Build
Projects, prompts, models, files, previews and the receipts behind a published app.
Last verified against production: . See current limits and availability.
Projects and conversation
Choose a project from the menu or create a new app. Rename the project with its rename control. Each prompt becomes a build in that project's conversation. Switching projects updates the conversation, steps, files and preview; a server-owned run continues while you view another project.
Preview shows the hosted app. Code shows grouped files, copy controls, changed-file markers and report or ABI artifacts when present. Chat gives the conversation a centered reading column. The selected mode is saved per project. Drag the desktop divider to widen the conversation up to half the workspace.
Each reply has one Build details disclosure with that run's steps and available tool output. It opens during a running build and collapses at completion. The working or reconnecting indicator describes the stream state; elapsed time alone does not prove that the model has produced another step.
Prompt and attachments
The composer accepts 24,000 characters. Text files can be MD, TXT, CSV, JSON or YAML: five files, 64 KiB each and 128 KiB combined. A large paste becomes a recoverable brief file. Open its chip to inspect the complete text. Attached text is included before needs analysis and is also placed in the sandbox workspace.
Paste, pick or drop PNG and JPEG images when the selected model supports image input. Limits are three images, 2 MiB each, 4 MiB combined, 4096 pixels per side and 16 megapixels. Unsupported input is refused with an error; the UI must confirm an attachment before you rely on it.
Enhance revises the prompt before a build. Its pending state blocks Build. Enter sends, Shift+Enter adds a line, Cmd/Ctrl+Enter also sends and Cmd/Ctrl+K focuses the composer. A failed submission preserves the draft for correction.
Models and your own key
Settings accepts OpenAI, Anthropic and Kimi API connections. Save a provider key, then choose that connection from the composer's model menu. An own-key build costs 0 Fletch credits; your provider charges for API usage. A chat subscription is separate from a provider API account.
Hosted keys are encrypted at rest and decrypted by Fletch's server for provider requests. Generated app code receives a scoped proxy credential rather than the upstream provider key. Operators with access to the server and its encryption material can access hosted keys. Revoke or replace a connection in Settings when needed.
A build records its selected provider, model and credential version. Changing the menu affects a future submission. A rejected or unavailable provider returns an error that needs to be resolved with that provider or another connection.
Queue, progress and retry
Two execution slots are coordinated through Postgres. Additional accepted runs wait in the queue and show their position. Queued Cancel and Edit handle the race with a worker claiming the run: if it has started, the interface refreshes its actual state.
Closing a tab does not cancel the server's run. Reopening a project reloads saved steps and reconnects to the stream. Source and session files are retained through supported maintenance. Current runtime and recovery limits are listed on the limits page.
A failed follow-up preserves the project's recorded published app. Run it again starts another attempt using the available project context. A refund is an append-only credit entry tied to the charged run; an own-key run has no Fletch debit to refund.
The sandbox and verdict
Model-authored commands run in the build container with a read-only root, restricted capabilities and a mounted workspace. Compilation, tests and security analysis produce stored reports. Read those reports to see which checks ran and their exact output.
The pipeline links to recorded contract and host receipts. The verdict is produced by the toolchain and remains above the workspace. A model's account of a change is useful context, while the compiler and transaction receipts establish their own results.
Audit for mainnet places a review instruction in the composer. Submit it to run a follow-up build through the same model and toolchain. Review its reports and remaining findings before any deployment decision.
Preview and wallet behavior
The preview toolbar has the live URL, Refresh, Open app and a settings gear. Fit is the default. The gear contains device widths and zoom; a fixed device viewport scales to fit the available pane.
Signing apps use the shared wallet component. It supports compatible injected EVM wallets, including Phantom's EVM provider, and WalletConnect pairing when a Reown project ID is configured. Wallet availability also depends on the browser, extension and chain support. Inspect the generated app's connection and transaction flows after publication.
Contract deployments currently target Robinhood Chain testnet, 46630. Front ends are hosted under apps.fletch.now. Use Code to inspect or copy the source, reports and ABI that the build actually produced.