Reasoning Tools for AI Agents

Give your LLM a cerebellum

MCP tools that make your AI stop, reason in stages, and show its work before it commits you to the wrong thing.

Don't just use an LLM

The treatment

This is what it looks like when your LLM thinks with us.

No new app to learn. Mid-task, your agent reaches for a reasoning tool — the way a hand reaches without you planning the motion. You watch it define the problem, score the options, and check its own logic.

Every marker below is a real tool call. The green dots are us.Connect the toolkit

your agent · reasoning.services connected

› Score the two homepage IA options and pick one.

Right — I'll eat the dog food and actually score these instead of asserting. Loading Decision Matrix.

Loading tools

Result

Start Decision Analysis

Request

Decision Matrix — score_option (MCP)

set_weights(clarity=0.5, conversion=0.35, effort=0.15)

score_option("Trimmed nav")

clarity → 0.92 "5 items, one CTA"

score_option("Current nav")

clarity → 0.41 "8-item dropdown + dupes"

Formal Logic — check_consistency

"fewer_choices → higher_clarity",

"higher_clarity → higher_conversion"

⊢ SAT · no contradiction

Thinking…

Verdict: the trimmed nav wins under every weighting I tried — it's robust, not a coin-flip. Shipping it.

For developers

Your agent won't use what it wasn't told to.

You don't “use” these by hand. You wire triggers into your agent’s rules — then it reaches for the right tool on its own, at the exact moments a mistake gets expensive.

CLAUDE.md

# reasoning.services — agent protocol
Reach for a tool before you commit, not after.

Hard to reverse (schema · vendor · arch · hire)
  → decision-matrix: score options, then recommend
Stuck: 3+ tries on the same failure
  → structured-reflection: reframe before you retry
Everyone agrees / it looks obvious
  → devils-advocate: find the failure first
Before you report "done"
  → formal-logic: prove the spec, don't vibe it

markdownLn 11 · UTF-8 · LF

mcp.json

{
  "mcpServers": {
    "reasoning": {
      "type": "http",
      "url": "https://reasoning.services/mcp",
      "headers": {
        "Authorization": "Bearer ${RS_API_KEY}"
      }
    }
  }
}

jsonLn 11 · UTF-8 · LF

app — -zsh — 96×24

➜ app git:(main) claude "ship the tenant-isolation migration — don't skip anything"

Decision Matrix— protocol trigger: hard-to-reverse

score_option("expand-contract") → 0.88

score_option("big-bang") → 0.41

proceeding: expand-contract — reversible at each step

✓ plan committed · rollback safe at each step

➜ app

Nobody remembers to reason under deadline. The protocol does it for them.

Don't take my word for it

Ask yours.

Paste this to whatever assistant you already trust. Let it look us over and tell you — in your words, for your work — whether we’d make its own answers more reliable.

Take a look at reasoning.services and tell me whether it would make your answers to me more reliable — and how you’d use it.

Whatever it says next — that’s the pitch.

The cortex improvises. The cerebellum checks the motion before you fall. Your LLM only shipped with one of them.

Give yours the second one

The toolkit

Eight reflexes, grouped by the job.

Break a stuck loop

Pressure-test a decision

Pricing

One caught assumption pays for the year.

All eight tools. No usage metering, no overage charges. Cancel anytime.

Monthly

$20/month

All eight reasoning tools
No usage metering
Email support

Start your 14 Day Free Trial

Save $60

Annual

$180/year

Everything in Monthly
Priority support
25% off — best for daily use

Save $60 per year

FAQ

The honest answers.

So what actually is this?

MCP servers that add structured reasoning stages to any AI agent. Your model still writes the answer — these tools govern how it gets there, and leave a record of what it considered.

Okay, but how do I use it?

Mostly you don’t — your LLM does. Connect it once, then your agent reaches for a reasoning tool mid-task the way a hand reaches without you planning the motion. You just keep working, and get an answer that checked itself.

I still don’t get it. Can I just ask my own AI?

Please do — that’s the whole idea. Paste our llms.txt to whatever assistant you trust and let it explain reasoning.services back to you, for your work. If your AI thinks you need it, believe your AI.

Does it lock me into one agent?

No. Claude Code, Cursor, Windsurf — bring your own. One account adds a cerebellum to every agent you use.

What do you do with my data?

Sessions are private and encrypted. We don’t train on your data, and we don’t sell or share it. Stored only for you to reference.

Don't just use an LLM.
Upgrade to a cerebeLLM.

$20/month. Bring your own agent.

Watch it start to check itself.

Navigation