Agentworks CLI

The operator's command-line interface for managing agentic workloads on Agentworks.

For the project's problem space, core concepts, key principles, and tightly-integrated tool set, see the top-level README. This document covers installing the CLI, the command surface, configuration, and operational details.

Getting Started

Install from PyPI:

uv tool install agentworks-cli
# or:  pipx install agentworks-cli

The everyday command is agw. The longer form agentworks is also installed if you ever want to type it out; examples throughout this document use agw.

# Initial setup
agw config init                          # creates ~/.config/agentworks/config.toml
agw config edit                          # opens the config in your $EDITOR (or $VISUAL) to fill in required fields
agw doctor                               # sanity-checks installed tools, Tailscale, config validity, and the local DB

# Create a VM, workspace, agent, and session to see how the pieces fit together
agw vm create my-vm
agw workspace create my-workspace --vm my-vm
agw agent create my-agent --vm my-vm
agw session create my-session --workspace my-workspace --agent my-agent

# Attach to the session's tmux session to drive it
agw session attach my-session
# Use tmux's 'detach' command (default Ctrl-b unless overridden by config) to disconnect while
# leaving everything running on the VM.
agw session attach my-session    # You'll pick up right where you left off
agw session stop my-session      # Sessions can be stopped (or can exit on their own)
agw session list
agw session restart my-session
agw session attach my-session
agw session delete my-session    # When you're done with it. Agent and workspace are preserved.

# Alternatively, you can create ephemeral workspaces and agents along with your sessions
agw session create my-ephemeral-session --vm my-vm --new-workspace --new-agent
agw session attach my-ephemeral-session
agw session delete my-ephemeral-session    # This will prompt you to delete the associated workspace and agent, too

# Finally, create two sessions and a named console
agw session create s1 --vm my-vm --new-workspace --new-agent
agw session create s2 --vm my-vm --new-workspace --new-agent
agw console create my-console s1 s2+1      # The + syntax gives you extra shells as that agent
agw console attach my-console

# Deleting a session drops it from any console that referenced it (no dangling
# references are left behind). session delete lists the affected consoles, and
# for any console left with no sessions it offers to delete the now-empty
# console (interactively). Under --yes it reports the empty console but leaves
# it for you to remove with `agw console delete <name>`.
agw session delete s1                      # Reports that my-console still referenced s1

agw console delete my-console              # Extra shells are lost but sessions are preserved

Prerequisites

• Python 3.12+ (uv will install one for you if needed)
• uv or pipx for installation
• Tailscale installed and connected (for VM workspaces)
• One of: Lima, Azure CLI (az), Proxmox, or WSL2 (for VM provisioning)

Global Options

When --non-interactive is set (or stdin is not a TTY), commands that would normally prompt for missing values (VM selection, workspace selection, name generation) will fail with a clear error indicating which flag is required. VM auto-selection still works: if there is exactly one usable VM, it is used without prompting. session create is an intentional exception -- it always prompts for workspace and mode (even when only one choice exists) since those are part of the session's identity and should be an explicit operator decision.

Domain errors (SSH timeouts, validation failures, missing resources, etc.) surface as a single clean line: Error: <message>. Truly unexpected failures (internal bugs, OS-level errors, third-party library failures) also get a clean single-line message, plus the full traceback appended to ~/.config/agentworks/logs/error.log for debugging. Pass --debug (or set AGW_DEBUG=1) to print the traceback to stderr instead.

On an interactive terminal, output is tastefully colorized by role so it is easy to scan at a glance: a yellow Warning: prefix, a red Error: prefix, bold section headers, a dim-green result line (the closing "VM deleted", "rekeyed", etc.), and dimmed secondary detail. agw doctor colors its per-check status labels the same way (green [ok], yellow [warn], red [FAIL], unstyled [info]), plus its summary line's fail/warn/ok counts. Color is a presentation aid only, never carried in the message text. It is suppressed automatically when the target stream is not a terminal (pipes, redirects, CI capture) and under --non-interactive, so scripted and captured output stays byte-plain. Set the NO_COLOR environment variable (any value, honored by its presence) to opt out of color even on a terminal.

Pressing Ctrl-C during a long-running operation triggers best-effort cleanup. Where the operation can roll back (e.g. vm create during the provisioning phase, workspace create, agent create, session create) it undoes the partial DB / on-VM state and prints Cancelling X... rolling back.. Where rollback isn't possible (vm reinit, agent reinit, the init phase of vm create) it prints a recovery hint: the next command to run (vm reinit, vm delete --force, ...). Every cancellation exits with the conventional SIGINT exit code (130).

Commands

Top-Level

VMs

Manage virtual machines across declared vm-sites (Lima local or remote, Azure, WSL2, Proxmox).

