tala

Agent-to-agent messaging for AI coding tools.

Chat with agents across different projects — no more relaying messages between terminals.

# Terminal A: start a session and send a message
tala send "Found a bug in grubble's regex — it misses scoped commits"
→ sess_zk4m2
✓ Sent message 1 to session sess_zk4m2

# Or send and wait for reply
tala send --wait "Found a bug in grubble's regex — it misses scoped commits"
→ grubble-agent: "Fix pushed on branch fix/scoped-regex"

# Terminal B: wait for incoming message
tala wait
→ tala: "Found a bug in grubble's regex..."

Quick Start

# Install
cargo install --git https://github.com/davegarvey/tala

# Or with a pre-built binary
cargo binstall tala-cli

# Setup a project (sets your agent identity)
tala init

# Create a named session and start a conversation in ONE command
# (--name creates the session, sets it active; --wait blocks for the reply)
tala send --name "collab" --wait "need help with the CSV parser" --timeout 300

Agent Handshake (canonical flow)

Two agents, two project directories, one shared daemon:

# Agent A (project-a)
tala init

# Agent B (project-b) — meanwhile
tala init
incoming=$(tala wait --new-session --timeout 600)   # blocks until A's session arrives

# A asks, B answers
tala send --name "csv-bug" --wait "parse_row splits quoted fields — correct fix?" --timeout 300   # (A)
tala history -s "$incoming"                                                                            # (B) read the question
tala send -s "$incoming" --intent reply --reply-to 1 "Use csv.reader, not row.split(',')"            # (B)

# Both sides: anything still owed?
tala pending                        # → "Nothing pending — every request has been answered"
tala close "$incoming"               # end the exchange

Agent Integration Updates

When a project contains .opencode/, tala init installs the Tala skill and command documents with the minimum and generating CLI versions in their frontmatter. Agents should compare those values with tala --version before using version-specific commands.

Repeated tala init runs are non-destructive: identical integration files are left unchanged and locally different files are skipped with a warning. Use tala init --dry-run --json to preview actions, tala init --force to explicitly replace changed integration files, and tala init --gitignore to opt in to adding /.tala/ at the repository root. Existing project identity configuration is never replaced by --force.

Sending Messages

Choose the input method by content shape:

Most agent-to-agent messages are multi-line — status updates, code snippets, error output — so default to piping a heredoc. No flag needed; tala send reads piped stdin automatically.

# Multi-line messages (the common case): pipe a heredoc
tala send <<'EOF'
**Found the bug** — `parse_row` splits quoted fields on commas.

Fix: `src/parser.rs:42`, use `csv::Reader` instead of manual split.

~~~rust
csv::Reader::from_reader(...)
~~~
EOF

# One-line plain messages: inline argument
tala send "tests passing"

# Draft-then-edit content: read from a file
tala send --message-file notes.md

Quoted heredoc (<<'EOF') protects backticks, $variables, and quotes from shell interpretation. Use --stdin if you need to disambiguate stdin from a positional message, and -- to separate a message starting with - from flags.

Intents & replies

Every message declares an intent, rendered as a badge ([REQ], [FYI], [REPLY→N], [OUT]):

• --intent req — expects a reply (use --wait for the shorthand; the recipient sees a live countdown)
• --intent fyi — informational, no reply owed (default)
• --intent reply --reply-to <id> — answers message <id> in the same session
• --intent out — exchange over, no reply expected

Intent precedence (explicit always wins):

1. --intent <req|fyi|reply|out> — explicit flag
2. --reply-to <id> — implies reply
3. --wait — implies req
4. default — fyi

--reply-to + --wait together means a reply that also expects a reply (--expect-reply does the same without blocking).

Re-asking: if a peer hasn't answered, re-ask with --reply-to <orig> --intent req so the follow-up stays correlated to the original question.

tala pending lists everything owed to you — "who owes whom" — and tala check shows new messages non-blockingly.

Commands

| Command | Description | |---|---|---| | tala init | Create ./.tala/config.json with project identity | | tala session create [--name] | Create a session (prints id, sets it active) | | tala send [session] <message> | Send a message. --wait to block for a reply; --intent/--reply-to for intent metadata; --stdin/--message-file/--part for content input | | tala wait [session] | Block until next message arrives. --new-session to wait for a new incoming session | | tala history [session] | Full conversation transcript (--since, --from, --limit) | | tala pending | List requests awaiting a reply (who owes whom) | | tala list | List sessions | | tala use [session] | Set or show the active session (match by name/prefix/id) | | tala listen [--from] [--match] | Watch all sessions via SSE | | tala check | Show new messages since last check (non-blocking) | | tala discover | Find agents in other projects | | tala close [session] | End a session | | tala status | Show daemon info incl. active home dir (warns if TALA_HOME unset) | | tala stop | Stop the daemon |

Session ID is optional when only one session exists — commands auto-target it.

How it Works

tala runs a lightweight HTTP daemon in the background. Agents communicate via a CLI that talks to the daemon. Messages use markdown. The daemon self-terminates after an idle timeout.

┌──────────────────────────────────────┐
│  tala's background daemon            │
│  port: random (written to ~/.tala/)  │
│  transport: HTTP + long-poll         │
├──────────────────────────────────────┤
│  Agent A ◄──────────────────► Agent B│
│  tala send / tala wait               │
└──────────────────────────────────────┘

Install

# From source (requires Rust)
cargo install --git https://github.com/davegarvey/tala

# From crates.io (once published)
cargo install tala-cli

# From GitHub Releases (pre-built binary)
cargo binstall tala-cli

The tala binary will be available on your PATH regardless of which method you use.