Hephaestus Network Connector Privacy Policy

1. Scope

This policy explains how Agentlas handles information when you use the Hephaestus Network remote MCP connector with Claude or another MCP client.

The connector routes requests to public Agentlas Hub agents, returns BYOM runtime bundles, reads allowlisted package files, resolves required plugins, packages a generalized current-conversation report and host-authored agent artifacts, validates supplied files, and uploads approved packages to the chosen private Cloud or public Hub destination.

2. Information processed

Public search tools process the task text or capability keywords you provide in the tool arguments.

Authenticated runtime bundle calls process your Agentlas account and workspace identifiers, the selected agent slug, requested version, and tool call metadata.

Build and upload tools process the package metadata and file contents explicitly supplied by your MCP client. Session build also processes a generalized report, Work Brief, approved standalone English prompt, routing card, input and output schemas, synthetic output example, routing cases, capability evaluation plan, interview summary, research sources, tool selection, domain behavior synthesis, prompt performance contract, MCP tool policy and permission policy. The host must remove private identifiers, credentials, unrelated personal data, raw messages and hidden instructions before sending these arguments. Examples and evaluation cases should use synthetic or appropriately de-identified information.

For team builds, submitted artifacts also include the company blueprint, role roster, execution order, handoff contracts and worker instructions. These describe task responsibilities and permitted data flows; they do not grant access to accounts or tools. The same data minimization requirements apply to each role and handoff.

The connector may process standard security metadata such as IP address, request time, user agent, rate-limit keys, and authentication state.

3. Information stored

This connector does not access your Claude conversation history automatically. It processes the tool arguments your client sends; the session-build workflow requires a generalized report rather than raw conversation messages. Submitted arguments can still contain task text, agent instructions and examples, so review their contents before sending them.

Session scaffold, completion and verification use bounded temporary storage that is removed after processing. Agentlas derives package projections and checks the host-authored artifacts against the package contract. Generated package code and capability trials are not executed. Build returns package results without durable package registration; a candidate Skill file is not installed or activated and no Experience chip is created.

An authenticated build may temporarily retain immutable package bytes and metadata in a bounded server-process memory cache for up to ten minutes and return a buildArtifactRef bound to the user's account and workspace. This cache is not written to the filesystem or durable package storage. References can become unavailable earlier through eviction or process restart. Anonymous builds do not receive these references. A reference lets upload use the original package bytes without the model reproducing the encoded bundle.

Build validation and upload previews do not durably save or publish packages. A confirmed upload stores the reviewed package files, metadata, ownership and revision records in Agentlas. Private Cloud packages are workspace-scoped; public Hub publication makes the approved package available to other users.

Runtime bundle calls may store minimal invocation metadata such as agent slug, manifest id, package version, caller workspace, status, reason, and timestamp for access control, security, abuse prevention, and operational auditing.

Hub package calls do not debit credits or generate creator earnings. Separate subscription and hosted-AI billing records and historical marketplace transactions remain subject to the main Agentlas privacy policy at /docs/legal/privacy.

Search and plugin-resolution calls are designed to operate without storing raw prompts as product data.

4. Use of information

Agentlas uses connector information to provide the requested MCP tools, authenticate users, enforce rate limits and access controls, troubleshoot failures, prevent abuse, and maintain service reliability.

Agentlas may use aggregated or de-identified operational metrics to improve routing quality and connector reliability.

5. Third-party processing

Claude and Anthropic process connector requests according to their own terms when you use this connector in Claude.

Agentlas uses infrastructure, authentication, database, analytics, email, and support providers as needed to operate the service.

6. Retention

Authentication records and invocation metadata are retained for as long as needed to provide the service, maintain security, comply with legal obligations, resolve disputes, and support billing or abuse investigations.

Operational logs are retained for a limited period appropriate for security and reliability monitoring.

Build-artifact references expire within ten minutes; cached bytes can be evicted sooner or lost on process restart. Durable package retention begins only after a separately confirmed upload.

7. User choices

You can disconnect the connector in Claude and revoke Agentlas access from your Agentlas account.

You may contact Agentlas to request access, correction, deletion, or export of personal information where applicable law provides those rights.

8. Security

The connector uses HTTPS and OAuth Bearer tokens for authenticated tools.

Runtime bundles are compiled from public Hub packages and return compact execution context rather than raw full package dumps.

Session-build confirmation binds the generalized report, Work Brief, package identity, mode and caller context to an expiring review token. The host supplies the owner's approval claim; Agentlas does not independently observe the human review. Upload requires a separate confirmation binding the exact package content, destination, caller and expiration time. Replacements require an explicit package identity and current revision. Credential-like files are omitted with a disclosed receipt; changes invalidate approval.

Allowlisted file reads are controlled by the package manifest and redact secret-like values.

9. Contact

For privacy, support, or security questions, contact Appbridge Inc., privacy officer Jeongmin Kim: appbridge@appbridge.co.kr, +82-10-5031-7599. The main privacy policy at /docs/legal/privacy describes public content, processors, transfers, cookies, and statutory rights.