One terminal for coding agents, conversation branches, and reviews of finished runs.
Use Pi, Codex, or OpenCode through the same interface, on your machine or in a Tangle cloud sandbox. Keep the instructions you give an agent, compare different approaches, and see which runner, model, and workspace each turn used.
Install and start · Cloud setup · Commands
A real run: the agent fixed a slugify function and passed eight tests.
Then /ask reviewed the result, citing the code and test output and identifying a coverage gap.
Read the captured example.
You need Node.js 22.19+, Linux or macOS, and a coding agent that is installed and signed in. Start CLI Bridge in a separate terminal; it connects Braid to your local coding tools. The setup guide has the complete commands.
With CLI Bridge running at http://127.0.0.1:3344:
npm install --global @tangle-network/braid
cd your-project
braid- Choose a profile from the runner and model combinations advertised by your bridge.
- Choose Local CLI Bridge, review the selection, and apply it.
- Type a task in the conversation, for example:
Find where this project handles authentication. Explain the flow and cite the files. Do not edit anything.
The transcript shows the response and tool activity.
The status line identifies the selected runner and model.
Use /help for commands, or /quit to leave.
A profile is the saved set of instructions and settings for an agent. Braid can generate one from your bridge's model catalog, so this local path needs no profile JSON. If setup says No AgentProfiles, check discovery or supply a profile.
Prefer a cloud workspace? Follow Tangle Sandbox setup. It requires a Tangle account, an API key, and available credit. New cloud connections delete the sandbox after one turn; choose retained execution before relying on files or background work.
| What you want to do | Command |
|---|---|
| See running work, waiting questions, and usage | /activity or F2 |
| Try another approach from this conversation | /fork |
| Inspect the conversation's branches | /graph |
| Add the next task while a run continues | /queue <task> |
| Stop the selected run | /cancel |
| Ask what a completed run established | /ask <question> |
| Change the next turn's runner or model | /runner, then /model |
/fork previews what carries over before you confirm.
A conversation fork shares the workspace by default; /fork --workspace requests a separate workspace when the provider supports it.
Changing runners uses a new provider session with an explicit context transfer.
It does not copy a runner's private memory.
/ask runs a separate analysis of a completed or failed run.
It can make additional model calls and charges; its findings cite the saved run without changing that conversation.
Read the analysis setup and commands before using it.
Commands that need an unavailable provider feature explain what is missing. See the conversation and control reference for continuation, permissions, queueing, and worker controls.
- Local work: CLI Bridge uses the coding tools and accounts you configure. Their subscription limits and API charges still apply.
- Cloud work: Tangle provides model access and sandbox compute. Review the selected connection, workspace source, and lifecycle before sending a task.
- Saved work: Braid keeps conversations and run records in an encrypted local database. A saved conversation does not imply that a remote process or its files still exist.
- Credentials: enter keys through the masked setup prompt. Keep them out of profiles and prompts; inspect exported run content before sharing it.
- Usage: reported, estimated, and unavailable costs remain distinct. An unavailable cost is not zero.
Security and storage · Profiles and connections · Recorded capability comparisons
braid rpc exposes JSON Lines commands over the same application core.
For scripts and integrations, start with the headless interface guide.
For terminal scrollback, use braid --inline; for plain text, use braid --plain.
Read the product contract, component map, and verification guide. Braid uses agent-runtime for execution and pi-tui for terminal rendering. See third-party notices for attribution.