Prized
PlatformPricingSign inGet started

The platform behind Prized.

Agents can write the app. Prized supplies the environment around it: isolated builds, scoped data, brokered credentials, authenticated distribution, and a durable audit trail.

01
Build

An isolated Linux sandbox runs the coding agent and Vite preview.

02
Scope

A tool-specific data path and signed tokens define what the build can reach.

03
Publish

The static bundle is uploaded and the tool's live version pointer flips.

04
Serve

The workspace gateway validates identity, serves assets, and proxies data calls.

One isolated sandbox per build.

Every Prized build starts in a private Linux sandbox with a coding agent, the real React source tree, and a Vite development server. The agent edits the application while the builder watches the live preview and iterates against the tool's scoped data path.

Engineers can also work with Claude Code in a Prized browser sandbox, or connect an external project with a tool-bound developer token. Those paths do not receive broad workspace access: they target one tool's data APIs and converge on the same publish pipeline.

Private Daytona sandbox
OpenCode + Vite in the builder
Claude Code available in browser sandboxes
Tool-bound developer tokens for custom code
build sandbox
Coding agent
React source
Vite preview
Ships through the publish pipeline
The agent, source, and live preview share one disposable boundary. Whatever wrote the code, it ships through the same publish pipeline.

The builder sandbox has no database or model credentials.

Generated code never receives a Postgres connection string or a model-provider key. Tool data calls go to Prized's authenticated SQL gateway. Model traffic goes to the LLM gateway. The sandbox receives derived credentials bound to its session or tool, not the platform secrets used behind those gateways.

That makes a leaked sandbox credential narrow by construction: it cannot be used to open an admin database connection or pivot into another tool.

No model-provider keys
No Postgres connection string
Session-scoped egress token
Tool-bound data token
in the sandbox
Session-scoped egress token
Tool-bound data token
never in the sandbox
Postgres connection string
Model-provider API keys
Derived, revocable tokens go into the sandbox. The platform secrets stay behind the SQL and LLM gateways.

Postgres grants are the isolation boundary.

Each tool is provisioned with its own schema, login role, and encrypted connection URL. The production data path opens Postgres as that role. There is no shared admin connection followed by a role switch for generated code to escape through.

The role's grants constrain access at the database engine. A separate SQL guard rejects cross-schema references, multiple statements, and statements that do not belong on the requested read or write path. Read calls also run in a read-only Postgres transaction.

Schema per tool
Login role per tool
Encrypted per-tool connection URL
Fail-closed SQL checks
Tool role
Its own schema
granted
Tool role
Every other schema
no grant
The data path connects as the tool's own role. Grants stop cross-schema access at the database engine, and the SQL guard fails closed in front of it.

Credentials stay in the broker.

Tool integration calls go through Prized's egress proxy and integration broker. The proxy authenticates the caller, binds the request to a workspace, checks the destination against the platform and workspace host allow-lists, and resolves the workspace's encrypted credential on the server side.

Generated code names an approved integration and operation; it never receives the raw credential. Reads are the default path. Writes require an enabled integration write path and a caller with write access.

Workspace-bound signed token
Host allow-list + SSRF checks
Encrypted credentials resolved server-side
Explicit write access
Tool call
Egress proxy
credential attached
Approved host
The workspace credential is attached inside the proxy, after the allow-list and SSRF checks. Generated code only ever names the integration.

Publishing is an atomic version flip.

Publishing validates the built output and uploads its files to S3-compatible object storage under a new version. A database transaction then inserts the version record and flips the tool's current version pointer.

The gateway serves the selected bundle at a workspace-scoped URL and proxies its API calls back through the governed data plane. Private workspace tools require a short-lived, workspace-scoped JWT; the gateway verifies the signature, workspace binding, and tool access before serving the request.

Validated static bundle
Versioned object-storage prefix
Atomic current-version pointer
Workspace-scoped JWT verification
v41previous
v42previous
v43current
Publishing validates the bundle, uploads it under a new version, then flips the current pointer in one database transaction.

The tool and its boundary move together.

Tool creation, sharing, and database access (down to the exact SQL run) are audit-logged. Workspace administrators can review that activity in the workspace audit view.

Forking copies source and lineage into a new tool. It also provisions a new schema, role, and credential. Access-control rows and integration grants are not copied from the parent, so a teammate reconnects the fork to data they are allowed to use.

Workspace audit view
New data boundary on fork
No inherited parent ACL
refund-review
Its own schema
Its own role
Its allowed users
fork
refund-review (fork)
New schema
New role
No inherited access
A fork copies the source and lineage. Access starts over: new schema, new role, nothing inherited from the parent.

Govern what a tool can access and do, not how every line was written.

Build with Prized

Build internal tools with AI. Safely.

PlatformGuardrailsAuditSign in
PrivacyTerms
Prized