Anthropic's June 15th pricing change is not just a billing update. It changes where Claude personal assistants should live.

If you want a Claude-powered assistant that runs beyond a single chat session, you now have two serious paths.

1. Managed Agents: hosted, scalable, API-native, and priced like infrastructure.
2. Local always-on Claude Code instance: close to your files, tools, repos, permissions, memory, and personal context.

Tools like Hermes Agent, OpenClaw-style assistants, T3 Code, Zed's Claude Agents, or OpenCode wrappers when they route unattended Claude work through subscription-authenticated Agent SDK usage or claude -p, now live inside the new Agent SDK credit model. That can mean cost increases of 12x to 25x typically, with extremes near 175x.

What changes on June 15th

Starting June 15th, 2026, Claude Agent SDK usage and claude -p no longer draw from the same subscription bucket as interactive Claude usage.

They move to a separate monthly Agent SDK credit.

That credit covers:

• Claude Agent SDK usage in your own projects
• claude -p / claude --print
• Claude Code GitHub Actions
• Third-party apps authenticated through the Agent SDK

It does not cover:

• Interactive Claude Code
• Claude chat
• Claude Cowork

Interactive Claude stays subscription-shaped. Unattended programmatic Claude becomes credit-shaped, then extra-usage/API-priced once the credit is gone.

Reported effective cost increases land in the 12x–25x range for typical programmatic workloads, with extreme cases (heavy harness usage on a Pro subscription) quoted as high as 175x. The right framing is not "the bill went up." It is "the subsidy went away."

Monthly Agent SDK credit by Claude plan tier, effective June 15th, 2026.

The local alternative: always-on Claude Code instance

An always-on local Claude Code instance keeps the assistant alive locally, with durable state in files and inbound channels or terminal sessions. That is the future. Run it under Docker so it survives reboots and stays available between sessions.

Claude Code already provides much of the substrate a personal assistant needs:

• Authentication and model access without a separate API key
• A permission-gated tool layer (Read, Write, Bash, Grep, Edit)
• Hooks at session start, tool pre/post, and session end
• Subagents with isolated context windows, so heavy reads do not poison the main session
• MCP server integration, skills, and a plugin system with marketplace
• CLAUDE.md as canonical project policy, read on every session
• Auto-memory keyed to the project working directory, persistent across sessions
• Session persistence, resume, and fork

A plugin loaded into that instance adds the operating layer on top: scheduled routines, background watches, a proposal pipeline that gates mutation, and channel-based notification. That is what turns the substrate into the scenarios below.

What that looks like in practice:

• An assistant that wakes at 7am, summarizes overnight Slack, and posts the digest to a Discord / Telegram DM
• A heartbeat that checks CI every twenty minutes, stays silent when green, and pings the operator when a job fails with the failing test pre-quoted
• A routine that picks gh issues when idle and tackles them

---

Disclosure: I maintain claude-code-hermit, an open-source Claude Code plugin that turns CC into an always-on personal AI assistant. The patterns described here are not theoretical for me.

Sources

• Anthropic Agent SDK credit: https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan
• Claude Code programmatic usage: https://code.claude.com/docs/en/headless
• Claude Code CLI reference: https://code.claude.com/docs/en/cli-reference
• Claude API pricing: https://platform.claude.com/docs/en/about-claude/pricing
• Hermes Agent (NousResearch): https://github.com/NousResearch/hermes-agent
• VentureBeat on the OpenClaw reinstatement and credit policy: https://venturebeat.com/technology/anthropic-reinstates-openclaw-and-third-party-agent-usage-on-claude-subscriptions-with-a-catch
• Effective price-increase math (MagnaCapax gist): https://gist.github.com/MagnaCapax/d9177e35b355853f03c730dfcaa693ef