Kill the bloat in Claude Code's system prompt, without turning a single tool off.
A lightweight Claude mod that strips the bloat from the system prompt without turning any tools off.
less-bloat.mp4
One install -> ~50% less tokens in your system prompt
| Tokens to start a session | Without | With less-bloat |
Savings (%) |
|---|---|---|---|
| CLI, out-of-the-box | 29,954 | 15,173 | 49.3% |
| Desktop app, out-of-the-box | ~61,900 | ~35,000 | ~44% |
Want to see it work? Ask Claude:
Start a new interactive Claude Code session with
less-bloat(https://github.com/jakobbarnwell/less-bloat) installed and one without it and send "hi" in each. Wait for the reply, then compare the first request's total input tokens in their transcripts (cached ones included). Turnless-bloatback on after, and tell me the results.
Recommended: Paste this into Claude Code, CLI or desktop:
Install the
less-bloatplugin from thejakobbarnwell/less-bloatmarketplace. Then, based on how I use Claude Code, tell me whether I should use default mode, and whether to turn back on any tools, MCP servers or connectors I turned off to save context. Call tools "described up-front" or "name-only", as the less-bloat README does.
Or install it yourself, in Claude Code:
/plugin marketplace add jakobbarnwell/less-bloat
/plugin install less-bloat@less-bloat
Or from a terminal
claude plugin marketplace add jakobbarnwell/less-bloat
claude plugin install less-bloat@less-bloatIt takes effect from the next session. Needs Claude Code 2.1.286 or later.
To uninstall
claude plugin uninstall less-bloat@less-bloatClaude works less well as context fills up.
- Most tools are name-only: the system prompt mentions only their name. This is by design: when Claude wants to use one, it fetches the full description with ToolSearch.
- But Claude Code describes some tools up-front: their name, description and schema are part of every session's system prompt, whether you normally use these tools or not.
less-bloatsimply takes most tools that are described up-front and makes them name-only too. They still work. Claude reaches for their description when needed.
If less-bloat moves a tool's full description out of the system prompt, it retains a small hint: this ensures Claude still knows what the tool does and when to fetch its description.
Default keeps these described up-front: Bash, Read, Edit and Write, Glob and Grep where Claude Code has them, Agent, Skill, AskUserQuestion and SendUserFile, plus the desktop app's chapter, side-task and widget tools when you're in the app, and a project thread's reply tools.
Custom is your own picks. Run /less-bloat to open settings: check a tool to describe it up-front, or uncheck one to make it name-only.
- In doubt? Ask Claude to set things up based on your session history.
- Wondering what a tool does? Ask Claude.
Your choice holds for every session, CLI and desktop alike.
A change applies from your next conversation: a new session or /clear.
- Fewer tokens every turn. The system prompt goes out with every request. Cutting it makes each turn cheaper, stretches your usage limits and leaves more room before compaction.
- More of Claude's attention on your work. Claude works less well as its context fills up. Less of that context goes to describing tools you don't need.
- Nothing is turned off. Settings can remove tools, but can't make a built-in tool or the desktop app's own servers name-only. Here every tool still works: Claude fetches its full description with one ToolSearch call the first time it needs it.
- No setup. The default works as is.
Does it break the prompt cache?
No. Each tool's placement is fixed for the conversation, and a change applies from the next one. See Limits for the one case that costs a request its cache.
Will Claude get worse at using tools?
No. The everyday tools stay described up-front, and the rest load when needed. Claude reaches for a name-only tool on its own when its name says what it's for. Keep a tool described up-front when its name doesn't, as Agent's doesn't say when to delegate. ToolSearch is always described up-front, since Claude fetches all the others with it.
Is it safe?
A plugin can read and write files, run commands and go online as you, so check what you
install. less-bloat is about 600 lines of TypeScript in hooks/, and does none of these.
(scripts/check.py and scripts/bench.zsh are a test and a benchmark you run yourself; the
plugin never runs them.) It keeps your choices,
and which tools it has told you about, in Claude Code's own plugin store.
