tribe

Tribe

Reusable coordination + memory infrastructure for coding-agent sessions.

Tribe lets multiple agent sessions discover each other, exchange messages, and share project context through a local daemon — with session-history recall built in. This repository owns the reusable infrastructure only: wire protocol, daemon, recall engine, and host integrations. Project workflow conventions such as @chief, @agent/N, beads, worktrees, and integration policy belong in a tent/SOP layer outside this repo.

Quick start (Claude Code)

/plugin marketplace add beorn/tribe
/plugin install tribe@tribe

Restart Claude Code; the tribe.* MCP tools (join, send, fetch, members, …) appear and the daemon autostarts on first use. Two sessions on the same machine can then talk: tribe.join({name: "a"}) in one, tribe.send({to: "a", ...}) in the other.

Components

Component	Package	Binary	Owns
Wire client	tribe-wire	tribe-wire	Unix-socket client, reconnecting transport, MCP stdio adapter, protocol CLI
Daemon	tribe-daemon	tribe-daemon	Broker process, SQLite state, session registry, message journal, plugins
Recall engine	tribe-recall	n/a	FTS5 session-history search, LLM planner/agent, hook handlers, lore summaries
Injection envelope	tribe-injection-envelope	n/a	Prompt-injection defense: envelope framing, sanitizer, hook JSON emission
Claude Code plugin	plugins/claude	n/a	Marketplace plugin: MCP registration + host-managed daemon lifecycle

tribe-wire is published to npm; the rest ship with this repository.

Mental Model

Claude / Codex / Hermes
  -> tribe-wire mcp            # from package tribe-wire
  -> local or SSH-forwarded Unix socket
  -> tribe-daemon
  -> SQLite + daemon plugins + recall engine

tribe-wire talks to an existing tribe. tribe-daemon is the broker process; it lazy-loads the in-repo recall engine for memory verbs (tribe.ask, tribe.brief, delta injection). Host plugins wire those pieces into an agent runtime.

LLM features (optional)

Coordination and FTS search work with zero configuration. Features that call an LLM (synthesis, query planner, session summaries) need an external backend: point TRIBE_LLM_DIR at a directory exposing lib/types.ts, lib/research.ts, and lib/providers.ts (e.g. bearly's plugins/llm/src). Without it those features degrade to their documented no-LLM paths — loudly, never silently.

TRIBE_RECALL_ENGINE_DIR and TRIBE_INJECTION_DEBUG_DIR override the in-repo engine/recorder locations for forks and experiments; you normally never set them.

Remote Agent Entry Point

Remote hosts should be able to run one of:

tribe-wire mcp --socket /tmp/tribe.sock
bunx tribe-wire mcp --socket /tmp/tribe.sock
npx -y tribe-wire mcp --socket /tmp/tribe.sock

The package name and binary name are both tribe-wire so package runners can use the direct form without a package/binary split.

Repository Layout

packages/
  wire/                # npm tribe-wire; bin tribe-wire
  daemon/              # tribe-daemon broker (private)
  recall/              # tribe-recall engine (private)
  injection-envelope/  # tribe-injection-envelope (private)
plugins/
  claude/              # Claude Code marketplace plugin (tribe@tribe)
docs/
  architecture.md

Development

bun install
bun run test        # vitest, all packages
bun run typecheck
bun run fmt:check

History

Extracted from github.com/beorn/bearly in June 2026 (km bead @km/bearly/19273-tribe-repo-split), with git subtree split history for the recall and injection-envelope packages. bearly retains the LLM toolkit (plugins/llm) and a migration pointer.