What is tokencode?

TokenCode is a terminal-based AI coding agent written in Go, similar in concept to Claude Code. You describe a programming task in plain text, and the agent reads files, edits code, runs shell commands, and searches the web in a loop until the job is done. It works with any AI model that speaks the Anthropic API format, defaulting to a DeepSeek endpoint so you can use it without an Anthropic account. The standout feature is a command called /race. Instead of running one agent on a problem, /race spins up multiple agents, up to 1,000, each working independently in its own isolated copy of your repository. When they finish, a judging step compares the results, scores them, and picks the best solution. Losers are discarded and their work is cancelled. The idea is that AI token costs keep falling, so throwing many parallel attempts at a hard problem and keeping only the best is more reliable than trusting a single run. You apply the winning result with one command, nothing commits automatically. Permission modes let you control how much the agent can do without asking. Plan mode is read-only. Review mode asks you to approve each action. Auto mode lets a smaller model decide what is safe. Yolo mode runs without interruption. You can switch modes mid-session. The agent also saves file checkpoints before every edit so you can roll back changes with a /rewind command without touching git. For teams, the project supports connecting agents to messaging platforms including Feishu, DingTalk, and WeChat. Each team member pairs their account with a working directory on their own machine. After that, they send a message to a bot and the agent runs a task in their local workspace, replying with the result. No shared server or public IP is needed. The server mode adds a web interface showing token usage stats, a chat panel for single-turn queries, and a team management page. It also exposes an A2A protocol endpoint so other AI agents can discover and call TokenCode as a sub-agent.