CODE MODE FOR JEV
Let it check more before it answers.
System One lets an agent call Jev inside a short program. It can rank evidence, branch on a result and check the next item without another agent turn. One MCP call returns the findings it needs.
Hosted Code Mode is in free early access with daily limits.
Email sign-in · No payment card or waitlist · Direct tools also available
The agent writes a bounded plan over relevant evidence.
Run a short plan with conditions and loops.
Every Jev call uses this connection's scopes and allowance.
Return a focused result to the agent, which keeps the final action.
These are service limits, not measured speed or cost savings. Hosted Code Mode has a separate daily execution limit.
YOUR DECISION CONTROL CENTER
Compare request volume and response times. Inspect uncertain results and identify the connection that made each call. Set limits or pause the engine from the same workspace.
Actual interface with demonstration fixture data. Your app's charts show its own receipts.
ONE PLAN · SEVERAL BOUNDED JUDGMENTS
The agent supplies evidence, calls Jev only where judgment helps, and uses ordinary conditions and loops to decide what to check next.
async () => {
const excerpts = [
{ id: "callback", text: "OAuth callback returned 504 after 12s." },
{ id: "database", text: "No database errors in this request." }
];
const match = await jev.find({
question: "Which excerpt explains the failed login?",
sources: excerpts
});
if ((match.value.answerExists ?? 0) < 0.7)
return { next: "inspect more evidence", reviewed: [] };
const reviewed = [];
for (const candidate of match.value.ranking.slice(0, 2)) {
const source = excerpts.find(item => item.id === candidate.id);
if (!source) continue;
const verdict = await jev.check({
claim: "The OAuth callback timed out.",
evidence: [source]
});
reviewed.push({ id: source.id, support: verdict.value.supported });
}
return { next: "agent reviews", reviewed };
}WHAT THIS CHANGES
This example can make one to three Jev calls inside the code run. Each inner call still uses the connection's scope and allowance.
Working from a file? With separate document consent, the agent can upload it by HTTP and call jev.searchDocuments inside the plan. Only that connection can search it, and it expires after 12 hours. The run returns an ID for inspecting a metadata-only trace.
Illustrative code, not a live evaluation or measured speedup. Jev returns probabilities, and the agent must treat an uncertain result as uncertain. Hosted Code Mode has separate daily execution limits.
THE AGENT WRITES THE PLAN
Which passage matters? Does it support the claim? Should the agent inspect another candidate? Code Mode lets the agent express that sequence with ordinary conditions and loops. Jev supplies bounded judgments; the agent keeps the final decision and action.
Read the Code Mode guideSTART WITH A REAL TASK
Use Jev to compare candidates, find relevant evidence or check a claim. Code Mode is useful when the next question depends on the first answer.
Find the code responsible for a form submitting twice.
Jev selects from the available tool descriptions. The agent supplies the arguments and runs the tool.
Give a repeated routing choice to a small model and keep the main agent on the repair.
Search found several files. Which ones explain this bug?
Jev rates the retrieved passages against the task. The original passages stay available for review.
Reduce irrelevant context before the next reasoning step.
Does this tool output support the claim that the release is live?
Jev checks the statement against supplied evidence and separates contradiction from missing information.
Catch an unsupported conclusion before it becomes an answer or another action.
Choose a calm palette for a coastal reading app.
Jev compares your described palettes with the brief. The agent applies the chosen tokens and checks contrast in code.
Reuse a selection recipe for design choices with a clear brief and a defined set of options.
Find a compact desk lamp with warm light and no subscription.
Jev compares the supplied item descriptions and returns a candidate or a reason to review the choice.
Match natural language to an existing catalog without generating an item that does not exist.
The service is running, but requests still fail. What should we inspect next?
Jev chooses among the checks the agent can actually run, using the symptoms already collected.
Make the next investigation step explicit. Compare it with a fixed diagnostic sequence.
WORKING IMPLEMENTATIONS
The browser companion and The Infinite Parable show how an application supplies context, uses Jev's answer and keeps control of the result.
A dedicated browser supplies accessible controls and current values. Jev chooses a small next step; the companion returns a screenshot and trace for the agent to review.
Recorded form runs and six accessibility decision cases are published with their failures and limits.
Explore the computer-use implementation ↗The game supplies eligible recordings and scene context. System One can recommend a cue or silence. The game retains cue history, spacing and playback authority.
An authored-dialogue integration. An independent playthrough study has not established an experience improvement.
Explore the dialogue implementation ↗CONNECT THE AGENT YOU USE
The skill explains when to delegate. MCP tools run the recipes. Your agent keeps control of the task and reviews the result before acting.
Connect the hosted MCP with OAuth, or keep using the verified local bridge.
See connection details Configuration availableRemote HTTP and OAuth setup instructions, plus the local bridge option.
See connection details Configuration availableUse the install link for remote MCP, then authenticate in your browser.
See connection details Configuration availableOpen the MCP install dialog, then authenticate in your browser.
See connection details Connection setup requiredAdd a custom MCP in developer mode, subject to account and workspace policy.
See connection detailsUse another agent? Start with its MCP support or call the HTTP API. Users can connect to our hosted MCP endpoint; local engines need a reachable transport for web clients. Check requirements and verification status.
MEASURE THE WHOLE WORKFLOW
Jev's published input list rate makes extra bounded checks practical to try. Code Mode can keep intermediate answers out of the agent's context. We have not measured a full-agent speed, cost or quality gain from Code Mode.
10,000 calls × 1,000 input tokens at Jev's $0.042/M published rate versus an editable $3/M example rate. This is not a bill or saved cost.
A two-step conditional path used one outer Code Mode call or two direct calls. Local wall time was slightly slower with Code Mode in the latest fixture.
One batched request per case made one more label error than two separate requests in a 12-case synthetic test.
The Code Mode fixture uses deterministic dialogue and no Jev inference. The batching result is a separate test, not a Code Mode benchmark. Neither demonstrates savings for a complete agent task. How to compare your own workflow ↗ Batched test method ↗
CONTROL STAYS WITH YOU
Code Mode uses only the tools granted to the connection. Every Jev call still counts against its allowance, and you can pause or revoke access.
Read the limits ↗Branch, loop and filter in one Code Mode run. Use the direct MCP tools when the task needs one judgment.
The agent supplies excerpts and chooses what to do next. Jev returns typed answers, not an action or a final explanation.
Compare final quality, Jev calls, agent tokens and end-to-end time. A low per-call price does not guarantee a cheaper or better answer.
Use hosted Code Mode for a dependent set of checks, or run direct tools in the local app.
The local app currently offers direct tools. Code Mode is available on the hosted MCP.