Where VMs are created is declared as vm-site resources: YAML manifests under ~/.config/agentworks/resources/ that pair a platform (the code that runs VMs on one backend kind) with its configuration. The lima-local and wsl2 sites ship built in. Every site registers on every host and disables itself when this host lacks what it needs (wsl2 is Windows-only; a local Lima site needs limactl; a platform may simply not be installed): a disabled site still lists and describes, using it is an error naming the requirement, and agw doctor shows each platform's and site's state with the reason. Run agw resource sample vm-site for commented, ready-to-edit examples (an Azure site, a remote-Lima site with platform_config.vm_host). The former agw vm-host registry is gone: a remote Lima host is now just a vm-site.

Note on WSL2: WSL2 distros share the Windows workstation's lifecycle. They idle-shut after ~60s of no wsl.exe activity (vmIdleTimeout in .wslconfig) and do not survive workstation shutdown or sleep. agentworks holds a wsl.exe keepalive for the duration of each VM-touching command, so individual agw operations work cleanly, but agents and sessions on WSL2 are not suitable for unattended background workflows. Use a site on a different platform that provides true long-lived VMs (e.g. Lima, Azure, Proxmox, etc.) if you need a VM that survives independent of your workstation.

Power-state semantics: a VM that stopped on its own (idle timeout, host reboot) is started automatically, on demand, by any command that needs it live. A VM stopped with agw vm stop is different: that records your intent, so it stays down and commands that would need it refuse with a hint until you run agw vm start, which clears the intent. agw vm describe shows which case a stopped VM is in: its status reads stopped (manual) versus stopped (idle).

vm create <name> takes the VM name as a required positional. Optional flags: --template (a declared vm-template), --admin-template (a declared admin-template; defaults to the reserved default admin-template, which always exists), and --site (a declared vm-site; falls back to defaults.site, else the one ENABLED site is inferred when there is exactly one, several prompt interactively, and non-interactive runs error naming the options). The selected admin-template is stored on the VM row (NULL = default) and drives its admin user on every later vm reinit, vm shell, and admin-mode session. An unknown --admin-template name fails before any provisioning or DB work. Hardware (cpus, memory, disk, swap) comes from the vm-template and the admin username from the admin-template; there are no per-create overrides, so to deviate you declare a new template. On Azure, cpus + memory select the smallest fitting VM size from the site's catalog (built-in B-series, or platform_config.vm_sizes); an off-ratio request rounds up and warns. These are immutable provisioning parameters stored in the database. All initialization behavior (packages, install commands, etc.) is driven by config. Templates carry no site: placement is per-host, so it never travels inside a shared template.

The first interactive vm create asks once for an optional system slug (3-20 chars, lowercase alphanumeric plus dash, no leading/trailing dash): a short identifier for this agentworks installation, used to namespace VM hostnames and backend-side names ({slug}-{vm-name}) so installs sharing a cloud account, Proxmox cluster, or Windows/Mac user don't collide. Leave it blank if this install is the only one using its sites' backends; a blank answer is remembered and it will never ask again. Non-interactive runs never prompt (a later interactive create still asks once).

vm reinit re-runs the initialization phase using the current config without reprovisioning the VM. Changes to config (new packages, different install commands, etc.) are picked up automatically.

vm delete requires --force if the VM has workspaces, agents, or sessions. The confirmation message shows what will be deleted. Pass --yes to skip the prompt.

agw vm shell is the agentworks-wrapped entry point; for raw SSH (VS Code Remote-SSH, scp, etc.), use the awvm--<vm> alias documented under Direct SSH aliases.

vm shell and vm exec both accept --workspace <ws> to root the admin session in a workspace directory on this VM. The workspace's template env joins the env chain (between vm and admin), and AGENTWORKS_WORKSPACE / AGENTWORKS_WORKSPACE_DIR are set in the session. The shell variant cds into the workspace; the exec variant runs the command from the workspace directory. A workspace that lives on a different VM is rejected with a ValidationError before any SSH work.

In both exec commands the -- separator is only required when the remote command's first token starts with - (it stops agentworks from reading the token as its own option); without it, a dash-led first token is rejected with a hint naming the recoveries. Bare commands need no --.

Combining --workspace with --platform works (the shell still cds into the workspace) but the workspace's template env and the AGENTWORKS_WORKSPACE identity vars are not delivered: the platform-native transports (limactl shell, wsl.exe) drop the env= kwarg by design. Treat --platform as a transport-repair escape hatch, not a routine combination.

agw vm shell --platform (legacy alias --provisioner, one release) opens the same shell over the platform-native transport (limactl shell for Lima, wsl.exe for WSL2, SSH via a temporarily-attached public IP for Azure) instead of Tailscale. Useful when Tailscale itself is the thing you need to reach the VM to fix (the issue #117 latched DNS state is the canonical case: its heal involves restarting tailscaled, which would terminate a Tailscale-SSH session mid-sequence). On Azure, a public IP is attached for the duration of the session and detached on exit. Proxmox isn't supported by this flag because the QEMU guest agent's exec interface is one-shot and non-interactive; use the Proxmox web UI's serial console (VM > Console in the Proxmox VE web UI) as the equivalent escape hatch.

