Trellis is a bulk execution framework for shipping large-scale software on a schedule. Install it once, wire it to cron, and your backlog ships itself.
It runs as a skill inside your agent harness. One install, six commands. Each command is stateless and idempotent. Running one twice produces no duplicate work, making the pipeline safe to run on a schedule.
Research before implementation. Tests before review. Comments resolved before merge. The full SDLC runs unattended.
Clone the repository, or copy the skills/trellis folder directly into your agent's skills directory.
Add LINEAR_API_KEY and GH_TOKEN to your environment. Each command lists which it requires.
Invoke /trellis vision-roadmap to generate your first roadmap and backlog, or drop any single command into an existing workflow.
$ git clone https://github.com/drewdoebereiner/trellis.git
$ cp -r trellis/skills/trellis ~/.claude/skills/
A trellis doesn't control the plant. It gives it structure to grow along.
Structure
The framework defines skills and shape. Agents traverse a lattice of composable capabilities, finding the path of least resistance through your codebase.
Autonomy
Six sub-commands cover the full SDLC loop, from roadmap to merged PR, with no manual handoffs between stages. Wire it to cron once and step away.
Scale
Most agent tooling handles one feature at a time. Trellis processes your entire backlog, PR queue, and test coverage gaps in a single automated pass, every night.
Most agent tooling helps you write one feature at a time. Trellis processes your entire backlog, PR queue, and test coverage gaps in a single automated pass, every night if you want.
Trellis calls Linear and GitHub directly via curl rather than through MCP servers. In bulk runs with 5 to 25 parallel subagents, MCP overhead compounds. Direct calls return exactly what is needed, leaving more room for actual work.
Skills are plain markdown files with no build step and no dependencies. Any agent harness that can load and execute skills can run Trellis. Claude Code, Cursor, Gemini CLI, and Codex all work out of the box.
Run the full SDLC loop end-to-end, or drop a single sub-command in wherever you need it. Each stage is independently useful and can run without the others. The pipeline is additive, not monolithic.
Each command maps to a stage of the software development lifecycle. Run them in sequence for a full automated sprint, or drop one in wherever you need it. All invoked as /trellis <command>.
Because every command is stateless and idempotent, the entire pipeline is safe to run on a cron schedule. No state files. No locks. No cleanup. Running a command twice never creates duplicate work.
Trellis tracks its own progress through Linear ticket states and GitHub PR metadata. Research agents skip tickets they have already commented on. dev-backlog only picks up unstarted tickets. fix-pr-comments only touches PRs with open feedback.
Wire the pipeline into cron once. Implementation runs every weeknight, PR review every afternoon, comment fixes before standup. No manual triggers needed.
Trellis is harness-agnostic. Copy the skills folder into your agent's skills directory and the commands are available immediately. No build step, no configuration beyond environment variables.