Claude Code loop auto-completes Linear tickets via Telegram chat

The post describes a production agentic coding setup built by the Niptao team, where a Claude Code session running on a loop acts as a software engineer managed entirely through a Telegram group chat. A plain-text message like "bug: demo revision limit" becomes a Linear ticket; the bot proposes taking it, receives a "go" approval, announces the start of work, and roughly forty minutes later posts a PR link. The PR includes a root-cause section, the fix, stated assumptions, and a regression test.

The architecture has four components: a Claude Code skill (a Markdown file describing the orchestration procedure), a 180-line stdlib-only Python script bridging Telegram, Linear via MCP, and git worktrees for isolation.

The architecture has four components: a Claude Code skill (a Markdown file describing the orchestration procedure), a 180-line stdlib-only Python script bridging Telegram, Linear via MCP, and git worktrees for isolation. The orchestrator never edits code itself — it reads, decides, and communicates. Each ticket is built by a fresh subagent spawned in its own git worktree branched off the integration branch; the worktree is discarded after the PR is opened. Linear handles three roles simultaneously through the MCP connection: it is the work queue (tickets labeled `agent` in Todo state, processed by priority), the state machine (three labels — `agent`, `agent-blocked`, and `manual` — drive the entire workflow), and the agent's persistent memory (every question and answer is mirrored as a ticket comment, allowing the loop to be killed and restarted cold without losing context). The team wrote zero lines of custom issue-tracker integration, inheriting all Linear changes automatically through the existing MCP connection.

Key facts

1. 01A Claude Code session running in a loop on a laptop acts as the autonomous engineer — no framework like LangChain or custom orchestrator is used.
2. 02The entire system is one Markdown skill file and one 180-line stdlib-only Python script.
3. 03Linear is connected via MCP (Model Context Protocol) with zero custom integration code written by the team.
4. 04Linear serves three roles: work queue, state machine (via three labels: `agent`, `agent-blocked`, `manual`), and persistent agent memory.
5. 05Each ticket is built in an isolated git worktree by a fresh subagent, which runs tests, opens a PR, and is then discarded.
6. 06The Telegram bridge uses only two Bot API calls — `sendMessage` out and long-polled `getUpdates` in — with no inbound infrastructure exposed to the internet.
7. 07Both the skill file and the Python script are publicly available at singlas/ai-dev-prompts.

Topics