Claude Code, Codex, Cursor

Your session limit is not the problem. The brief is.

Sessions do not disappear into hard work. They disappear into re-running work that was mis-specified the first time, on a reasoning tier that was never needed. Choose the coding tool you use; Metis decides how to spend its intelligence and writes the prompt.

Run a thought free no signup
metis
What you type
"remove AI filler, make the KPI bar clearer in the PDF report, add more outcomes to the table"

Half-formed is the correct input. That is where the routing decision actually gets made, and it is the state your tool never sees you in.

What your tool gets
PLAN, cleared to execute
Codex regular prompt LOW reasoning continue session one pass, small spend
01Three tasks, not oneContent, layout and data touch different code paths. Bundling them is what produces the eleven-file blast radius and the forty minutes of correction after it.
02"Clean" now means somethingNamed KPIs stay, units shown once, spacing fixed. The vagueness that generates the correction loop is gone before the agent sees it.
03Why low, not maxMechanical once specified. A higher tier costs multiples of the quota for an identical diff. This is the default nobody checks.
04A prompt you pasteBoundaries, acceptance criteria, and the mistake this tool will most likely make on this exact task, named in advance.

Not a prompt beautifier

Plenty of tools will make your sentence more polite and put "you are an expert" in front of it. That is not the problem.

PROMPT IMPROVERS

They take a request you have already decided on and rewrite it more formally. They cannot tell you it should have been three requests, that it belongs in a fresh session, or that it should not be built yet.

METIS

Works on the decision, not the wording. What the task actually is, where it runs, at what tier, whether it is ready at all, and what your tool will get wrong. The prompt is the last step, not the product.

It does not always give you a plan

A confident plan built on a vague thought is the expensive failure, because you find out four corrections later. When the work is not ready, Metis says so instead of guessing.

PLAN

Cleared to execute

Real and scoped. You get the route, the split, the criteria, and prompts you can paste.

INTERROGATE

Answer first, by tapping

Only what lives in your head, never what your repo could answer. Tap the options and the plan builds itself from your answers.

SPEC FIRST

Held on the ground

"Simplify this without losing features" is a design problem in implementation clothing. Metis says so, and tells you what the spec has to pin down.

The calls you are guessing at

Nobody keeps up with which tier suits what, when a goal beats a prompt, or when a fresh session costs less than dragging context. Metis picks, and shows its reasoning.

Which reasoning levelMatched to real difficulty, not to how important the task feels. A high tier on a mechanical edit burns multiples of your quota for an identical diff, and it is the most common way a session ends early.
New session or continueLong sessions do not ignore your early instructions, they stop retrieving them. Metis calls it per task, and when it says new, you get the handoff note to carry over.
Prompt, plan mode, or goalA verifiable exit condition means a goal, with criteria written and the gaming clauses fenced. No exit condition means it is not a goal yet, whatever you were about to type.
What to attach, and what not toScreenshots for visual work, never a raw log dump. Reference designs get transplanted literally rather than matched, which is the single most expensive mistake in agent coding.
Where it will go wrongEvery prompt names the mistake your tool will most likely make here, and the guard against it. Predicting the failure is cheaper than correcting it.
Your codebase, specificallyTen lines about your project, once. It stops asking what it could already know, names your real files, and guards the traps your tool keeps falling into.

Free while it proves itself

No checkout, account, or subscription. The beta is intentionally small until repeat use proves it deserves more infrastructure.

Public beta

€0
  • 3 successful engine calls per day
  • PLAN, INTERROGATE, and SPEC FIRST
  • Project context and Tool notes
  • Brief history saved in this browser
  • No payment details
Open Metis

Failed requests do not use the allowance. The counter resets at 00:00 UTC.

Questions

Does my code get sent anywhere?

Metis never reads or connects to your repository. It sends the thought you type, plus any selected Project context and Tool notes, to Cloudflare Workers AI to produce the brief. Do not paste credentials, secrets, personal data, or confidential source code.

Why not just use plan mode?

Plan mode plans the request you already gave it. Metis works on whether that was the right request, on the right tool, at the right tier, in the right session. It runs before the tool opens, which is also why it can tell you not to build something yet.

Is this a model router?

No. Routers sit in the request path and classify traffic a machine already understands. Metis sits in front of you, on work that is not a request yet, where the intent is still in your head and the wrong call costs a rerun rather than a few cents.

Will it be wrong sometimes?

Yes. It is judgement, not arithmetic. The reasoning is shown for every call so you can overrule it, and tiers and limits change often enough that it flags where you should confirm against current docs rather than pretending to know.

Do I need to write good input?

The opposite. Lazy, fragmentary and unpunctuated is the intended input. If a thought is too vague to route, you get tappable questions rather than a confident guess.

Privacy, in plain language

Metis is built to need little data and no account.

Saved on this device

Your briefs, Project context, Tool notes, and preferences are stored in this browser. “Delete local data” removes them from the device.

Sent for each run

The thought and any selected context or Tool notes are processed by Cloudflare Workers AI. Metis does not send your repository because it never has access to it.

No tracking profile

There are no accounts, ads, analytics, or tracking cookies. To enforce the free allowance, Metis stores short-lived, bucket-specific hashes derived from the request IP—not the raw IP. An optional useful/not-useful vote is tied only to an anonymous run token. Operational records are automatically pruned.