isphere-ai-bridge/docs/mcp-core-business-plan.md

MCP Core Business Plan

Date: 2026-07-09

Business correction

项目核心不是继续做 UIA selector/report，而是通过 MCP 暴露 iSphere 的联系人、群组、消息、文件能力。

后续研发主线切换为 MCP 核心通信能力：搜索联系人、搜索群组、收发消息、收发文件。现有 UIA/selector/report 只作为底层辅助和证据，不再作为主线继续扩展。

Core business capabilities

The product roadmap is limited to six business capabilities:

1. Search contacts.
2. Search groups.
3. Receive/read messages.
4. Send messages.
5. Receive/download files.
6. Send files.

Everything else is infrastructure or fallback support. Selector catalog/report validation may support UI automation if needed, but it is not the main product deliverable.

Capability source priority

Use this priority order when searching for real implementations:

1. Existing bridge / API / local service.
2. Local database / log / cache read-only access.
3. UI automation.
4. Network/API analysis.
5. Reverse engineering, only for explanation and validation, not as the default route.

Implications:

• Prefer an existing bridge/API/local service before reading local storage or using UIA.
• Prefer read-only database/log/cache access before UI automation when it can answer the business question.
• UIA selector work is a fallback access mechanic, not the primary roadmap.
• Do not continue investing in selector report hardening unless a concrete business tool is blocked by selector quality.
• Reverse engineering is last resort and should explain formats or validate assumptions; it should not become the default delivery path.

Stage A: Freeze MCP core tools contract

Goal: lock the business-facing MCP tool surface before adding more helper or selector work.

Deliverables:

• docs/mcp-core-tools-contract.md exists and defines:
  • isphere_search_contacts
  • isphere_search_groups
  • isphere_receive_messages
  • isphere_send_message
  • isphere_receive_files
  • isphere_send_file
• Write-capable tools require approval_id for real execution.
• Write-capable tools support dry_run=true and default to dry-run until approved.
• Read-only tools do not mutate the iSphere client state.

Acceptance:

• Contract names, input parameters, output JSON, error codes, audit fields, read-only/write status, approval behavior, and implementation priority are explicit.
• No new selector/report node is introduced as the next primary step.

Stage B: Find real capability sources

Goal: identify the best source for each read-only capability.

Tasks:

1. Inventory existing bridge/API/local service artifacts and current Go/C# boundaries.
2. Re-check whether existing bridge endpoints or local services can provide contacts, groups, messages, or files.
3. Re-check local database/log/cache artifacts only in read-only mode.
4. Use UIA only if higher-priority sources cannot satisfy the read-only tool.
5. Document source decisions per tool before implementation.

Acceptance:

• Each read-only tool has a selected primary source and a fallback source.
• The decision is evidence-backed and does not rely on stale prior notes without current verification.
• No automatic login, hook, injection, memory read, or endpoint-security bypass is introduced.

Stage C: Read-only loop

Goal: deliver useful MCP read capability before any real write capability.

Implementation order:

1. isphere_search_contacts
2. isphere_search_groups
3. isphere_receive_messages
4. isphere_receive_files

Acceptance:

• Each tool can return normalized JSON through MCP.
• Each tool records audit metadata.
• File receiving/download writes only to an approved local output directory and never mutates source client state.
• Tests and verification do not send messages, upload files, click destructive controls, or alter production data.

Stage D: Approval-after-read-only write capability

Goal: add write behavior only after the read-only loop is proven.

Implementation order:

1. isphere_send_message with dry_run=true preview first.
2. Approval store and audit validation.
3. isphere_send_message real send only with valid approval_id.
4. isphere_send_file with dry_run=true preview first.
5. isphere_send_file real send only with valid approval_id.

Acceptance:

• No valid approval_id, no real send.
• Every real send records target, content/file hash, approver, time, connector, and result.
• Idempotency prevents accidental duplicate sends.
• Operator can review dry-run output before approval.

Stage E: Pre-production safety, audit, and deployment acceptance

Goal: make the MCP service safe and operable before broader use.

Checks:

• Tool allowlist and source allowlist are explicit.
• Audit records are durable and redacted where needed.
• Approval records are traceable.
• Deployment docs include exact commands, rollback, logs, and known limits.
• Security boundaries are still enforced: no automatic login, no bypass, no injection, no hook, no target-process memory read.

Acceptance:

• Full test suite passes.
• Operator runbook covers normal operation and failure modes.
• A business user can understand which of the six capabilities are supported and which remain staged.

Immediate next action

Current next step is:

MCP core tools contract freeze + read-only message source discovery + N12-pre evidence package organization.

Do not start N16/N17 selector/report hardening as the mainline. If UIA is needed later, use the existing N13/N14/N15 results as auxiliary evidence only.