Does this work for normal Claude, or just Claude Code?
Just Claude Code: the CLI and the desktop app's Code tab. Chats on claude.ai or in the Claude apps don't run Claude Code plugins.
How do I see which tools are described up-front?
Run /less-bloat. Its pane shows what new
conversations get. To see this conversation's, ask Claude. Where the pane can't open, such as in
claude -p or without ToolSearch, /less-bloat prints the list instead.
Does a name-only tool still cost tokens?
Only its name, in Claude Code's list of name-only tools, and the first sentence of one that asked for its full description. The rest of its description and its schema come in when Claude fetches them.
How does less-bloat work with subagents?
The same way. Each subagent type comes with its own system prompt and set of tools, and less-bloat places those tools as it does yours. What you save depends on the type: a general-purpose subagent's first request is about half the size, like the main conversation's. Explore and Plan have no tools to trim, and cost about the same as without it. Each subagent gets hints only for the tools it has. If you run subagents with tools allowed but ToolSearch off, less-bloat won't work and all your tools will be fully described in the system prompt up-front.
How do I turn a tool off completely?
less-bloat makes tools name-only; it doesn't turn any off.
To remove one:
- Deny it in
~/.claude/settings.json, which the CLI and the desktop app both read. A tool's name removes that tool,mcp__<server>every tool of a server:"permissions": { "deny": ["WebSearch", "mcp__playwright"] }. - Turn an MCP server off with
/mcp, or remove one you added withclaude mcp remove <name>. - In the desktop app, turn connectors off in Settings → Connectors.
I turned tools off to save context. Should I turn them back on?
If you'd rather have them name-only than gone, yes: remove them from permissions.deny, or turn the
MCP servers or connectors back on. They're name-only, unless your mode keeps one described up-front.
Ask Claude to set up less-bloat, and it checks your settings for tools you turned off. It can't see
connectors turned off in the desktop app: check Settings → Connectors.
What about skills?
Skills add bloat too: every installed skill (with model invocation enabled) has its name and "when-to-use" frontmatter in the system prompt too. Run /skill-doctor to see which skills you use, and disable the rest.
Will I know what it changed?
Your first session shows a toast with how many tools less-bloat
made name-only that Claude Code would describe up-front, and a line in the transcript pointing to
/less-bloat. After that, you get one only for a new tool, such as a newly added MCP server's, once.
What happens when I add new MCP servers or connectors?
Almost all added tools are name-only in Claude Code anyway,
so nothing changes. If you install one that asks for its full description in the system prompt, less-bloat makes it name-only too, with a notice.
Does /clear reset less-bloat?
The new conversation starts fresh. less-bloat keeps each
earlier conversation's placements in memory until Claude Code exits, so /resume back to one keeps
its prompt cache.
What happens when a conversation compacts?
Nothing changes for less-bloat. Compaction
summarizes the messages, but the tools at the start of each request stay as they were, so their
cache still holds. Claude Code lists the name-only tools again, and less-bloat their first sentences (the hint).
- If you fully quit a session, change your
less-bloatsettings, and reopen (/resume) it within the cache's lifetime, the cache gets invalidated. - A name-only tool's first use takes a ToolSearch call to fetch its full description. That's why the basic and most used tools stay described up-front. If you use a tool in every session, consider putting it up-front.
- An MCP server that connects late (after your first message) isn't placed by
less-bloatin that conversation. Claude Code adds its tools in a message instead, which keeps the prompt cache: described up-front if they ask for their full description, name-only otherwise. - Needs ToolSearch. Claude Code turns it off with
ENABLE_TOOL_SEARCH=false, and by default behind a customANTHROPIC_BASE_URLsuch as a gateway. Then every tool is described up-front. - Not in cloud sessions, WSL, normal Claude chats, or Cowork, which don't load your plugins. In a desktop app's Projects, threads that run on your machine load it; cloud threads and the project's coordinator don't.
Inspired by Matt Pocock's thread and article on killing the bloat in Claude Code's system prompt.
MIT