Manual server proof. Two clean commit worktrees, vp i, T3CODE_DEV_SEED=0 vp run dev:server --port --no-seed. Apply seed-driver.py before starting a devReplay thread. Add project with t3 project add, start one devReplay/showcase thread with prompt Populate synthetic history, --json into -start.json. The seed driver emits 3200 completed synthetic reads and 32 message chunks via ACP; server persists 3234 timeline items. It retains the running agent session for 10 minutes. Wait for seeded- and server projection catch-up, then node read-driver.mjs . MCP session config contains credentials: keep private. Read script calls real /mcp via the agent bearer. For Claude Code create private HTTP MCP config from that endpoint and authorization. Run claude -p asking exactly one t3_thread_read(threadId,view:activity,limit:100,maxCharsPerItem:1), --model haiku --strict-mcp-config --mcp-config --allowedTools mcp__proof__t3_thread_read --max-turns 3 --output-format stream-json --verbose. Published transcripts omit private local file paths; response JSON is synthetic. No application service implementation was changed. Stop the server via its captured process session.