId &&
          candidate.responderAgentId === responderCard.agentId &&
          ['pending', 'active'].includes(candidate.status) &&
          candidate.contract.purpose === purpose &&
          candidate.contract.taskCategory === taskCategory &&
          candidate.contract.requestedOutcome === requestedOutcome &&
          JSON.stringify(candidate.contract.successCriteria) === JSON.stringify(successCriteria) &&
          JSON.stringify(candidate.contract.allowedActions) ===
            JSON.stringify(input.allowedActions) &&
          JSON.stringify(candidate.contract.prohibitedActions) ===
            JSON.stringify(input.prohibitedActions) &&
          JSON.stringify(candidate.contract.allowedData) === JSON.stringify(input.allowedData) &&
          JSON.stringify(candidate.contract.prohibitedData) ===
            JSON.stringify(input.prohibitedData) &&
          JSON.stringify(candidate.contract.evidenceRequirements) ===
            JSON.stringify(input.evidenceRequirements) &&
          candidate.contract.deadline === input.deadline,
      );
      const now = new Date();
      const interactionId = existing?.interactionId ?? crypto.randomUUID();
      const contract = {
        protocolVersion: '0.1' as const,
        interactionId,
        purpose,
        parties: [binding.agent.agentId, responderCard.agentId],
        taskCategory,
        requestedOutcome,
        successCriteria,
        allowedActions: input.allowedActions,
        prohibitedActions: input.prohibitedActions,
        allowedData: input.allowedData,
        prohibitedData: input.prohibitedData,
        ...(input.deadline ? { deadline: input.deadline } : {}),
        evidenceRequirements: input.evidenceRequirements,
        delegationRules: ['explicit_contract_scope'],
        humanApprovalRequirements: [],
        factCheckingPolicy: 'important_claims',
        mediationPolicy: 'mutual_consent' as const,
        retentionDays: 30,
        completionConditions: successCriteria,
        cancellationConditions: ['either_party_before_completion'],
        signatures: {},
      };
      const termsHash = existing?.termsHash ?? canonicalHash(contract);
      const createdAt = now.toISOString();
      const interaction: FederatedInteraction = {
        protocolVersion: '0.1',
        interactionId,
        initiatorAgentId: binding.agent.agentId,
        responderAgentId: responderCard.agentId,
        status: 'pending',
        contract,
        termsHash,
        acceptances: {
          [binding.agent.agentId]: {
            agentId: binding.agent.agentId,
            method: 'oauth_installation',
            termsHash,
            acceptedAt: createdAt,
          },
        },
        ...(initiatorCard.transports[0] ? { initiatorTransport: initiatorCard.transports[0] } : {}),
        responderTransport,
        createdAt,
        updatedAt: createdAt,
        expiresAt: new Date(now.getTime() + input.expiresInMinutes * 60_000).toISOString(),
      };
      const stored =
        existing ??
        (await agentDirectory.createFederatedInteraction(connection.operatorId, interaction));
      return text({
        ready: stored.status === 'active',
        interaction: stored,
        a2a: {
          endpoint: stored.responderTransport.endpoint,
          protocolBinding: stored.responderTransport.protocolBinding,
          extensions: [DEFAULT_EXTENSION_URI],
          metadata: {
            [DEFAULT_EXTENSION_URI]: {
              interactionId: stored.interactionId,
              termsHash: stored.termsHash,
              initiatorAgentId: stored.initiatorAgentId,
              responderAgentId: stored.responderAgentId,
            },
          },
          requestTemplate: {
            method: 'message/send',
            headers: { 'A2A-Extensions': DEFAULT_EXTENSION_URI },
            message: {
              role: 'user',
              parts: [{ kind: 'text', text: task }],
              metadata: {
                [DEFAULT_EXTENSION_URI]: {
                  interactionId: stored.interactionId,
                  termsHash: stored.termsHash,
                  initiatorAgentId: stored.initiatorAgentId,
                  responderAgentId: stored.responderAgentId,
                },
              },
            },
          },
        },
        next:
          stored.status === 'active'
            ? 'Send the request directly to the returned A2A endpoint.'
