UCLI is an experimental universal command-line interface for working with AI agents. Its purpose is day-to-day comfort: one discoverable place to choose models, switch accounts, set effort levels, and return to saved work without memorizing each vendor's command syntax and session flags.
This repository is a clean public source checkout for people deciding whether to try the local interface or contribute to its development. No public binary release is promised, and no private development history or host verification evidence accompanies this checkout.
UCLI requires Go 1.27.2 and uses the Python 3.13 standard library for build and verification scripts. You can explore the interactive interface immediately in demo mode, which creates isolated local state and simulated agent responses without using any AI quota:
go run ./cmd/ucli --home .build/demo-home --demoType a prompt and press Enter to receive a simulated local reply. Press Tab to discover actions, run /models to browse model choices, or run /sessions to inspect saved activity. The demo uses no provider quota.
To run the repository checks and package binaries from a single command, invoke the developer runner with your Go executable path:
python tools/dev.py all --go PATH_TO_GO --jobs 1Packaged outputs are written to .build/bin/ucli-windows-amd64.exe (Windows), .build/bin/ucli-linux-amd64 (Linux), and .build/bin/ucli-darwin-arm64 (macOS). These cross-builds verify that the source compiles cleanly for each target, but they are not native host proof.
The fixed rich header shows the static cat mascot, version, and task context, while the framed bottom editor places notifications above the input box and the active model and working path below it. PageUp and PageDown scroll the output body. Enter submits a nonempty prompt, Alt+Enter, Shift+Enter, or Ctrl+J adds a newline, and an empty Enter never starts work. Restored unfinished text is editable, and paste or voice input inserts reviewable literal text without auto-submitting. Tab discovers actions and completes slash commands, and /send provides an explicit submit equivalent. Display flags preserve layout and input: --no-color and NO_COLOR strip color only, --no-motion keeps cats static, --ascii switches only cats to ASCII, and --plain uses the static line interface. Additional interface details are in internal/cli/README.md.
Startup and native admission checks still run synchronously. After a prompt is admitted and saved, the rich shell returns to editing and menus while an in-process worker owned by this UCLI invocation runs it. Pressing Enter on a task that is already working saves the next prompt without queuing, steering, or sending it. /sessions inspects saved activity, and /stop requests a stop only for active work on the selected task owned by this process. Pressing Ctrl+C twice within the two-second confirmation window exits and requests a stop for this invocation's workers, as does closing UCLI. There is no detached service or promise that work survives terminal or process closure. See the terminal UI research for design context rather than usability evidence.
Local state is persisted in SQLite, covering saved tasks, drafts, and results. Each task tracks its requested model, account, and effort settings independently, supported by replay protection, search and paging, explicit backups, and backed-up schema migrations. Core design boundaries are described in docs/architecture.md.
You can validate and inspect workflow definitions locally using workflow check and workflow view examples/review.workflow.json. These commands only validate and display the graph; they never execute a workflow. Optional local voice input starts only on explicit request and inserts reviewable text into the composer rather than sending automatically. Only synthetic voice fixtures are tested today; real microphone hardware, ASR quality, and backend downloads remain unverified.
/models reports installed binaries, cached or unchecked sign-in state, and experimental or unavailable execution separately. Blocked entries can be inspected without changing the current selection. Ordinary listing is local, while explicit setup or refresh reads native metadata. Claude model rows come from discovered native metadata rather than a hardcoded alias list, and models require a compatible reviewed build and supported capabilities. AGY can refresh and cache its native catalog only; UCLI sign-in verification and execution remain unavailable. GitHub reads local gh identity and provides no AI models or hosted UCLI account. Credentials stay in native stores. See ucli login codex, ucli login claude, ucli login github, and docs/native-signin.md for details.
Prior bounded native checks on Windows x64 completed small read-only requests with Codex 0.163.0 and Claude 2.1.304. That provides evidence for those reviewed builds and requests, not universal model, account, or OS admission, and not proof for this new worker integration. Native execution remains experimental. The read-only default disables native tools, while explicit --permission sandbox grants full local computer access through the native client without added OS isolation. Unknown outcomes remain guarded: native reconciliation is unavailable, and UCLI does not blindly retry. Follow-up prompts send only the new input into fresh native sessions without automatic provider conversation memory. auto currently selects the default supported model at its lowest supported effort. Account cycling, provider quota reservations, workflow execution, and a background model/effort decision agent remain future work.
Contributions are welcome, especially on small issues, offline protocol fixtures, task recovery, cross-platform portability, and terminal usability. See CONTRIBUTING.md, docs/verification.md, and docs/next-development.md before opening a pull request.
Released under the MIT License (LICENSE), copyright 2026 BlueRaddish.