SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI agentsTrellis

Trellis

Trellis is an out-of-the-box engineering framework for AI programming developed by Mindfold, hosted on GitHub (mindfold-ai/Trellis). It addresses the problem that AI writes code quickly but every session starts from scratch, not remembering project conventions and team requirements: by persisting specs, tasks, and memory into the code repository, it makes any coding agent work according to your engineering standards. Its workflow is a four-stage closed loop—Plan (trellis-brainstorm asks questions step by step and generates prd.md), Implement (the trellis-implement sub-agent writes code), Verify (the trellis-check sub-agent reviews against the specs and runs lint, type checking, and tests), and Finish (trellis-update-spec writes new experience back to .trellis/spec/). Users only need to describe requirements in natural language, let the AI gradually sort out the PRD, and then have the AI call Trellis Implement to automatically verify the results. After completion, enter /trellis:finish-work to archive the task and update the log. Trellis is a project layer across multiple coding agents and IDEs. Individual developers can rely on its memory and repeatable process, while teams gain shared standards, task boundaries, reviewable context, and platform portability. It uses the AGPL-3.0 license, and the official offering provides entry points for Simplified Chinese, documentation, quick start, supported platforms, and use cases.

Open website

About this tool

Purpose

Trellis is an out-of-the-box engineering framework for AI programming. The pain point it addresses is that although AI writes code quickly, every session starts from scratch and does not remember your project, conventions, or team requirements. Trellis persists specs, tasks, and memory into your code repository, so that any coding agent works according to your engineering standards.

Main capabilities

Trellis runs a four-stage loop, accompanied by automatically invoked skills and sub-agents:

  • Plan: trellis-brainstorm sorts out requirements one question at a time and writes prd.md. Research-intensive items are handed off to the trellis-research sub-agent. The result is curated specs and research files referenced from implement.jsonl / check.jsonl.
  • Implement: the trellis-implement sub-agent writes code according to the PRD, with curated context automatically injected, and does not make git commits.
  • Verify: the trellis-check sub-agent reviews the diff against the specs and runs lint, type checking, and tests, fixing what it can on its own.
  • Finish: run the final checks, after which trellis-update-spec promotes new experience back into .trellis/spec/, so that the next session starts smarter.

Around these files, Trellis adds scoped specs, task PRDs, workflow gates, workspace memory, and platform-aware generated files. The official description says these files are useful entry points, but they often become monolithic.

Usage

The workflow is simple:

  1. Describe what you want in natural language.
  2. Brainstorm with the AI one question at a time until the PRD is clear, then start implementation.
  3. Let it run—the AI will call Trellis Implement and automatically verify the results against the specs, lint, type checking, and tests.
  4. When the work is complete or the session context fills up, enter /trellis:finish-work. Trellis will archive the taskand update journals.

The official site provides entry points for Simplified Chinese, documentation, quick start, supported platforms, and use cases. For configuration details, see the quick start and supported platforms guides.

Applicable scenarios

  • Individual developers: use it to gain memory and repeatable workflows.
  • Teams: gain greater benefits, including shared standards, task boundaries, reviewable context, and platform portability.
  • Individual workspace journals are stored separately per developer, while shared specifications and tasks remain in the repository and can be reviewed and improved like other project artifacts.
  • Many teams start by having AI draft specifications from existing code, then manually tighten the important parts. Trellis works best when high-signal rules are kept explicit and placed under version control.

Limitations and boundaries

  • Trellis is not a tool exclusive to any single coding agent: it is a project layer that works across multiple coding agents and IDEs.
  • Specifications need not be entirely handwritten: Trellis works best when high-signal rules are kept explicit and versioned.
  • Regarding the ablation feature: trellis ablate temporarily removes all supported, project-owned Trellis surfaces after creating a verified recovery transaction outside the project. Then start a brand-new agent session for comparison, and run trellis restore to restore the exact previous state. Use --dry-run to preview either operation. The private recovery transaction contains the exact .trellis tasks, specifications, and workspace bytes, which may contain sensitive text written by users; the transaction is retained until restore successfully verifies.
  • Ablation is different from trellis uninstall (permanent removal) and TRELLIS_HOOKS=0 (hooks only). Ablation does not start agents, manage worktrees, hide Git changes, or remove the global CLI, channel logs, or host transcripts. If managed paths change during ablation, restore will refuse all writes until the conflict is resolved.
  • The above are explicit statements in the current materials; facts such as pricing, the specific list of supported platforms, and version numbers are not yet provided in the materials.

License and source

Trellis is built by Mindfold, is licensed under AGPL-3.0, and its code is hosted in the GitHub repository mindfold-ai/Trellis. The official site also provides entry points for community and resources.

Related tools

Browse category