agentic-loadout

One command that gives a project a coding-agent setup, for Claude Code, Cursor and Codex.

• Memory: agemem keeps current state, decisions and durable notes in memory/, so a new session picks up where the last one stopped. For Claude Code, agemem also disables Claude's built-in auto-memory for the project, to avoid double memory. The agent is told to suggest a memory-wrapup skill when a task is finished and before a push; it works agemem's queue of conflicting or repeated notes (agemem 0.3.0+). Subagents can read that queue but never resolve it: only the main agent writes memory.
• Spec-driven development: the sdd-feature skill takes a feature through design, spec, tests, implementation and an independent grade, with an approval gate between phases.
• Session history: when the Entire CLI is installed, capture agent sessions as checkpoints linked to commits. Checkpoint uploads are disabled by default.
• Subagents: an orchestrator delegates to architect (design), coder (implementation), code-grader (review) and utility (commits, checks), following one shared workflow.
• Language coders: coder-go, coder-python, coder-typescript and coder-swift come with code-navigation MCPs, so they rename and refactor through the compiler or type checker instead of grep and sed.

Each piece is handed to the tool that already installs it; agentic-loadout only writes what no tool covers.

Needs uv (agemem, Python coder), Node.js 20+ (skills, TypeScript coder), and Go with gopls for the Go coder.

--python 3.12 makes uvx use a tested Python instead of whichever is newest on your machine. To pin the loadout version, run uvx --python 3.12 agentic-loadout@0.6.0 .; individual dependencies may have their own version policy.

Use

uvx --python 3.12 agentic-loadout .                  # detects your clients and languages, asks before installing
uvx --python 3.12 agentic-loadout . --dry-run        # prints the commands and the files they write; changes nothing
uvx --python 3.12 agentic-loadout . --personal-settings --yes

Project configuration stays inside the project. Nothing is written to ~/.claude, ~/.codex or ~/.cursor. Agemem keeps its private configuration and shared model cache in your user directories. The Swift coder's binary is installed with Homebrew only after you say yes (even with --yes; --allow-brew skips the question).

Rules go to AGENTS.md, or to CLAUDE.md if the project already has one that isn't just an @AGENTS.md import. Each section sits between <!-- agentic-loadout:begin NAME --> markers, so re-running updates it in place.

What gets installed

Worktrees

The worktree skill runs a task in an isolated git worktree. /worktree create branches from the current branch and sets up the environment, save commits, apply merges into the branch the worktree came from, and delete removes it. agentic-loadout adds a short rule to your instructions: when your checkout has uncommitted changes the agent didn't make, it asks once whether to use a worktree, and after the task it offers save, apply and delete. It never merges, pushes or deletes without your yes. Leave it out with --without worktree.

Entire session history

Entire records agent sessions and connects captured work to Git commits with checkpoints. It gives an agent or reviewer historical context about the request, discussion and files behind a change. Agemem complements this with the current project state and selected durable facts and decisions: an agent can inspect a relevant checkpoint, verify a useful lesson against current sources, then capture it in agemem with provenance. Neither a transcript nor a memory note replaces the code, tests or an independent review.

Entire is included by default if its CLI is already installed and the target is a Git repository. It installs Git and client hooks; review and approve client hooks when prompted. Loadout does not install Entire itself or initialize Git. Use --without entire to omit it. --no-hooks skips Entire as well as agemem's hooks. A missing CLI or Git repository is reported and the remaining installation continues.

Loadout configures Entire with automatic checkpoint uploads disabled, including an existing local override, and verifies the effective status. Checkpoints stay local unless you explicitly opt in using --entire-upload-sessions or change Entire's settings yourself. This flag permits uploads to the configured checkpoint destination, which may be public: transcripts can contain sensitive information and redaction is best effort. Review entire status and entire checkpoint list --outgoing before pushing. This setting does not prevent manual pushes or later changes. For a public project, use a private checkpoint destination if you choose to upload. Existing worktrees and local settings can differ; check the effective status in each one.

