A plugin for Claude Code and Cowork that lets your main session hand work to the cheapest model that can do it well. A file search goes to Haiku. A gnarly refactor, or a careful read of conflicting reports, goes to Opus. An architecture or strategy call goes to Fable.
It ships two things:
- Four subagents, one per model tier.
- Routing rules that tell Claude when to use each one. A session-start hook loads them into every session, so you don't paste anything into your
CLAUDE.md.
What it runs. The only command is the session-start hook: cat "${CLAUDE_PLUGIN_ROOT}/rules/delegation.md". It reads one file from inside the plugin and prints it into the session. The plugin sends nothing anywhere, fetches nothing, and collects no data. The subagents use the same tools and permissions your session already has.
| Agent | Model | Effort | Use for |
|---|---|---|---|
tierwise:quick-task |
Haiku | medium | File and folder lookups, grep/search, reading documents and web pages, pulling out facts, batch lookups in one script, doc lookups, quick Q&A, status checks, trivial one-liner edits |
tierwise:standard-task |
Sonnet | high | Drafting documents/emails/reports, summaries, proofreading, spreadsheets, routine data work, research with a clear question, code edits, tests, refactoring, clear-spec implementations, docs |
tierwise:deep-task |
Opus | medium | Complex debugging, multi-file refactors, code review, security, perf, analysis of sources that disagree, data analysis whose figures feed a decision or get published, thorough document review — the default for heavy reasoning |
tierwise:architect-task |
Fable | high | Architecture, strategy, hardest debugging, high-stakes tradeoffs — ceiling-level, when deep-task isn't enough |
The full routing rules are in rules/delegation.md. Beyond "pick the cheapest tier that works", they cover how to brief a subagent, how to verify what comes back, when to escalate, and when to use a Workflow instead of dispatching agents by hand.
Renamed. This plugin was called
claude-tiered-agents, thentiered-agents, and its marketplace wastiered-agents-marketplace. If you installed an old name, remove it first. In Claude Code, run/plugin uninstall <old-name>@tiered-agents-marketplaceand then/plugin marketplace remove tiered-agents-marketplace. In Cowork, go to Customize → Plugins, remove the old plugin, and remove the old marketplace. Then installtierwisebelow. Old copies no longer get updates, and keeping one alongside the new one loads the rules twice.
In a Claude Code session:
/plugin marketplace add bejaminjones/tierwise
/plugin install tierwise@tierwise
Then start a new session (or run /clear) so the routing rules load.
Auto-update is off by default for marketplaces you add yourself. To turn it on, open /plugin, go to Marketplaces, select tierwise, and choose Enable auto-update.
To update by hand instead:
/plugin marketplace update tierwise
- Open Customize in the sidebar, then Plugins.
- Select Add, then Add marketplace, and enter
bejaminjones/tierwise. - Select tierwise, then Add.
- Start a new task so the routing rules load.
To get updates, select Check for updates on the marketplace, or turn on Sync automatically.
A plugin you install in Cowork is saved to your claude.ai account. It also reaches Claude Code the next time you start a session signed in to that account (Claude Code v2.1.273 or later; run /reload-plugins if it hasn't appeared), so you don't need to install it twice. This only works one way: a plugin installed from Claude Code's command line doesn't reach Cowork.
- Model access.
architect-taskruns on Fable,deep-taskon Opus. If your plan doesn't include a model, that agent won't run. Edit itsmodel:line in a local copy, or ask Claude to use the next tier down. - macOS or Linux. The session-start hook uses
cat.
Claude Code: /plugin → Installed → tierwise → Disable.
Cowork: Customize → Plugins → tierwise → Disable plugin.
Each switch turns off both the agents and the routing rules, but only for the copy installed there. If you installed it in both Claude Code and Cowork, turn it off in both.
MIT