How it works
Plan, run, prove.
Pick your depth
Short version
Two or three sentences per part, no jargon.
Under the hood
Which component hands what to which, for builders.
Exact values
Formats, limits and numbers, for anyone checking our claims.
The idea
A plan first, a record after.
The sections that follow zoom into this one line.
Short version
You describe a task. Cortiq writes it as steps, runs them, asks you before anything acts outside the run, and gives you a file you can prove later.
Under the hood
The planner produces a JSON plan from a fixed toolset. The runner executes it step by step, pausing at action steps for a decision. The finished run is hashed so a short fingerprint can be anchored on chain.
Exact values
- Plan format
- JSON, ≤ 24 steps
- Tools
- 13 built in
- Hash
- keccak-256
- Chain
- Base · 8453
Architecture
Where each part runs.
Your browser does the work; this site relays reads; Base keeps receipts.
Short version
Almost everything happens on your own device. Our server only passes along requests to read public chain data, and the chain only sees what you choose to anchor.
Under the hood
Read tools call this site's relay, which forwards a short allowlist of read-only JSON-RPC methods to Base and rejects anything that could write. Compute and action tools run in the page. Wallet transactions are signed in your wallet, never on the server.
Exact values
- Relay methods
- Reads only
- Signed transactions
- Never relayed
- Run storage
- Your browser
- Server logs of runs
- None
Reasoning
Rules by default, AI by choice.
Two planners, one set of checks.
Short version
The built-in planner reads your words and picks steps from templates. If the site has an AI planner switched on, you can use that instead. Either way you see and edit the plan before it runs.
Under the hood
The rule planner looks for addresses, transaction hashes, quoted text, numbers and verbs, then assembles fixed templates. The AI planner receives only the task text and a description of the tools; its answer is parsed and rejected if it names a tool that does not exist.
Exact values
- Rule planner
- In the browser
- AI planner
- Optional
- AI sees
- Task text only
- Model price cap
- $1 / $4 per 1M
Approval
Nothing acts without a yes.
The gate sits between the plan and the world outside the run.
Short version
When a step would save a file, copy to your clipboard or ask your wallet to sign, the run stops and shows you exactly what it will do. You approve or reject; both are written down.
Under the hood
Action tools are marked in the toolset. The runner fills their inputs, sets the step to 'needs approval' and waits on a promise that only your click resolves. A rejected action is logged as rejected and skipped; the rest of the run continues.
Exact values
- Action tools
- 3
- Decision logged
- Always
- Timeout
- None, waits
- Funds movement
- Not possible
Receipt
A fingerprint on Base.
Proof that a run file existed, unchanged, at a given block.
Short version
Your finished run becomes a file. Two short fingerprints of that file can be written to Base from your wallet. Later, anyone with the file can check the fingerprints still match.
Under the hood
Plan and log are serialised as canonical JSON (keys sorted, no spaces) and hashed with keccak-256. CortiqRunReceipt stores both hashes with the step count, the approved-action count, the block and the sender. The verify page recomputes the hashes and looks up the receipt.
Exact values
- Contract
- CortiqRunReceipt
- Steps per receipt
- 1 to 64
- Gas, first anchor
- ~210,000
- Owner / fee
- None