Skip to main content
Glama

Dials — a behavior control panel for Claude

Give Claude a set of 0–10 sliders — rigor, verification, verbosity, autonomy — that it reads over MCP and actually follows. Tune how Claude works from the outside, as data, instead of re-typing "be more thorough / just ship it / double-check everything" in every chat. The settings persist across sessions and work in Claude Code, Claude Desktop, and claude.ai.

Free. MIT. Zero dependencies.

On claude.ai? Get a free hosted connector (no signup) at https://dial.kestudios.dev — generate a token, paste the endpoint + token into a custom connector, done.

In Claude Code / Desktop? One line:

claude mcp add dials -- npx -y github:willykeenan/claude-dials

Why

How Claude behaves is normally buried in prose — system prompts, CLAUDE.md, whatever you remember to say. Prose can't be dialed. Dials turns the tunable dimensions into numbered sliders exposed as MCP tools + a resource. Slide one, and every task after it changes. That's a control plane for the model, not a notes file.

Related MCP server: claude-orator-mcp

The dials

Dial

0

10

Rigor

quick first-pass

exhaustive reasoning

Context depth

act on the prompt alone

read the full surrounding context first

Verification

assert it's done

run it and show evidence

Self-review

ship the first draft

adversarial critique before delivering

Verbosity

one-liner

exhaustive explanation

Autonomy

ask before acting

just do it, report after

Novelty

proven patterns only

bleeding-edge

Caution

move fast

treat as if it'll be reviewed

Each value maps to a concrete instruction (see dials://schema), so Claude gets an unambiguous directive, not a vibe.

Presets

  • default — balanced baseline

  • fable — maximum rigor: read everything, verify with evidence, self-critique; terse replies, meticulous work

  • ship — fast and autonomous, for low-stakes velocity

  • careful — high-stakes: verify everything, ask before acting, max caution

  • explore — research/prototyping: reach wide, try new things

Tools

get_dials · set_dial(dial, value) · apply_preset(preset) · list_presets · explain_dial(dial) · reset_dials

And a resource dials://current (the live settings as Markdown) + dials://schema (full JSON), plus a load-dials prompt.

Install

Client

How

Claude Code

claude mcp add dials -- npx -y github:willykeenan/claude-dials — details

Claude Desktop

add the snippet in configs/claude_desktop_config.json

claude.ai

free hosted connector — generate one at dial.kestudios.dev, or self-host src/http.mjs (details)

Drop CLAUDE.md into a repo to make Claude read your dials at the start of every task.

Use it

"get my dials" · "set rigor to 9" · "apply the careful preset" · "explain the autonomy dial"

Or just talk: "double-check everything from now on" → Claude bumps verification.

Design notes

  • Zero dependencies. The MCP wire protocol (JSON-RPC 2.0 over stdio + Streamable HTTP) is implemented directly. No SDK, no supply-chain surface, instant npx.

  • State lives at ~/.claude-dials/state.json (override with DIALS_STATE_FILE). Corrupt or missing state falls back to defaults — a config tool must never take down the client.

  • Remote auth. The HTTP transport takes an optional DIALS_TOKEN bearer token; unset = open (local dev only).

Verify it yourself

npm test

Spawns the real server over stdio and drives a full JSON-RPC handshake — 30+ assertions across initialize, tools, resources, prompts, persistence, schema enforcement, and error paths. Also: npm run eval (mock obedience gate) / npm run eval:live (real Claude, costs tokens).

License

MIT — built by KE Studios. https://dial.kestudios.dev

Available Tools

6 tools
apply_presetB

Apply a named preset. One of: default, fable, ship, careful, explore.

Inputs: preset (required).

explain_dialC

Explain what a dial means and how each value band changes behavior.

Inputs: dial (required).

get_dialsA

REQUIRED at the start of every task: load the user's binding behavior dials (rigor, verification, verbosity, autonomy, …) and follow them exactly for the rest of the task. Higher is not always better.

Inputs: no named parameters.

list_presetsA

List the available presets and what each one does.

Inputs: no named parameters.

reset_dialsA

Reset all dials to the balanced default preset.

Inputs: no named parameters.

set_dialA

Set one behavior dial to a value 0–10. Persists across sessions.

Inputs: dial (required), value (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool has a distinct purpose: applying presets, explaining dials, loading dials, listing presets, resetting, and setting individual dials. No overlap.

    Naming Consistency5/5

    All tools follow a consistent verb_noun pattern (e.g., apply_preset, explain_dial, get_dials), making them predictable.

    Tool Count5/5

    Six tools is ideal for a dial and preset management server; neither sparse nor bloated.

    Completeness4/5

    Covers all major operations: get, set, reset, preset apply/list, and explanation. Missing a granular 'get single dial' but get_dials covers it.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      C
      quality
      D
      maintenance
      An MCP server that enables persistent memory, structured thinking sessions, and project-based knowledge management for Claude. It includes specialized coding tools for package discovery and reinvention prevention by validating code against existing libraries and APIs.
      17
      3
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      An MCP server for deterministic prompt optimization in Claude Code. Score prompts across 7 quality dimensions, auto-select from 11 Anthropic techniques, and return a structural scaffold.
      1
      29 npm
      2
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      An MCP server for Claude Code that catches vague prompts, applies triage and correction pattern learning, and provides semantic search, cross-service contracts, and scorecards to reduce wasted tokens.
      13
      MIT