Loadout adds project instructions for using checkpoints as untrusted evidence; no separate Entire skill is required. Agents can use entire agent-help for the commands supported by the installed CLI. Session capture depends on the client hooks being active; it is not a guarantee that every action or subagent is recorded.

Language coders

Each standard coder comes with a -small counterpart for bounded work and an independent code-reviewer. For example, use coder-python-small for a clearly specified Python change and coder-python for more complex work. Both use the same navigation MCP. When usage limits or parallel work are a concern, the orchestrator can suggest small coders and follow your preference. The reviewer checks the diff and complements the SDD final grade.

Claude uses Haiku for small coders and Sonnet for the reviewer. Codex uses Luna and Terra for these new roles; existing roles keep their current settings. Cursor uses configured models but may fall back when a model is unavailable: check which model actually ran. Model availability varies by account. Edit a generated agent file to change its model; re-runs preserve that edit unless --force is used. Read-only reviewer instructions are a convention, not a sandbox boundary.

Claude and Codex small-coder files explicitly set high reasoning effort: effort: high in Claude's agent frontmatter and model_reasoning_effort = "high" in Codex's agent TOML. Claude invocation settings or the CLAUDE_CODE_EFFORT_LEVEL environment variable can override its agent setting. Cursor's generated agent files do not specify reasoning effort. Existing customized agent files are preserved on re-runs; use --force to replace them with the generated defaults after reviewing your edits.

Re-running an existing installation adds the small agents, reviewer and .agents/delegation-tiers.md. Re-running with components omitted does not remove previously installed files; use trash on unwanted generated agents and guidance manually after reviewing their paths.

A coder is one unit: the coder-<lang> subagent, the language's navigation MCP and a routing rule. They are always installed together, because without the subagent's instructions the agent doesn't use the MCP.

uvx --python 3.12 agentic-loadout coders                    # detect languages, confirm, install
uvx --python 3.12 agentic-loadout coders go typescript  # or name them (go, python, typescript, swift, all)
uvx --python 3.12 agentic-loadout coders --list             # catalogue, what is detected, what is installed

The Go MCP runs via go run github.com/illescasDaniel/codenav-go-mcp/cmd/codenav-go-mcp@latest. Go fetches the public module when needed; @latest can select a newer release over time. Re-running the installer migrates the exact old WIP binary path to this launcher while retaining its environment settings; custom MCP launchers are left alone. Cursor's configuration sets the workspace explicitly. Projects with build tags should configure CODENAV_GO_MCP_BUILD_FLAGS for their build, so gopls sees the right files.

Keeping it out of git

--personal-settings writes rules to CLAUDE.local.md (Claude Code only; it starts with @AGENTS.md if you have one), Claude settings to .claude/settings.local.json, and adds every file the run created to .git/info/exclude. A file that already existed and was changed (for example a committed .mcp.json) can't be hidden by git, so the final message lists it with your options. With Cursor or Codex selected the rules go to AGENTS.md and you are told. --git-exclude does the exclusion without the personal files.

Flags

--tools claude,cursor,codex · --with go,python,swift,typescript|all · --without sdd|agemem|agents|coders|worktree|entire · --memory-model jev|qsiv · --no-hooks · --model ALIAS · --claude-compat · --dry-run · --no-input · --yes · --allow-brew · --entire-upload-sessions · --force. With --no-input and no --yes/--with, no language coder is installed and the command to run is printed.

The Jev key is never asked for here. Set AGEMEM_JEV_API_KEY or edit ~/.config/agemem/config.yaml.

When re-running in an existing Agemem project, leave out --memory-model to preserve its configured provider, including cascade (QSIV with Jev fallback), and avoid repeating provider setup. An explicit --memory-model that conflicts with the project's configured provider is rejected before installation; change provider settings separately.

Development

uv run pytest

The design and the decisions behind it are in original implementation plan. The real Homebrew path can only be checked on macOS.