Workspaces

Manage workspaces on VMs.

workspace create <name> takes the workspace name as a required positional. Optional flags: --vm, --template, and --open-vscode.

workspace copy <source> <name> copies a workspace to a new VM workspace. Accepts --vm. Source and destination can be the same VM (a clone) or different VMs.

workspace delete requires --force if the workspace has sessions. Running sessions are killed during deletion. Pass --yes to skip the confirmation prompt.

There is deliberately no workspace shell: a shell rooted in a workspace is always somebody's shell. Use agw vm shell <vm> --workspace <ws> for an admin shell or agw agent shell <agent> --workspace <ws> for an agent shell. For curated tmux views over a workspace's sessions, use agw console create + agw console attach.

Agents

Manage agents (isolated Linux users) on VMs. Agents are VM-scoped and access workspaces via grants.

agent create <name> takes the agent name as a required positional. Optional flags: --vm, --template, and --grant-all-workspaces.

agent reinit --update-template <tmpl> re-points the agent to a different declared template (validated against the resource registry, then persisted) before re-running setup. An unknown template name is rejected up front, leaving the stored binding unchanged.

agent shell and agent exec both SSH directly as the agent's Linux user. agent shell opens an interactive login shell (sources the agent's profile). agent exec runs a single command non-interactively but still wraps it in the agent's login shell so the agent's PATH (mise shims, ~/.local/bin, etc.) is in scope. Useful for scripted invocations like agw agent exec myagent -- claude -p "...".

Both accept --workspace <ws> to root the session in a workspace the agent has access to. The workspace's template env joins the env chain (between vm and agent), and AGENTWORKS_WORKSPACE / AGENTWORKS_WORKSPACE_DIR are set in the session. The shell variant cds into the workspace; the exec variant runs the command from the workspace directory. A workspace on a different VM is rejected with a ValidationError; a workspace the agent lacks access to raises AuthorizationError with a hint to run agent grant-workspaces.

agent delete requires --force if the agent has running sessions. Pass --yes to skip the confirmation prompt.

agw agent shell / agw agent exec are agentworks-wrapped entry points; for raw SSH access to an agent (e.g. from VS Code Remote-SSH or scp), use the awagent--<agent> alias documented under Direct SSH aliases.

Direct SSH Aliases

Agentworks maintains operator-side SSH config entries for both VMs and agents under ~/.ssh/config.d/agentworks.conf (or inline in ~/.ssh/config if ssh_config_dir = false):

The agent alias is keyed on the agent's operator-facing name (the same name you use in agw agent ... commands), not on the on-VM Linux user (which is an implementation detail). The prefixes are configurable via operator.ssh_host_prefix (default awvm--) and operator.ssh_agent_host_prefix (default awagent--).

Entries are rebuilt declaratively from the database on every agent / VM lifecycle operation, so a fresh agw agent create or agw vm delete keeps the file in sync without manual intervention. Run agw config sync-ssh-config to force a rebuild.

Sessions

Manage sessions (persistent tmux sessions running in workspaces). Session names are globally unique -- no --workspace flag needed for most commands.

session list accepts --workspace, --vm, --agent, and --admin to narrow the result set. Filters compose with AND. The name filters (--workspace, --vm, --agent) accept a single value or a comma-separated list (--vm vm1,vm2); commas within a filter are OR-ed together. --agent <name> matches agent-mode sessions only; --admin matches admin-mode sessions only (the two are mutually exclusive).

session stop and session restart operate on a single session by default. Pass --all (session stop/session restart) or --all-stopped (session restart) to batch over the sessions on the VM. The batch form accepts --vm <vm>, --workspace <ws>, --agent <agent>, and --admin to narrow the set; filters compose with AND and require one of the batch flags. The name filters accept a single value or a comma-separated list (--vm vm1,vm2); commas within a filter are OR-ed together. --agent matches agent-mode sessions only; --admin matches admin-mode sessions only (the two are mutually exclusive). Pass --force to stop/restart broken sessions via PID kill.

session create <name> takes the session name as a required positional. Optional flags: --workspace, --template, --admin, and --agent. If --workspace / --new-workspace is omitted, you are prompted to pick from the existing workspaces or [Create new workspace] -- filtered to the known VM when --vm or --agent already pins one (the prompt prints Only showing workspaces on VM 'X' when a filter is active). If --admin / --agent / --new-agent is omitted, you are prompted with admin, the existing agents on the resolved VM, and [Create new agent]. The prompts always fire when the flags are absent -- there is no single-option auto-select for workspace or mode, since both are part of the session's identity. --vm works differently: it auto-selects when exactly one usable VM exists (logged as Using VM 'X'), prompts when multiple, and is required only in non-interactive mode when no workspace or agent anchor pins the VM. In non-interactive mode (--non-interactive or no TTY), any required prompt raises a ValidationError directing you to pass the corresponding flag. Pass --new-workspace to create a workspace on the fly (with optional --workspace-name, --workspace-template, and --vm; --workspace-name defaults to the session name). Pass --new-agent to create a new agent for the session (with optional --agent-name and --agent-template; --agent-name defaults to the session name); the new agent is provisioned on the workspace's VM. When a session created with --new-workspace or --new-agent is later deleted, you are offered the option to delete the workspace and/or agent as well -- the workspace if no other sessions remain on it, the agent if it has no other sessions and no explicit grants.

