Skip to main content
TRW
Skip to content
TRWDocs — AI Agent Engineering Framework

Docs

Install TRW, choose your clients, and preserve the work worth carrying forward

TRW is the engineering operating layer that sits above your coding client. Install it once, choose the surfaces you actually use, and give each session access to explicitly recorded project knowledge, structured execution, and validation receipts.

Read Quickstart Get Started
Tracked PRDs
3000+
Repository tests
44000+
Recorded sprints
800+

Read in this order

QuickstartLifecycleRequirementsTools

1-line installer

Run the shell bootstrap first

The fastest path to a working TRW setup — installs trw-mcp from PyPI, records the target profiles you select, and applies dedicated client wiring where it exists.

shell bootstrap
curl -fsSL https://trwframework.com/install.sh | bash
  • Downloads the Python installer and authenticates when needed.
  • Lets you select more than one client surface for the repo.
  • Bootstraps the local TRW layer before you open your first session.

Client coverage

One repo can carry more than one client surface

Keep the tools already used by the repository. Select one profile or several in the same setup pass, then review the files each integration can actually write.

7 current installer profiles

Claude Code, Cursor IDE, Cursor CLI, OpenCode, Codex CLI, GitHub Copilot CLI, and Antigravity CLI can be selected for new setups.

Multi-select supported
Claude CodeCursor IDECursor CLIOpenCodeCodex CLIGitHub Copilot CLIAntigravity CLIAider (retired)

What changes after install

Sessions can resume from recorded state

Agents explicitly record learnings and checkpoints; later sessions call TRW to retrieve a bounded, relevant set and re-check it against the current repository. Validation receipts keep the handoff inspectable instead of asking you to trust narration.

  • Knowledge available for recall

    Explicitly recorded learnings can surface as bounded candidates when a later session calls trw_session_start or trw_recall.

  • Right-sized ceremony

    Quick fixes stay fast. High-risk changes get checkpoints, validation, review, and durable handoff.

  • Verification before narration

    TRW treats build results, review gates, and persisted state as the definition of done.

  • Recovery after interruption

    Checkpoints and run state survive context compaction, interrupted sessions, and handoffs.

Without TRW vs With TRW

See how work changes once project state survives the session

Without durable context, every session spends energy reconstructing state. With TRW, the repo can retain selected learnings, checkpoints, and validation history for explicit recall.

Without TRW
  • Learnings lost every session
  • No checkpoints, rework from scratch
  • Skip tests, ship bugs
  • No quality gates, defects cascade
  • Context evaporates on compaction
With TRW
  • Knowledge compounds across sessions
  • Checkpoints survive interruption
  • Automated validation gates
  • Quality audit before delivery
  • Discoveries persist permanently

Install flow

From install to first session

  1. 1

    Choose package-only or full-framework access

    The source-available trw-mcp package supports an unauthenticated install. Create or sign in to an account when you want the authenticated full-framework installer path.

  2. 2

    Run the bootstrap

    The shell bootstrap can install trw-mcp with --allow-unauthenticated, or authenticate and download the full Python installer for guided setup.

  3. 3

    Select your clients

    First-time installs let you select every client surface you want TRW to configure in this repo, including more than one.

  4. 4

    Start your first session

    Open a configured client, call trw_session_start(), verify recalled candidates, and explicitly record learnings, checkpoints, and delivery evidence worth preserving.

What gets added to your repo

The local operating layer lands in the repo

  • .trw/config.yaml

    Project-local TRW configuration, including your selected target_platforms.

  • .trw/

    Runs, learnings, analytics, and the local operating layer that survives session boundaries.

  • Client-specific instruction surfaces

    Examples: CLAUDE.md, .cursor/rules/, .codex/INSTRUCTIONS.md, ANTIGRAVITY.md, or .github/copilot-instructions.md.

Supported clients

Choose the client profiles that match your workflow

Current profiles are separated from retired ones, which cannot be selected and remain listed only so existing installs stay cleanable. Integration depth varies by client: inspect the files each profile writes before standardizing a team workflow.

7 current profiles7 installer-selectable1 retired profile, uninstall onlyNo client switch requiredOne repo can select several

Claude Code

Hooks & agents

End-to-end TRW integration for local CLI workflows that lean on hooks and durable instructions.

Best for
Local CLI workflows that want repo-native hooks, agents, and guidance

TRW writes

CLAUDE.md.claude/hooks/.mcp.json

Cursor IDE

IDE-native

TRW translated into Cursor’s IDE rules, hooks, MCP config, and editor-first automation.

Best for
Interactive editor-first workflows inside Cursor

TRW writes

.cursor/rules/.cursor/hooks/.cursor/mcp.json

Cursor CLI

Headless-safe

TRW adapted for headless Cursor runs where explicit permissions and CLI-safe defaults matter.

Best for
cursor-agent, CI, and scripted terminal workflows

TRW writes

AGENTS.md.cursor/cli.json.cursor/mcp.json

OpenCode

Command-first

TRW translated into OpenCode’s instruction, command, and agent surfaces.

Best for
Terminal workflows built around commands and repo-local instructions

TRW writes

opencode.json.opencode/commands/.opencode/agents/

Codex CLI

Codex-native

TRW adapted to Codex’s workflow and context constraints without changing the core loop.

Best for
Codex sessions that want tighter context control and repo-local guidance

TRW writes

.codex/config.toml.codex/INSTRUCTIONS.md.codex/agents/

GitHub Copilot CLI

GitHub-native

TRW translated into GitHub-native repo instructions, hooks, and agent surfaces.

Best for
GitHub-native repos adopting agent workflows and path-scoped guidance

TRW writes

.mcp.json.github/copilot-instructions.md.github/instructions/.github/agents/

Antigravity CLI

Antigravity-native

TRW writes Antigravity-native instructions, MCP registration, subagents, and supported hooks.

Best for
New Google CLI workflows that need large context and native agent surfaces

TRW writes

ANTIGRAVITY.md.antigravitycli/settings.json.antigravitycli/agents/

Aider

Retired; manual only

The Aider target was retired because it never had a native adapter. Current installation rejects the identifier.

Best for
Existing manual workflows that do not select Aider during TRW initialization

TRW writes

manual local CLI

Next stops

Read the docs in the order you will use them

Quickstart gets you in. Concepts explains the model. Configuration and troubleshooting answer the operational questions that usually show up next.

Quickstart

See the installer flow, the verification checklist, and the first-session walkthrough.

Core Concepts

Understand sessions, runs, learnings, ceremony, and the compounding memory loop.

Configuration

Learn how target_platforms, build thresholds, and platform settings actually behave.

Troubleshooting

Fix install issues, MCP connection problems, and client-specific startup failures.

Quickstart