Experimental

TokenDanceCode

TokenDanceLab/TokenDanceCode as an experimental Codex / Claude Code style CLI project.

Positioning

TokenDanceCode is a CLI/agent experiment from TokenDanceLab. It explores the interaction model, execution boundary, and engineering feel of command-line agents such as Codex and Claude Code.

It is not a replacement for AgentHub. AgentHub focuses on multi-agent collaboration, Desktop, and team workflows; TokenDanceCode is a lightweight CLI proving ground.

What it is good for

Trying command-line agent task input, progress feedback, and result presentation.
Validating model API calls through TokenDance Gateway.
Capturing CLI interaction patterns that may inform future AgentHub runtime adapters.
Keeping an open, low-commitment product idea without claiming unfinished capabilities as stable.

Current product shape

Topic	Notes
Entry	Global command-line Coding Agent; target command is `tokendance`
Interface	Scrolling CLI experience, not full-screen TUI, GUI, Web, or IDE plugin
Models	Current CLI can auto-enable an Anthropic-compatible provider; OpenAI provider mapping and tests exist, but CLI auto-selection is not a default promise yet
Architecture	Core Runtime is separated from CLI Shell; Runtime emits structured events and CLI renders them
Tools	File, shell, git/diff/review capabilities are implemented around permission and review boundaries

Local trial path

The source repository remains authoritative. The example below is safe public guidance for an experimental trial path; OpenAI provider mapping exists, but Gateway/OpenAI-compatible CLI auto-selection is not a default-entry promise yet, and exact commands should follow the repository README, docs/, and current implementation.

powershell

git clone https://github.com/TokenDanceLab/TokenDanceCode.git
cd TokenDanceCode
python -m pip install -e .
$env:ANTHROPIC_API_KEY = "<your-anthropic-compatible-key>"
tokendance

Configuration and credentials

Configuration	Recommendation	Boundary
Anthropic-compatible key	Use a local environment variable for the current default experimental path	Do not commit it to repositories, issues, screenshots, or sample config
OpenAI/Gateway provider	Treat it as a provider-mapping direction and follow the current repository README and implementation	Gateway auto-selection is not a stable public promise yet
Working directory	Run inside a local repository and start with a small read/diff task	Do not publish private absolute paths in public reports
Command execution	Begin with read-only commands before write or shell tasks	Writes, shell, and git actions need explicit review and confirmation boundaries

What to observe during a trial

Whether task input reaches the runtime clearly, and whether progress and final output are understandable.
Whether file reads, diffs, command output, and error states have structured boundaries.
Whether failures distinguish model credentials, network, commands, permissions, and code issues.
Which CLI interaction patterns should feed back into AgentHub Desktop or Runtime adapter design.

Relationship to AgentHub

Dimension	TokenDanceCode	AgentHub
Use case	Personal command-line experiment and local agent UX	Multi-agent collaboration, Desktop, Web, Hub/Edge, and team workflows
Status promise	Experimental project with low-commitment public wording	Product line with continuously aligned site and docs
Event value	Captures CLI runtime events and interaction patterns	Normalizes runtime events into collaborative, reviewable, auditable product events
Gateway use	OpenAI provider mapping exists; CLI auto-selection for Gateway is not promised yet	Configures model API boundaries through adapters or server-side settings

Experimental boundary

The GitHub repository is the current public entry point; installation, commands, and configuration should follow that README.
OpenAI provider mapping exists, but Gateway/OpenAI-compatible CLI auto-selection is not a default promise yet; do not describe the TokenDance API key trial path as stable.
Do not treat it as a replacement for AgentHub Desktop. It is better framed as a command-line agent interaction experiment.
Real model keys should use local environment variables or TokenDance Gateway, not repository files, issues, or public screenshots.
Keep capability language low-commitment: exploratory, usable for feedback, and not an enterprise SLA.