Named Consoles

Named consoles are persistent, curated tmux views over sessions on a VM. Each console is its own tmux session (aw-console-<name>) containing one window per included session, plus any extra shell panes you want preloaded into a session's window.

console create accepts:

• --vm -- target VM. Inferred from the listed sessions when omitted; if the listed sessions span more than one VM, console create errors and asks you to pick one with --vm. When no sessions are listed (e.g. with --all and no explicit specs), VM selection falls back to the standard prompt (auto-picked if you have a single VM, prompted otherwise).
• --all -- include every session on the VM with 0 shells, appended after the explicit specs (alphabetical).
• --all-running -- like --all but restricted to sessions whose live tmux state on the VM is OK (one SSH round-trip; same probe agw session list uses). Mutually exclusive with --all. Requires the VM to be reachable.
• --add-admin-shell -- include a top-level admin-shell window as window 0, matching the legacy vm console behavior.

console list accepts --vm, --workspace, and --agent to narrow the result set. Each filter takes a single value or a comma-separated list (--workspace ws1,ws2); commas within a filter are OR-ed together. The --workspace and --agent filters use "any session matches" semantics: a console is listed if at least one of its member sessions belongs to the given workspace / runs as the given agent. When --workspace and --agent are both passed, the SAME session must satisfy both predicates. The session count displayed is the total membership, not the count of matching sessions. Filters compose with AND.

Session specs use name or name+N shorthand, where N is the number of default shell panes to pre-open in that session's window (running as the session's agent user, cwd = workspace root):

# A console with three sessions; the first two get extra shells.
# VM is inferred from the sessions.
agw console create backend auth-server+2 auth-tests+1 docs

# Same, but also include a top-level admin-shell window (window 0).
agw console create backend auth-server+2 auth-tests+1 docs --add-admin-shell

# Everything currently running on the VM, after the explicit specs.
agw console create live auth-server+2 --all-running

# All sessions on the VM (running or not). Needs --vm since no sessions are
# named explicitly to infer from.
agw console create everything --vm aw-private --all

# Add an admin shell rooted in a sub-path of the workspace.
agw console add-shell backend auth-server --cwd src/api --admin

console restore-session repairs a single session window in a running console: it re-adds shell panes you killed by accident (each back in its configured position) and rebuilds the window from config if it is gone entirely. It is additive and never kills a live pane or window, so it refuses, pointing you at console attach --recreate, when the fix would require destroying live state: more panes live than configured, shell panes it can't map back to the config (untagged, duplicated, or out of range), or a window whose session pane itself was killed (the console then shows a plain shell where the session should be).

Memberships and shell layouts persist in the database. agw console attach builds the tmux session on first attach (or with --recreate); subsequent attaches reuse the running tmux session. Adding or removing sessions/shells while a console is attached updates the live tmux state immediately (best-effort); when the console isn't running on the VM, only the DB is updated and changes appear on next attach.

tmux Architecture

Each session runs in its own locked-down tmux session on the VM. There are several ways to interact with sessions, at different scopes:

Session tmux Sessions

Each session gets a locked-down tmux session using the session name directly as the tmux session name. The user's ~/.tmux.conf (customizable via dotfiles) is loaded first so that familiar keybindings (prefix, detach, copy mode, scroll) work for direct session attach. Window/pane creation, session management, and the command prompt are selectively unbound.

Agent-mode sessions each get their own tmux socket, so every session runs as its own tmux server rather than as a window in a shared one. The sockets are grouped in a per-agent directory whose ownership keeps one agent from reaching another's (cross-agent isolation), and giving each session its own server means it inherits the environment delivered over its own SSH connection instead of leaking env across sessions through a shared server. The agent's shell attaches directly to the tmux pane PTY, and each socket path is persisted in the database.

Named Console

console attach <name> creates or attaches to the aw-console-<name> tmux session. Each member session becomes a window running the same wrapper used by the VM console, plus a configurable number of extra shell panes (default user = session's agent user, default cwd = workspace root; override per pane with --cwd / --admin on console add-shell).

aw-console-backend
  Window 1: auth-server                attached session + 2 agent shells (workspace root)
  Window 2: auth-tests                 attached session + 1 agent shell
  Window 3: docs                       attached session only

Membership and per-session shell layout are stored in the database; see Named Consoles above for the lazy-build and live-update semantics. Of note here is the VM-boot behavior: the mutation commands (add-sessions, remove-sessions, reorder-sessions, add-shell) never auto-boot the VM, whereas the explicit attach/repair commands (attach, restore-session) do start a stopped VM, since their job is to bring live state up.

Workspace tmuxinator Config

Workspaces with tmuxinator enabled in their template (the default) carry a tmuxinator config (.tmuxinator.yml in the workspace root, symlinked as ~/.config/tmuxinator/ws-<name>-console.yml) describing a ws-<name>-console session: an admin-shell window plus one window per session. It is regenerated whenever sessions change. The agw workspace console command that attached to it was removed (superseded by named consoles); the config remains usable directly on the VM via tmuxinator start ws-<name>-console (e.g. inside VS Code's integrated terminal).

