Agent Chorus

Let your AI agents talk about each other.

Ask one agent what another is doing, and get an evidence-backed answer. No copy-pasting, no tab-switching, no guessing.

If you use 2+ AI coding agents (Codex, Claude, Gemini, Cursor CLI, Cursor IDE), Chorus gives them shared visibility — no orchestrator required.

chorus read --agent claude --include-user --json

Two problems, one tool:

• Silo Tax — multi-agent workflows break when agents cannot verify each other's work. Chorus gives every agent read access to every other agent's session evidence.
• Cold-Start Tax — every new session re-reads the same repo from zero. A Context Pack gives agents instant repo understanding in 5 ordered docs.

See It In Action

The Handoff

Switch from Gemini to Claude mid-task. Claude picks up where Gemini left off.

The Status Check

Three agents working on checkout. You ask Codex what the others are doing.

From Zero to a Working Query

chorus setup wires every agent on the box in under a minute.

Quick Start

# 1. Install
npm install -g agent-chorus      # requires Node >= 18
# or
cargo install agent-chorus       # requires Rust >= 1.74

# 2. Wire your agents
chorus setup                     # patches CLAUDE.md / GEMINI.md / AGENTS.md, adds .gitignore entries
chorus doctor                    # verify session paths, provider wiring, updates

# 3. Ask any agent in natural language
#    "What is Claude doing?"  /  "Compare Codex and Gemini outputs."  /  "Pick up where Gemini left off."

Or call chorus directly:

chorus read --agent codex --include-user --json

Every response is structured, source-tracked, and redacted:

{
  "agent": "codex",
  "session_id": "session-abc123",
  "content": "USER:\nInvestigate the auth regression...\n---\nASSISTANT:\nI am tracing the auth middleware...",
  "timestamp": "2026-06-02T10:30:00Z",
  "message_count": 12,
  "source": "/home/user/.codex/sessions/2026/06/02/session-abc123.jsonl"
}

Source file, session ID, and timestamp on every response. Secrets auto-redacted before output. Prefer --format markdown for human review.

To reverse everything setup did: chorus teardown (add --global to also drop ~/.cache/agent-chorus/).

What's New in v0.16.0

• Cursor IDE adapter. Chorus now reads both the cursor-agent CLI transcripts and Cursor IDE app sessions through one adapter. If you use the Cursor app, your sessions are now first-class.
• --history=on-demand default. chorus read now returns just the latest session for the current cwd. Closes the 2.5x token-inflation issue measured in the v0.15 field study. Provider snippets carry the contract so consumer agents inherit it automatically.
• cwd_mismatch is now explicit. When --cwd matches no session, the output says so. No more silent fallbacks that read like real data.
• Doctor honesty pass. New info severity, env-var dangling-path detection, git-aware hooks checks, stale-snippet detection. Doctor tells the truth or stays quiet.
• Codex search parity fix. chorus search --agent codex no longer silently returns empty. The read ⊆ search invariant is now enforced for every adapter.
• --help overhaul. Per-subcommand help leads with that subcommand. chorus report --help ships a copy-pasteable handoff JSON schema.

Full changelog and upgrade notes: RELEASE_NOTES.md.

How It Works

1. Ask naturally — "What is Claude doing?" / "Did Gemini finish the API?"
2. Your agent runs chorus — chorus summary, read, timeline, compare, search, diff, send, messages, checkpoint, etc.
3. Evidence-backed answer — sources cited, divergences flagged, no hallucination.

Tenets:

• Local-first — reads agent session logs directly on your machine. No data leaves.
• Evidence-based — every claim tracks to a specific source session file.
• Privacy-focused — auto-redacts API keys, tokens, and passwords.
• Dual parity — Node.js + Rust CLIs ship identical output contracts, conformance-tested against shared fixtures.

Key Capabilities

A taste — see docs/CLI_REFERENCE.md for the full surface.

# Structured digest — files, tools, duration. No LLM calls.
chorus summary --agent claude --cwd . --json

# Chronological view across every agent on the project
chorus timeline --cwd . --format markdown

# What an agent actually touched (Read/Edit/Bash/Write)
chorus read --agent codex --tool-calls --json

# Verify one agent's claim against another
chorus compare --source codex --source claude --cwd . --json

# Audit what got redacted, and why
chorus read --agent claude --audit-redactions --json

# Coordinate without switching tabs
chorus send --from claude --to codex --message "auth module ready" --cwd .
chorus messages --agent codex --cwd . --json

# Broadcast where you left off before ending a session
chorus checkpoint --from claude --cwd .

Supported agents: Codex, Claude, Gemini, Cursor CLI, Cursor IDE. Full capability matrix in docs/CLI_REFERENCE.md.

Context Pack

A context pack is an agent-first, token-efficient repo briefing for end-to-end understanding tasks. Instead of re-reading the full repository on every request, agents start from .agent-context/current/ and open project files only when needed. Local-first, no need to make your repo public.

chorus agent-context init    # creates .agent-context/current/ with templates
# ...agent fills in the <!-- AGENT: ... --> sections...
chorus agent-context seal    # validates and locks the pack

Ask your agent: "Understand this repo end-to-end using the context pack first, then deep dive only where needed."

CI gate: chorus agent-context verify --ci exits non-zero if the pack is stale or corrupt. Internals, sync policy, enforcement: AGENT_CONTEXT.md.

Architecture, in one diagram

Chorus sits between your agent and other agents' session logs. Read-only, evidence-first, no central control plane.

Boundaries: no task router, no scheduler, no autonomous chaining, no live sync stream. Snapshot-based reads from local logs, by design.

Easter Egg

chorus trash-talk roasts your agents based on their session content.

Go Deeper

Every agent session is evidence. Chorus makes it readable.

Found a bug or have a feature idea? Open an issue. Ready to contribute? See CONTRIBUTING.md.