huddle

One session. Many AI brains. Supervised. @codex, @gemini, @opencode — answer inline in the agent session you're already in. And when one of them works, another one checks the work.

You're in opencode. You want a second brain on a diff, a second opinion on an architecture call, a worker to fix the file while a supervisor verifies it did what you asked. Today you switch terminals and copy-paste context. Huddle puts every agent at one table: you @mention, that agent answers in your conversation, every agent shares the same transcript — and any agent can audit any other's work.

you:     @codex what does add(2,3) in calc.ts return? Just the value.
codex:   5
you:     @gemini check what codex did — was it right?
gemini:  VERDICT: ON-SPEC — add(2,3) returns 5, matching the request exactly

That second exchange is the point of huddle: AI watches AI. A supervisor agent audits a worker's run against your task and returns a verdict with findings — not vibes.

Install

huddle init --host opencode --guests codex,gemini

Detects your installed agents, writes the MCP config into your host, drops the routing skill into your project. Then just start your host and @mention a guest.

Requires: Node ≥ 22 (native TypeScript + SQLite, zero npm dependencies) and one MCP-capable host (opencode). Guests just need their CLIs installed and logged in — huddle never touches your auth or billing.

Related MCP server: Hive

Any agent. Any model. Your combo.

• Guests: any agent with a terminal CLI — ships with codex, gemini and opencode adapters; anything else is a 15-line JSON manifest (docs/ADAPTER.md). PR yours in.

• Models: per-call override — @codex[o4-mini], @gemini[gemini-3-flash]. Free, paid, or local: it's your CLI, your keys, your choice. models lists every catalog with free/paid/local tiers.

• Combos: every user picks their own host + guests at init. No hardcoded pairs.

The tools (MCP)

Advise by default. Act on command.

• @codex review this → advise mode: guest reads your project with full agentic context (spawned in your working directory), answers, suggests diffs. Read-only by adapter flag.

• @codex! fix it → act mode (!): the guest edits files directly, sandboxed to the workspace. Git guards record the diff after every guest turn, and one command recovers:

huddle undo   # reset the working tree to HEAD — guest edits gone

Security

Zero npm dependencies — Node built-ins only, nothing to audit but us. Args-array execution (never a shell), binary allowlist (only registered adapters run), secret redaction on every transcript write, per-call timeout kill, per-session spawn caps, delimited guest output. Full model in SECURITY.md. Windows-native: npm .cmd shims are resolved to their underlying scripts and spawned shell-free.

Verified working (real run, 2026-09-09)

From huddle history — actual bus transcript, all free models:

[07:41:23] codex(gpt-5.6-terra) answer: In ordinary arithmetic, yes—though programming
           languages can produce different results through overflow, floating point…
[07:45:20] gemini(gemini-3-flash) answer: I agree with codex because dynamic type
           coercion (such as '2' + 2 evaluating to '22')…
[07:46:11] gemini(gemini-3-flash) answer: VERDICT: ON-SPEC FINDINGS: 1. Position Provided…
[08:20:06] codex(gpt-5.6-terra) answer: 5          ← via opencode host, inline

FAQ

Does huddle itself need an agent? The host does (opencode — anything MCP-capable). Guests just need a CLI. A standalone huddle chat TUI (you as the host, zero agents required) is on the v1.5 roadmap.

Does it touch my API keys or billing? Never. It invokes your installed CLIs; their own auth and models carry over.

How is this different from parallel multi-agent routers? Those dispatch tasks to agents working in parallel and synthesize one answer. Huddle is turn-based and conversational: agents share one transcript, talk in sequence, review and audit each other — a session, not a dispatch queue.

Roadmap

• v0.3 — model pinning in config, review across sessions

• v1.5 — huddle chat standalone TUI (human host), during-run drift guard

• v2 — live interjection (supervisor course-corrects a worker mid-run)

License

MIT