VM Console (deprecated)

vm console creates or attaches to the vm-console session, which spans all sessions on the VM. Built dynamically (not via tmuxinator). Superseded by named consoles, which let you curate which sessions are in scope at any moment instead of seeing every session on the VM. Will be removed in a future release.

Shells

vm shell and agent shell open plain login shells with no tmux (optionally rooted in a workspace via --workspace <ws>). Use these when you just need a terminal without the console structure.

Key Behaviors

• Direct attach (session attach): the user's prefix key, detach, copy mode, and scroll all work normally. Status bar is hidden since there is only one pane.
• Consoles (console, vm console): the console's prefix key eclipses the inner session's prefix, so window switching, detach, etc. all operate at the console level. Session windows use a wrapper that re-attaches if the inner session disconnects and shows a message when the session ends.
• Nesting protection: the console commands refuse to run inside an existing tmux session to avoid prefix key conflicts. Pass --allow-nesting to override.
• Console lifecycle: consoles are independent of sessions. Killing or detaching a console does not affect running sessions. --recreate rebuilds from scratch.

Session Templates

A session template selects the harness that runs the session's workload. The harness is a capability that owns starting/restarting the tool and checking its required executables; the template names one with spec.harness and hands it a spec.harness_config block it validates. A template that names no harness runs the built-in shell harness (a login shell, $SHELL --login, or an operator-supplied command), which is the built-in default template's behavior. Define custom templates as session-template resources:

apiVersion: agentworks/v1
kind: session-template
metadata:
  name: htop
  description: Live process monitor
spec:
  harness: shell
  harness_config:
    command: htop
    required_commands: [htop]

shell's harness_config vocabulary is the command surface every template used to spell at the spec top level:

• command: the pane command (empty/omitted is a plain login shell). Supports {{session_name}} and {{workspace_name}} variable substitution (double-brace syntax).
• restart_command: used by session restart, for a tool that needs a different invocation on restart. If omitted, command is used. (To run Claude Code, prefer the dedicated claude-code harness below, which resumes the previous conversation on its own.)
• required_commands: executables the command needs, checked on the session's launch target (the agent, or the VM admin for admin sessions) before any state mutation, so launching a session whose tool is not installed fails fast with a clear error instead of a cryptic downstream tmux failure. Merged (de-duped, order-preserving) across template inheritance.

In a YAML manifest these three keys live only under harness_config; spelling any of them at the spec top level is a load error that points you at the nested shape. That check is one instance of a general deprecated-field notice: any resource kind can flag retired or relocated spec fields with an actionable message (a hard load error when ignoring the field would change behavior, otherwise a warning that agw doctor also surfaces). It is separate from the TOML flat-field handling below, which is a permanent supported spelling, not a deprecation.

The claude-code harness runs Claude Code as the session: session create starts a new Claude session and session restart resumes the same conversation when its transcript still exists on disk (launching fresh when Claude never wrote one). It needs only that claude is installed on the launch target, and announces the chosen action (resume vs new session) in the pane, so the decision is never silent. Its harness_config vocabulary is three optional fields:

• permission_mode: forwarded verbatim to claude --permission-mode (its choice set is Claude's, not validated here).
• model: forwarded verbatim to claude --model.
• extra_args: a list of raw argv tokens appended last, the escape hatch for any flag the harness does not model. Each element is one argv token (shell-quoted, never re-split), and elements support the {{session_name}} / {{workspace_name}} variables.

apiVersion: agentworks/v1
kind: session-template
metadata:
  name: claude
  description: Claude Code session
spec:
  harness: claude-code
  harness_config:
    permission_mode: acceptEdits
    model: opus

The (harness, harness_config) pair inherits as a unit: a child that restates the same harness merges its config block into the parent's (child wins per key; shell unions required_commands), while a child naming a different harness starts from a fresh config (the parent's block was addressed to a different tool). env, inherits, and the description merge as usual.

TOML ([session_templates.<name>] in config.toml, deprecated but supported): the same harness string and a nested harness_config table are accepted:

[session_templates.htop]
harness = "shell"
[session_templates.htop.harness_config]
command = "htop"
required_commands = ["htop"]

For shell, the legacy flat keys command / restart_command / required_commands keep working at the section top level and are hoisted into harness = "shell" + the equivalent harness_config. The flat form is the documented default TOML shape; YAML manifests are the primary authoring surface (run agw resource sample session-template). The flat fields cannot be combined with a non-shell harness or with an explicit harness_config table (one spelling per declaration), and both conflicts are load errors. One inheritance interaction worth noting: a legacy flat-field child under a harness: claude-code parent hoists to harness = "shell", which (per the different-harness rule) switches the lineage back to shell with a fresh config.

