Skills for coding agents

Toft publishes two agent skills: the instructions a coding agent loads before it touches Toft. They live beside the client, in larchway-labs/toft-cli, and are updated with every toft release.

Skill	Load it to	Source
toft-cli	Ship an app: sign in without a browser, the deploy loop, exit codes and --json shapes, jobs and their runs, what toft refuses and why	SKILL.md
toft-sdk	Write the app: PORT, HOST and BASE_PATH, how the build detects a framework, toft.toml, and the SDK for Python, Node, R and Go	SKILL.md

Install them

toft carries both skills, matched to the commands that binary has. Write each into the folder your agent reads skills from:

Agent	Folder
Claude Code	~/.claude/skills
Codex, Cursor, and other agents that read the Agent Skills format	~/.agents/skills

dir=~/.claude/skills        # or ~/.agents/skills
for s in toft-cli toft-sdk; do
  mkdir -p $dir/$s
  toft skill $s > $dir/$s/SKILL.md
done

For one project only, use .claude/skills or .agents/skills in that repository instead. toft skill with no name lists the skills, and it needs no sign-in. Run the loop again after upgrading toft.

To fetch them without toft installed, download them from the repository:

for s in toft-cli toft-sdk; do
  mkdir -p $dir/$s
  curl -fsSL -o $dir/$s/SKILL.md \
    https://raw.githubusercontent.com/larchway-labs/toft-cli/main/skills/$s/SKILL.md
done

A copy from the repository describes the latest release, which may be newer than your toft; toft skill is the copy that matches.

Give the agent a key, not a browser

toft login opens a browser and waits for a person. An agent signs in with a key instead:

printf '%s' "$TOFT_KEY" | toft login --token-stdin https://toft.example.com

Use a personal key (toftu_, Your account → Credentials) for an agent working on your behalf, or a deploy key (toftd_, Admin → Deploy keys) for one running unattended. See Signing in.

Have the agent run toft whoami first. It says what the key may do and what the install allows, so the agent learns its limits before trying a deploy.

What an agent gets back

• toft check says whether a deploy would be refused, without one.
• toft deploy --verify /path fails unless the page answers.
• toft fetch /path shows what the app returns, with the body capped at 8 KB so a large page does not flood the agent’s context.
• Flags go anywhere on the line, so an agent does not have to reorder a command it composed.
• With --json, every refusal carries code, retryable and often a hint. An agent should retry only when retryable is true, and pass the hint on to a person when the next step is theirs. See Errors as JSON.

Tip

No local runtime is needed to deploy. The install builds the app, so an agent with toft and a key can create an app, push a directory, read the build log and roll back without Node, Python or Go on its machine.