mythoscode

On the JembatanAI gateway

Agents that prove their work.

One engine, four products, one AI gateway. Every claim on this page traces to a recorded test.

terminal output
$ bun scripts/mythoscode/static-scan.ts --providersproviders: jembatanai onlystatic-scan: 0 hit(s) in 0 file(s)$ bun test packages/ai/test/jembatanai-errors.test.tsbun test v1.4.4 (38afee6c8)144 pass0 fail1047 expect() callsRan 144 tests across 1 file. [131.00ms]$ bun test packages/superagent-panel/test -t securitybun test v1.4.4 (38afee6c8)15 pass13 filtered out0 fail30 expect() callsRan 15 tests across 1 file. [552.00ms]
Recorded output, not a mockup. Commands and lines are copied from the verification records of proofs M0-GW-06, M0-ACCT-03, M3-PANEL-02 in feature_list.json.

Four products, one engine

They share one provider, one sandbox, one audit log and one validation pass. None is generally available. Each page carries its real status.

MythosCode

Release candidate in private build

The engine itself: a terminal coding agent with one gateway and one key.

  • Terminal-first coding agent with the IDE wired in
  • Single gateway, key from the environment only
  • Offline test suite on a mock gateway
  • Clean provenance: an MIT fork plus our own work
MythosCode →

MythosReview

Launching first

Pull request review where unconfirmed findings are dropped before anyone sees them.

MythosReview →

MythosGuard

Pilot by request

A security agent on your infrastructure, with a human gate on every consequential action.

MythosGuard →

MythosHunt

Early access

Bug-bounty testing inside your configured scope, with nothing reported that cannot be reproduced.

MythosHunt →

How a claim earns its place

  1. Produce

    An agent does the work inside written policy: scoped credentials, sandboxed execution, secrets kept out of model context.

  2. Validate

    A second pass confirms the result is real and reproducible, running the build or the tests where it can. Unconfirmed output is dropped.

  3. Approve

    A person approves before any external submission and before any consequential production action. The audit log replays the whole path.

Why it is built this way

One gateway
Every model call goes to the JembatanAI gateway. There is no second provider, no vendor login and no per-provider proxy, and a static scan counts what is left to strip.
Findings confirmed twice
MythosReview posts nothing until a second model from a different family confirms it. The same family is a configuration error, not a warning.
Humans on consequential steps
A missing policy entry is forbidden, a consequential action stops at an approval, and an out-of-policy attempt is denied and logged.
Claims with a record
Numbers on this site come from the repository's test records or from the benchmark file. A number nobody measured is not printed.

Who it is for

The Super-agent panel, running

Real screenshots of the web panel from this repository, served on a local machine with fixture data. Nothing is mocked up.

Super-agent panel, Needs you view, three items waiting for a person
Needs you: work that stopped at a human gate.
Super-agent panel, Learning view, a proposed memory change shown as a diff
Learning: a proposed change shown as a diff before anyone approves it.
Super-agent panel, Watchdog view, all checks passing
Watchdog: the agent is watched too.

Counted, not narrated

These counts are computed from the repository's feature list when this site is built. A status there changes only when its proof command runs.

130
tracked requirements
127
proofs passing
3
skipped, waiting on a live key

The benchmark round so far is offline: a scripted model on a mock gateway. It checks the pipeline, not capability, so this site quotes no capability score. Status

Pricing is not set yet

Every product will open with a free or trial tier, a flat team rate and a self-host option. The numbers are the owner's decision and are not printed until they exist.

Set at launch

See the pricing shape

Be told when it is ready

Leave an email. You get a note when a product opens, nothing else.

Questions

Can I use MythosCode today?

No. There is no public release. The installers on this domain already exist, and until release day they print the status and exit with an error instead of pretending to install.

Which AI models does it use?

Only what the JembatanAI gateway serves. The key comes from one environment variable and never enters a file, a log or a fixture.

Does my code go anywhere else?

Model calls go to the gateway and nowhere else. MythosGuard is designed to run on your own infrastructure, with inference through an endpoint you control.

What will it cost?

Not decided. The pricing page shows the tiers and says Set at launch where a price will be.

Can an agent post a wrong finding?

It can propose one. A second model has to confirm it before it is posted, and a person approves anything consequential. The false-positive rate is measured, not claimed.

Is MythosHunt for any website?

No. It works only inside a scope you configure and reports nothing it cannot reproduce. Out-of-scope requests are denied and logged. It never submits a report itself.