Config

Resource Registry

Cross-kind inspection of the Resource Registry. The registry is the framework that owns every operator-declared, auto-declared, and built-in resource the CLI knows about: secrets, VM templates, agent templates, workspace templates, apt / install-command entries, git credential providers, secret backends, etc. The two commands below stop at the framework-uniform fields (kind, name, origin, references, used_by, description). For kind-specific detail (secret backend mappings, template inheritance chains, resolution previews), reach for the per-kind command (e.g. agw secret describe).

resource list accepts --kind <csv> (e.g. --kind secret,vm-template) and --origin <variant> where variant is operator, auto, or builtin. --names-only emits kind/name per line and backs shell completion (/ cannot appear in resource names, so the split is unambiguous). The kind/name token is the one grammar across the resource group: resource describe secret/npm-token and resource migrate vm-template/dev take the same shape.

resource migrate is a recurring, incremental mover -- run it any time you want to move resources (or a subset) from TOML to YAML manifests. Selectors scope the run: KIND one kind, KIND/NAME one resource (overlaps union), or --all for everything TOML-declared -- a bare invocation errors rather than migrating the whole config by accident. --layout per-kind|single|per-resource picks the file mapping (default one multi-document file per kind, e.g. resources/vm-templates.yaml). Output is append-only: existing YAML files are never parsed or rewritten, new documents are appended after a --- separator. Because a resource declared in both sources is a hard load error, the migrated TOML sections are commented out in place with a # migrated to resources/<file> marker (default) or removed with --toml delete; either way the original config.toml is backed up to paths.backups first and the rewrite is atomic. Deprecated [secret_backends.*] sections are dropped (with a note) by any run, including a bare run with nothing else to migrate. Every real run finishes by rebuilding the registry and verifying it is row-for-row identical to the pre-migration one -- on mismatch the run rolls back and reports. --dry-run prints a summary of what would migrate where and writes nothing; add --full for the complete YAML documents and the config.toml diff.

resource sample prints a kind's fully-commented-out sample manifest (--all for every kind) -- the YAML teaching surface, mirroring agw config sample for the settings file. --write <file> saves under the resources directory instead (relative .yaml/.yml path; appends if the file exists). Written samples are inert until you uncomment them (delete one leading # per line), so --write can never create a live resource or a duplicate.

Configuration

Configuration splits into two surfaces:

• Settings live in ~/.config/agentworks/config.toml: your identity, paths, defaults, and the secret backend chain. Run agw config init to generate a sample; see sample-config.toml for the full reference.
• Resources (secrets, templates, git credentials, vm-sites, apt / install-command entries) are declared as YAML manifests under ~/.config/agentworks/resources/, auto-loaded whenever a command needs them. agw resource sample <kind> prints a commented starter (--all for every kind). The classic TOML resource sections keep working (deprecated, with one aggregated load warning naming the sections present; silence it with the global --no-deprecations flag); agw resource migrate moves them to YAML whenever you like. See docs/guides/resources.md.

Settings sections (config.toml, permanent):

• [operator] -- SSH keys (required), additional authorized keys, SSH config management
• [paths] -- VM workspace, VS Code workspace file, and backup directories
• [defaults]: site, the default vm-site for vm create (platform is the deprecated alias)
• [session.config] -- session defaults (history limit)
• [secret_config] -- active secret backend chain ([secret_backends.*] sections are deprecated no-ops; see Secret Backends below)

Resource kinds (YAML manifests; the deprecated TOML section is noted for each):

• vm-site ([azure] / [proxmox], flat legacy shape): a configured place to create VMs. spec.platform names the backing platform, spec.platform_config carries its settings (Azure subscription/resource-group/region, Proxmox API endpoint + token secret, remote-Lima vm_host). The lima-local and wsl2 sites ship built in (on hosts where their platform can run) and their names are reserved
• vm-platform: read-only capability rows for the in-tree platforms (lima, wsl2, azure-vm, proxmox); listed by agw resource kinds, never declared
• vm-template ([vm_templates.*]): VM resources, apt packages, system install commands, mise, and the target site
• admin-template ([admin.config]) -- admin user shell, dotfiles, git credentials, user install commands, mise
• agent-template ([agent_templates.*]) -- agent user shell, dotfiles, git credentials, user install commands, mise
• session-template ([session_templates.*]) -- session commands with variable substitution
• workspace-template ([workspace_templates.*]): workspace repo, tmuxinator, optional git identity (git_user_name / git_user_email, stamped into the cloned repo), inheritance
• named-console-template ([named_console]) -- named-console layout (tmux preset names + aw-session-vertical)
• git-credential ([git_credentials.*]) -- git credentials; spec.provider selects github or azdo (TOML also accepts the legacy type)
• secret ([secrets.*]) -- secret declarations referenced by {secret: name} env entries
• apt-source / apt-package / system-install-command / user-install-command ([apt_sources.*] etc.): apt / install-command extensions
• Env vars ride their owning resource: an env map in the template's spec (TOML: [<scope>.env] subsections) at vm / workspace / admin / agent / session scope

Environment Variables and Secrets

Env tables can be declared at five scopes; for any given session the merged value is computed in this precedence order (highest scope wins; identity vars win over everything):

session > (agent | admin) > workspace > vm           (AGENTWORKS_* identity overrides all)

Admin and agent scopes are mutually exclusive: a shell opened as the admin user (e.g. agw vm shell) sees admin scope; an agent-mode session sees agent scope. Each scope is an env map on the owning resource, mapping env-var name to either a plaintext string or a secret reference:

apiVersion: agentworks/v1
kind: vm-template
metadata:
  name: default
spec:
  env:
    HTTP_PROXY: http://proxy:3128
    NPM_TOKEN: { secret: npm-token }

(TOML equivalent: [vm_templates.default.env] with NPM_TOKEN = { secret = "npm-token" }.)

