touchmarkon Stellar

The StudioDesk 2 · the back room

Bring the whole agent. Keep the dossier.

The quick test on the home page reads one public repository at the counter. The Studio is the bench behind it: private files that stay in memory, a GitHub repository pinned to one commit, a .tmark held to the code, reviewers you choose, and an evidence dossier you download and keep.

Open the workbench Quick test instead

No wallet. No payment. Nothing runs unless you consent to it.

What the Studio adds to the quick test

i.

Private code, kept private

Upload files or a project folder. An immediate review stays in this server's memory and in your tab.

ii.

GitHub, pinned to a commit

Import a public repository at one immutable commit. Entry points are discovered; nothing is executed.

iii.

A .tmark, held to the code

Declare what the agent may do. Every claim is checked against the source that signs and pays.

iv.

A dossier you keep

Choose reviewers and depth, then download the designed report, the evidence JSON and a capsule.

01

The workbench

File the agent on the ticket: source, a .tmark, its intended behaviour and the reviewers you want. The dossier fills in as Punch reads.

Nº 02 · Agent to file
ASource

Pinned to one commit. Source, dependencies and possible entry points are discovered; no special filename is needed and nothing is executed.

On the bench

empty

A root veritas.json locates the entry point, instructions, tools and policies precisely. Discovery can draft one for supported projects.

Just exploring?
BDeclaration .tmarkoptionalnone

A .tmark says what the agent may do. The review holds the code to each claim, so a broken promise becomes a precise finding. What goes in one

.tmarkPaste it, upload it, or drop it with the source.
    Open the builder
    CIntended behaviourbe specific
    Declared capabilities what it can actually do
    Advanced inputs 0 tools declared

    Imported manifests keep their tool names and configuration. A declared permission is not proof that a limit is enforced.

    Architecture & trust boundaries optional

    Describe mcp_servers, delegates, memory_stores, allowed_domains and controls. Declarations are checked separately from source evidence.

    DDesign your auditTouchmark official

    One accountable operator. Specialist reviewers for the parts that matter to you.

    Choose the depth

    Loading available review plans…

    Choose your reviewers

    All official reviewers are operated by Touchmark. Baseline source checks always run; your selection shapes the investigation. Independent operators will join later.

    Your Touchmark capsule .tmcapsule

    A portable audit contract: source, intended behavior, authority boundaries and scenario expectations travel together. Fingerprints tie the plan to the exact evidence.

    No capsule loaded. You can create one from your project.

    The capsule contains your source. Keep it private unless you intend to share it.

    Execution settings explicit consent

    Use the detected entry point or the one declared in veritas.json. Imports and execution occur only inside an isolated microVM.

    An immediate source review stays in memory. Official orders store the submitted source with your explicit consent. Optional AI review sends redacted source to the named provider only with your consent. Never upload credentials.

    No wallet connection. No payment. No transaction.

    Order an official Testnet reviewPilot

    Create your private order, then fund its scoped review with test XLM. Your wallet signs each payment step. Available automation runs the review and completes report delivery after funding is verified.

    Get the wallet

    Preview your audit plan above, then request a quote.

    One order. The whole review.

    Checking official order availability…

    Get a quote first. Keep the private order receipt to resume later.

    Resume an existing order

    Use the pilot key above and your private receipt. Source, access tokens and pilot keys are never saved in browser storage.

    Load the same project and audit selection, get its quote, then prove requester ownership. Recovery rotates the private token; the previous receipt stops working.

    Evidence dossier

    Ready when you are

    Trust needs a paper trail.

    Bring an agent to the bench. The dossier shows concrete findings, what is missing, and what to check next.

    1. Find risky permissionsSigning, spending, code execution and data access.
    2. Understand the evidenceSource locations and actionable checks, not just a score.
    3. Keep your reportAn integrity-checked download. Stellar publication is a separate step.
    Address ≠ audit

    A wallet balance cannot tell us how an agent behaves. Source and runtime evidence answer different questions.

    02

    Look past the prompt

    An agent is a system of permissions, dependencies and decisions. Its review should be, too.

    i.

    What it intends

    Instructions, goal conflicts, prompt injection, tool manipulation and delegation.

