AI Agent Integration

Use Claude Code, Cursor, Codex, or any terminal-based AI agent to control milmil via the milmil CLI.

milmil exposes its agent automation surface through a single CLI binary (milmil) — the same tool you use as a human admin. Because every modern AI agent can call shell commands, no agent-specific integration is required.

Quick start

Install the CLI

Download the latest release for your platform from the GitHub Releases page — the cli-v* tags hold platform binaries for linux, darwin, and windows on amd64 and arm64.

# Example: Linux amd64
VER=0.1.0
curl -L https://github.com/milmil-dev/milmil/releases/download/cli-v${VER}/milmil_${VER}_linux_amd64.tar.gz | tar xz
sudo mv milmil /usr/local/bin/
milmil version

Generate an API token

In milmil's web UI: Settings → API Tokens → Create. Name it after the agent that will use it (e.g. claude-code-laptop). Copy the mlml_… token immediately — it's shown only once.

Authenticate the CLI

milmil auth login --server http://localhost:8080
# > Token (mlml_...): mlml_...
# Logged in as "admin" to http://localhost:8080

Verify:

milmil auth status

The credentials are stored at ~/.config/milmil/credentials with mode 0600. Override per-command with --server / --token flags or the MILMIL_SERVER / MILMIL_TOKEN env vars.

Tell your agent about milmil

Run this from the directory you'll work with the agent in:

# Claude Code:
mkdir -p .claude/skills/milmil
milmil generate-skill --format claude > .claude/skills/milmil/SKILL.md

# Codex / Aider / any agent that reads AGENTS.md:
milmil generate-skill --format agents-md >> AGENTS.md

# Cursor:
milmil generate-skill --format cursor > .cursorrules

Each shim is a 3-5 line redirect telling the agent to run milmil agents-guide whenever the user mentions milmil. The full guide (roughly 140 lines, embedded in the binary) covers recipes, pitfalls, and the audit workflow.

What the agent can do in v0.1

The CLI surface in v0.1 maps each subcommand to a single server endpoint:

Subcommand	Effect
`library list / add / scan / stats`	Manage libraries
`search anime <q>`	Local DB fuzzy title search
`episode list / watch-url`	Inspect a Bangumi-matched anime's episodes
`watch resolve <title> --episode <n>`	Title + episode number → web watch URL + stream URL
`match auto / list / apply / undo`	Manual + scripted matching, with surgical and window undo
`audit list / show`	Browse and inspect every autonomous action
`token list / revoke`	Per-agent token management

Run milmil agents-guide for the recipe-oriented version. Every action that mutates server state is recorded in milmil's audit log with the calling token's name, timestamp, and a JSON snapshot.

Coming in v0.2

These commands are scaffolded but not yet wired:

subscribe namespace — subscribe add / list / undo will land with a server-side macro endpoint that records a parent audit entry linking the per-action children. For v0.1, use the web UI's plan-to- watch toggle plus the RSS rule editor.
match auto --confidence-floor X — accepted but no-op in v0.1 (warns to stderr). The server's existing matcher uses internal acceptance thresholds; the autonomous floor lands with the /match/auto macro endpoint in v0.2.
match suggest --file <id> — backing endpoint not yet defined.

Audit and undo

milmil audit list --since 1h
milmil audit list --action match.apply --since 7d
milmil audit show <audit_id>

# Surgical undo
milmil match undo --id <audit_id> --dry-run    # preview
milmil match undo --id <audit_id> --yes        # commit

# Bulk window undo (reverts every match.apply in the window — including
# manual ones from the web UI)
milmil match undo --since 5m --dry-run
milmil match undo --since 5m --yes

External-tracker reverters (Bangumi, AniList) report failed: dep not configured in v0.1. Wiring lands in v0.2 alongside the macro endpoints.

Supported skill formats

Format	Output	Conventional location
`claude`	Skill markdown with frontmatter	`~/.claude/skills/milmil/SKILL.md` (global) or per-project
`cursor`	`.cursorrules` body	project root
`agents-md`	`AGENTS.md` section	append to project's `AGENTS.md`

Adding a new format is a one-file change inside the CLI repo — api/cmd/cli/skill_templates/<format>.md is automatically picked up by milmil generate-skill --format <format>.

One token per agent

Best practice is to mint a separate token per agent and per device. Each token is independently revocable and the audit log records which token performed each action:

milmil token list                    # see all tokens
milmil token revoke claude-code-laptop --yes

milmil token revoke refuses to remove the token used by the current CLI session — switch credentials with milmil auth login first if you need to revoke that one.

AI Agent Integration

On this page