Every secret reference points to a secret resource declaration (auto-declared with a framework-synthesized description if you skip it). Active backends (and their precedence order) are listed in [secret_config].backends. Today the implemented backends are:

• env-var -- reads from the operator's process env. Default convention is AW_SECRET_<UPPER_SNAKE_CASE>, overridable per secret via the secret's backend_mappings (env-var: CUSTOM_NAME).
• prompt -- interactive prompt; you are never asked for the same secret twice in one command, and all prompting happens before the command starts changing anything.

Resolve before any mutation: a command resolves all the secrets its plan needs up front, before it starts changing anything. A secret that cannot be resolved by any active backend fails at preflight with a hint (agw secret describe <name> shows how each backend looks it up), before any prompt and before any VM is started. The set of secrets is computed from the command's static filters (positional targets, --vm, --workspace, --agent, etc.) -- dynamic predicates like --all-stopped apply later, so the prompted set may over-approximate. Non-interactive mode (no TTY or --non-interactive) surfaces missing secrets as SecretUnavailableError with a per-secret hint naming which backends were tried. Commands that join existing shells (session attach, session list, console attach against a live tmux session, console add-sessions) consume no secrets.

Miss semantics: what "not found" means depends on the backend. Conventional sources (env-var, prompt) treat a missing value as a soft miss and fall through to the next backend in the chain -- a GITHUB_TOKEN env var that isn't set is just-not-set, not a config error. Persistent-store backends (1Password, Vault when implemented) will treat an explicit mapping that doesn't resolve as a hard miss: they raise SecretMappingError and the chain halts so a wrong op:// URI doesn't quietly fall through to a prompt that masks the real problem.

Inspect the merged result for any context with agw env show:

agw env show --session my-session              # secrets redacted as <from secret: name>
agw env show --vm my-vm --reveal-secrets       # resolves through the active backend chain

Inspect how each active backend would resolve each declared or auto-declared secret (e.g. "which env var name does this secret read from?") with agw secret list:

agw secret list
# 4 secrets (2 operator-declared, 2 auto-declared)
#
# NAME                 DESCRIPTION                                                                env-var                       prompt
# ----                 -----------                                                                -------                       ------
# api-key              OpenAI key for the operator's service                                      OPENAI_API_KEY                enabled
# force-prompt         Always prompted at command time                                            disabled                      enabled
# git-token-github     (auto) the auth token for git_credentials:github                           AW_SECRET_GIT_TOKEN_GITHUB    enabled
# tailscale-auth-key   (auto) the Tailscale auth key for vm-template:default (and 1 more)   AW_SECRET_TAILSCALE_AUTH_KEY  enabled

Columns are the active backends in [secret_config].backends precedence order. Cells show each backend's static lookup identifier (env var name, vault path, op:// URI) or disabled / enabled for backends with an explicit opt-out or no static identifier (prompt). The Description column shows the operator-supplied text for operator-declared secrets, or a framework-synthesized (auto) <usage> for <kind>:<name> (plus (and N more) when more than one source requires the secret) for auto-declared ones. The synthesized text reads as "what this secret is for, and who's asking." The summary line breaks the rows down by origin. Values are never resolved.

For the full per-secret detail view, including the structured origin block, usage list (who requires this secret), per-backend mapping table, and a resolution preview, use agw secret describe:

agw secret describe tailscale-auth-key
# Secret: tailscale-auth-key
#   Kind: secret
#   Description: (auto) the Tailscale auth key for vm-template:default (and 1 more)
#   Origin: auto-declared (vm-template:default)
#
# Referenced by:
#   - vm-template:default -- the Tailscale auth key
#   - vm-template:heavy -- the Tailscale auth key
#
# Backend mappings:
#   - env-var: AW_SECRET_TAILSCALE_AUTH_KEY
#   - prompt: (prompt at resolution time)
#
# Resolution preview:
#   would resolve via env-var

describe reports state -- it does not prompt and does not resolve the secret's value.

agw doctor's Secrets group emits exactly one row per registry secret -- operator-declared and auto-declared alike (auto-declared rows, e.g. tailscale-auth-key and the git-token-* family, carry an (auto) marker; they are exactly the secrets most likely to prompt at command time):

• OK when at least one active backend would resolve the secret at runtime (would resolve via env-var, would resolve via prompt, ...). would resolve via prompt is the heads-up that the next command needing this secret will ask for it interactively.
• WARN when nothing in the chain would resolve it (config-valid but no path to a value, e.g. env-var has no matching env var set and prompt is opted out via backend_mappings.prompt = false).
• FAIL when backend_mappings references an unknown backend name (not a registered backend like env-var / prompt).