    Intent & control
    ii.

    What it can touch

    Signing authority, spend limits, credentials, network access, memory and private data.

    Security & authority
    iii.

    What you can prove

    Source locations, build provenance, failures, report commitments and evidence gaps.

    Evidence & accountability
    Explore all 24 control areas
    03

    Evidence that travels. Authority that doesn’t.

    Private source stays off-chain. A report binds to an exact agent version, its issuer and its lifecycle. Understand the trust model

    01

    Authorization, at the contract

    Soroban authorization binds signatures to invocation arguments. Application policy still defines what is allowed.

    02

    Payment for work. Not a verdict.

    SAC escrow separates requester, provider and evaluator. Payment does not buy a passing result.

    03

    A credential with an expiry date

    Non-transferable report credentials bind to a version, expire and can be revoked. The Testnet pilot links scoped review evidence to registry-bound payment and a report credential. Check the published receipt below.

    The Stellar proof

    From a funded review to a verifiable record.

    Checking the pilot receipt

    Follow the official Testnet pilot: a scoped source report, registry-bound escrow and a non-transferable credential.

    This is a published protocol demonstration using our fixture. It does not issue a credential for an agent you upload.

    04

    Check the record

    Verify the exact bytes of a downloaded report against a fingerprint you obtained independently. Your file stays in this browser.

    Nº 04 · Report to checkSHA-256 · in this tab

    Integrity confirms a file match. It does not prove issuer identity or safety.

    05

    Give autonomy a boundary

    Try the same deterministic rules the enforcement SDK uses. This simulation never signs or sends a transaction.

    Policy · Testnet · demo

    Authority is explicit.

    10 XLM per action. 50 XLM total. One permitted destination. One delegation level.

    Rule evaluation only; no signed approval or execution.
    06

    Pay for the review. Not a passing score.

    Choose the depth of investigation. The verdict never changes the price.

    Source triage

    Preview

    Free

    Map your project, inspect the supplied evidence and find what needs a closer look.

    Build your audit ↗
    Bounded execution

    Assurance

    Testnet pilot

    Run a supported entry point, challenge selected behaviors and keep reproducible observations.

    Choose your reviewers ↗
    Extended investigation

    Deep

    Testnet pilot

    Expand the test plan and examine more boundaries. Every report distinguishes executed checks from open questions.

    Set the scope ↗

    Loading current plan budgets and pilot prices. Production checkout is not active.

    07

    Good questions. Straight answers.

    What the Studio establishes, and what it does not.

    Can I add any Stellar agent?

    You can submit a public G-account or C-contract address on Testnet or Mainnet, or provide source without a wallet. The address check is read-only. A private agent needs its owner's source or a later authorized runtime test; no service can fully audit hidden behavior from an address alone.

    Does Stellar have an agent key?

    There is no universal agent ID required by this product. Agents may use ordinary Stellar accounts or smart accounts with delegated permissions. We ask for public identifiers only. A private key is authority to act, not evidence of trustworthy behavior.

    What does an audit establish?

    Official reviewers inspect the supplied source and declared boundaries. Authorized runtime plans execute supported Python entry points in isolated microVMs. The dossier keeps source findings, runtime observations, AI hypotheses and chain evidence distinct. Each claim is bound to the evidence actually collected; an untested surface remains an open question.

    Does a credential mean the agent is safe?

    No. It records an issuer's assessment of a specific version. It cannot reveal hidden behavior or guarantee future behavior. Always check the report scope, issuer, version, expiry and revocation state.

    What is available today?

    The workspace checks the available reviewers, providers and official order service on this deployment. Immediate source review works without keys. An enabled official order service saves the consented scope before funding, runs the review and completes Testnet delivery. Private receipts control access to the report; its hash and token record are public. Runtime and AI review keep their separate consent and access requirements. Testnet uses test assets only.

    When should I use the quick test on the home page instead?

    When the agent's code is in a public repository and you want a rules verdict in about a minute, with an optional model review that reads every file and a testnet stamp for a pass. Come to the Studio for private code, uploads, a declaration you are still drafting, chosen reviewers, or a dossier you keep.