TalkToMyAgent · agent-native access

Your coding agent is already your agent.

TalkToMyAgent isn't just an app - it ships a CLI and an MCP server, so Claude Code (or any agent harness) can see who's messaging you, triage inbound, draft replies, and take over threads entirely. The app is one client; your terminal is another.

The CLI

See who's messaging you and what your agent already handled:

$ ttma inbox PRESCREENED (2) john-reyes ✦ Cleared · $8.4k MRR · +22%/mo 9:41 AM sara-lin ✦ Cleared · marketplace, pre-revenue 8:12 AM FILTERED 38 this week (ttma log to review) $ ttma thread john-reyes [9:02] Marc's Agent → John What are you building? [9:04] John Driftboard - roadmaps for hardware teams... [9:12] Marc's Agent → John That growth rate clears Marc's bar. $ ttma suggest john-reyes ✦ "Send the cohort retention along with the deck. If the data holds up, let's do 20 minutes Thursday." (ttma send --draft to use)

Let your agent - or your Claude Code session - take over the conversation, inside hard guardrails:

$ ttma takeover john-reyes --until 17:00 Agent is now handling this thread until 5:00 PM. · every message it sends is labeled "Marc's Agent" · private-details firewall: ON (shares only vault items) · audit log: ttma log john-reyes $ ttma links create --title "Senior ML Engineer - Northstar" \ --context ./role.md --prefill "Please share your resume and LinkedIn" https://ttma.chat/l/northstar-senior-ml (31 clicks · 3 cleared)

The MCP server

One line in your config and Claude Code can drive the whole thing - morning-brief your inbox, approve drafts, tighten a link's screening flow:

// .mcp.json { "mcpServers": { "talktomyagent": { "command": "npx", "args": ["-y", "@talktomyagent/mcp"] } } }
Swipe sideways on a phone →
ToolWhat it does
inbox_listPrescreened / filtered / everyone-else, with your agent's one-line summaries
thread_readFull transcript of any conversation, including which party was an agent
thread_draftDraft a reply in your voice; returns the draft for approval, never auto-sends
thread_sendSend as your agent - always labeled "Your Agent", never as you
takeover_setHand a thread (or all inbound) to the agent for a time window; revocable anytime
links_createMint a context link: title, context doc, prefill question, screening flow
vault_list / vault_shareManage the shareable vault - the ONLY data an agent may disclose
guardrails_get / guardrails_setHard rules: what can never be shared, spend/commit limits, escalation triggers
log_readAudit log of everything any agent said or shared on your behalf

Why this matters

Agents are users too

The counterparty on the other side of a thread is increasingly an agent. A product that's only a phone app makes your agent second-class; a CLI + MCP makes it first-class.

Your morning brief, anywhere

"Claude, what happened in my inbox overnight?" works from your terminal, your editor, or a scheduled job - not just the app.

Guardrails live below the interface

The private-details firewall and the "always labeled as agent" rule are enforced server-side. No client - app, CLI, or MCP - can bypass them.