TeLLMphone ☎️

Let your LLMs call each other.

TeLLMphone is a local MCP server that lets coding agents place calls to each other, or to another headless session of the same agent with a different personality or model. Claude Code can ring Codex, Gemini, or Grok with a question about the current project and get an answer back, keep that conversation going across multiple turns, or leave a voicemail for the next agent session to pick up. Conversations survive interruptions on both sides, and the callee can be given a saved personality and a specific model.

The second opinion you want is usually installed on the same machine, one terminal over — and you're tired of being the copy-paste layer between two AIs. Now they can just call each other.

Currently supports Claude Code, Codex, Gemini, and Grok; other agents can be added as plugins.

Requirements

• macOS or Linux
• Python 3.11+ and uv
• The agent CLIs you want to connect (claude, codex, agy, grok), installed and logged in

Install

uv tool install tellmphone   # or: pipx install tellmphone
tellmphone install

install registers the MCP server with every agent CLI it finds — via claude mcp add, codex mcp add, grok mcp add, or Gemini's shared MCP config — and sets the config needed for non-interactive tool approval where the CLI requires it. It is idempotent; rerun it if you move the checkout. tellmphone uninstall removes the registrations (it does not delete ~/.tellmphone/).

Usage

You talk to your agent; your agent works the phone. In a Claude Code session:

Call codex and get a second opinion on this migration. Use the grumpy-reviewer personality.

Claude will place the call, relay the answer, and can keep the thread going with follow-ups. In the other direction, start a Codex session in the same project and ask it to check its messages.

That's what the demo above shows: Claude, asked for a second opinion, places the call itself and then argues with the answer. You can also work the phone yourself:

tellmphone call codex "Sanity-check the file locking in src/store.py." --personality grumpy-reviewer
tellmphone reply "Fair. Would any of it break on Windows?"
tellmphone messages
tellmphone show call-7f3k9q2m
tellmphone gc --days 30

The tools the agents get:

How it works

┌─────────────┐   MCP (stdio)   ┌──────────────────┐   headless CLI   ┌─────────────┐
│ Claude Code │ ──────────────► │   TeLLMphone     │ ───────────────► │ codex exec  │
│  (caller)   │  call/reply/…   │   switchboard    │  spawn / resume  │  (callee)   │
└─────────────┘                 └──────────────────┘                  └─────────────┘

There is no daemon. Each agent runs its own TeLLMphone instance over stdio; shared state lives under ~/.tellmphone/. A live call records the message, spawns one detached callee turn (codex exec, claude -p, agy --print, grok -p) in the project directory, and returns the stable call id. When the turn finishes, its answer lands in the caller's mailbox. Replies start another one-shot turn and resume the callee's native session where the CLI exposes one. If a native session is lost or not available, the stored transcript is replayed into a fresh one. Callees run in their CLI's read-only/sandboxed mode unless you allowlist a project for writes, and a hop limit keeps agents from chaining calls indefinitely.

The internals lean into the name: the switchboard routes calls, the phonebook lists who you can dial, a busy line means the callee is still thinking, voicemail holds messages for agents that aren't running, and the hop limit stops two agents from playing telephone forever.

Details, including the security model, are in docs/DESIGN.md.

Configuration

~/.tellmphone/config.toml:

[defaults]
timeout_s = 300   # how long the CLI waits for a live turn before returning ringing
max_hops = 2      # agent-to-agent chain depth limit

# default model per callee; a call's explicit model argument wins.
# Use whatever model ids your CLI accepts.
[agents.codex]
model = "gpt-5.5"

# standing write grant; callees run read-only unless a project is
# allowlisted here or the caller passes write=true for a specific call
[permissions."/path/to/project"]
write = true

The file is optional — everything has a sensible default. For one-off delegation you don't need it at all: a top-level caller can pass write=true on a single call (--write on the CLI). Agents that are themselves callees can't grant write, so access never spreads down an agent-to-agent chain.

Personalities are Markdown files with a small frontmatter block (name, description) followed by the system prompt. Twelve builtins ship with the package and update with it — critics (grumpy-reviewer, security-auditor, sycophancy-cop), thinking partners (rubber-duck, devils-advocate, architect, the-algorithm, transaction-cost-accountant), and builders (debugger, test-engineer, evidence-engineer, tiny-hacker); your own live in ~/.tellmphone/personalities/. A user file with the same name as a builtin replaces it, and one with disabled: true hides it. Callers select personalities by name and never send system prompts inline.

Data

Call records and transcripts live under ~/.tellmphone/projects/. Transcripts contain whatever flowed through the conversation, which for a coding agent usually includes your code. Everything is plain JSON owned by your user; delete a call directory (or all of ~/.tellmphone/) to purge.

Development

uv sync
uv run pytest

The test suite runs against fake claude/codex/agy/grok executables, so it needs neither CLI installed nor network. From a checkout, uv run tellmphone install registers your development version with your agents.

TeLLMphone is a project powered by The California Open Source Company. Licensed under Apache 2.0.