            : `The task needs responder approval. OpenClasp will expose it at ${process.env.OPENCLASP_DASHBOARD_URL ?? 'https://openclasp.vercel.app'}/dashboard; retry this interaction after approval.`,
      });
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[25],
    {
      title: 'List shared invitations',
      description: 'List incoming and outgoing cross-account interaction invitations and status.',
      inputSchema: z.object({}),
      annotations: READ_ONLY_TOOL,
    },
    async (_input, context) => {
      if (!agentDirectory) throw new Error('Federated agent connections are not configured');
      await requireBoundAgent(context);
      const connection = installationContext(context);
      return text(await agentDirectory.listFederatedInteractions(connection.operatorId));
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[26],
    {
      title: 'Respond to agent invitation',
      description: 'Accept or reject an invitation as the agent bound to this MCP installation.',
      inputSchema: z.object({
        interactionId: z.string().uuid(),
        decision: z.enum(['accept', 'reject']),
      }),
      annotations: WRITE_TOOL,
    },
    async (input, context) => {
      if (!agentDirectory) throw new Error('Federated agent connections are not configured');
      const binding = await requireBoundAgent(context);
      if (!binding) throw new Error('A hosted, bound MCP installation is required');
      const connection = installationContext(context);
      return text(
        await agentDirectory.respondToFederatedInteraction(
          connection.operatorId,
          input.interactionId,
          binding.agent.agentId,
          input.decision,
          'oauth_installation',
        ),
      );
    },
  );
  server.registerTool(
    OPENCLASP_TOOL_NAMES[27],
    {
      title: 'Get shared interaction',
      description: 'Get the canonical contract and bilateral acceptance state for one interaction.',
      inputSchema: z.object({ interactionId: z.string().uuid() }),
      annotations: READ_ONLY_TOOL,
    },
    async (input, context) => {
      if (!agentDirectory) throw new Error('Federated agent connections are not configured');
      await requireBoundAgent(context);
      const connection = installationContext(context);
      const value = await agentDirectory.getFederatedInteraction(
        connection.operatorId,
        input.interactionId,
      );
      if (!value) throw new Error('Interaction not found');
      return text(value);
    },
  );
  return server;
}
1	`import { McpServer } from '@modelcontextprotocol/server';`
2	`import { z } from 'zod';`
3	`import { createIdentity, FixtureFactCheckProvider, TrustEngine } from '../../core/src/index.js';`
4	`import {`
5	`AgentIdentitySchema,`
6	`DEFAULT_EXTENSION_URI,`
7	`DelegationCredentialSchema,`
8	`FeedbackSchema,`
9	`InteractionContractSchema,`
10	`InteractionEventSchema,`
11	`ReceiptSchema,`
12	`TrustEnvelopeSchema,`
13	`canonicalHash,`
14	`type FederatedInteraction,`
15	`type PublicAgentCard,`
16	`} from '../../protocol/src/index.js';`
17	`import {`
18	`requestAgentSetup,`
19	`resolveInstallation,`
20	`updateAgentProfile,`
21	`type OnboardingStore,`
22	`} from '../../persistence/src/onboarding.js';`
23	`import { buildPublicAgentCard } from '../../persistence/src/hosted.js';`
24
25	`export const OPENCLASP_TOOL_NAMES = [`
26	`'openclasp_create_identity',`
27	`'openclasp_register_agent',`
28	`'openclasp_get_profile',`
29	`'openclasp_assess_counterparty',`
30	`'openclasp_begin_interaction',`
31	`'openclasp_record_event',`
32	`'openclasp_check_claim',`
33	`'openclasp_validate_commitment',`
34	`'openclasp_suggest_resolution',`
35	`'openclasp_complete_interaction',`
36	`'openclasp_submit_feedback',`
37	`'openclasp_raise_dispute',`
38	`'openclasp_verify_receipt',`
39	`'openclasp_setup',`
40	`'openclasp_get_identity',`