@tangle-network/braid

Braid

Select the published Discovery Operator without cloning a repository:

npm install -g @tangle-network/braid
braid --profile discovery-operator --connection connection-local-cli-bridge

The existing connection picker configures CLI Bridge or Tangle Sandbox. --connection selects its configured id. The operator needs disco read access, gh-drew, and a subscription seat on the execution host. The published name selects the immutable Lab profile snapshot bundled in the package and validates it with the canonical profile importer. Selection works offline; local profile paths remain supported. Play approval names the exact play id and commit. A status question authorizes observation only. The operator's approval rule is an instruction, not a host command permission boundary.

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. Record which results you accept, turn corrections into portable agent profiles, 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.

Start with a local coding agent

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.

While you 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>
Record whether a finished task met your needs	/feedback accept [reason...] or /feedback reject [reason...]
Review a profile draft from your corrections	/profile learn [new-file]
Request a runner recommendation for the next task	/runner advice <task>
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.

Keep corrections for the next task

After a finished run, explain what to repeat or fix:

/feedback reject Repair the implementation instead of weakening the regression test.
/profile learn learned-agent.json

The profile preview shows lessons from this conversation, their feedback sources, and the exact changes to a portable AgentProfile. The draft is unmeasured. Private values and metadata follow profile export redaction rules, with a notice when redaction changes the profile. Press Ctrl+S to save a new file, then explicitly select it with /profile learned-agent.json when ready.

/runner advice <task> uses this conversation's feedback, the latest finished run, and compatible runner/model pairs in your profile catalog. It uses the existing analysis connection and can incur model charges. The result is advice based on recorded observations; it does not automatically switch runners or establish which runner is best.

The next step is to compare candidates on held-out tasks through Runtime and Eval before using results to learn runner selection. Profile drafts can already be saved and reused locally; profile sharing and cloud publication are not implemented in this workflow. See profile learning and runner advice.

Your work, accounts, and costs

• 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

Use Braid from another program

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.

Contribute

Read the product contract, component map, and verification guide. The vision records ranked directions, their evidence, and longer-term research bets. Braid uses agent-runtime for execution and pi-tui for terminal rendering. See third-party notices for attribution.

MIT license.