Backend-applicability detail (per-backend soft-skip reasons, inactive mappings, per-secret references) lives in agw secret list and agw secret describe. AGENTWORKS_* identity overrides surface in the Configuration group (they're a config-load warning). Broken { secret = "..." } references are caught earlier as a hard config-load error before doctor runs. Git-credential tokens are just secrets: their resolvability reports as ordinary git-token-<name> rows in the Secrets group, like any other secret. Doctor is preflight-only and never prompts, so it does not authenticate tokens against their provider; live verification (a token expired, revoked, or wrong-scope) happens at the capability runup() stage inside provisioning ops, and on-demand via the planned agw doctor --runup (which may prompt). The Tailscale group checks only workstation connectivity; the auth key is the tailscale-auth-key secret row.

When the config or a resource manifest fails to load, the groups that depend on them (VM sites, Secrets) do not vanish: each renders a single [info] ... skipped (config or manifests unavailable; see the Configuration group) row, so a degraded run keeps the same section skeleton as a healthy one and the Configuration group carries the actual failure.

Secret Backends

A backend is a capability resource that produces secret values (env-var, prompt; future backends like onepassword): a read-only row backed by registered code, listed by agw resource list --kind secret-backend and activated in precedence order by the chain ([secret_config].backends). Per-secret behavior -- identifier overrides, structured store addressing, opt-outs -- lives in each secret's backend_mappings.<backend>.

Mise (Polyglot Tool Manager)

Agentworks installs mise by default on all VMs for managing CLI tools (terraform, adr-tools, node, etc.) with optional lockfile-based integrity verification. See Using mise for the full guide.

Claude Code Plugins

Agentworks can register Claude Code marketplaces and install plugins automatically per user (admin and per-agent). Configure via claude_marketplaces and claude_plugins on the admin template or any agent template. Requires the claude CLI on PATH (typically installed via user_install_commands). To install nerftools this way:

apiVersion: agentworks/v1
kind: admin-template
metadata:
  name: default
spec:
  claude_marketplaces: ["https://github.com/WayfarerLabs/nerftools#4.1.0"]
  claude_plugins: [nerftools-default@nerftools]

(TOML equivalent: [admin.config] in config.toml, deprecated but supported.)

Built-in Apt / Install-Command Entries

Agentworks ships built-in entries for common tools (apt sources, apt packages, system install commands, and user install commands), bundled as YAML manifests under agentworks/manifests/builtin/. Run agw resource list --kind apt-package,system-install-command,user-install-command,apt-source to see what is available (or filter to any single kind). Reference these entries by name from VM, admin, and agent templates. User-defined entries override built-in entries with the same name.

VM Initialization

VM creation follows a two-phase lifecycle tracked by separate status columns:

1. Provisioning (provisioning_status) -- one-time, platform-specific, over the provisioning transport (Lima shell, SSH, or WSL2 exec): create user, install system packages, add SSH key, install and join Tailscale

2. Initialization (init_status) -- repeatable via vm reinit, over Tailscale SSH: configure apt sources, install apt packages, install snap packages, install mise, set shell, reconcile SSH authorized keys, run system install commands, write mise config, configure PATH, configure git credentials, sync dotfiles, fetch mise lockfile, run mise install, run user install commands for the admin user

Initialization is fully declarative, driven entirely by config. vm create only accepts a name, --template, --admin-template, and --site; the immutable provisioning parameters (resources, admin username) come from the selected templates. vm reinit takes only the VM name and re-runs initialization using the current config.

Non-fatal initialization failures (packages, dotfiles) produce a partial status rather than aborting. Fatal failures prompt for deletion or reinit. Use vm describe to view the full event log.

Shell Completion

agw completion install

The shell is autodetected from $SHELL; pass --shell {bash|zsh|powershell} to override (or when autodetection isn't unambiguous, e.g. on Windows). completion install writes the script to the standard location for that shell. For PowerShell it also appends a dot-source line (. "...") to $PROFILE. For bash and zsh, if your rc file is missing the loader (bash-completion for bash, fpath=(~/.zfunc $fpath) for plain zsh without a plugin manager), the installer prints a one-line note telling you what to add.

To print the script without installing, use agw completion show (handy for piping into your own config-management flow). To remove completions installed here, use agw completion uninstall --shell {bash|zsh|powershell}. For PowerShell, uninstall also strips the dot-source line the installer appended to $PROFILE; user-authored lines around it are left untouched.

Completions include dynamic VM, vm-site, workspace, session, and template name lookups.

State

All state is stored in ~/.config/agentworks/agentworks.db (SQLite). Schema migrations are forward-only and run automatically.

Environment Variables

Secret values are read from the operator's shell via the env-var backend, which follows the convention AW_SECRET_<UPPER_SNAKE_CASE> derived from the secret's name. The Tailscale auth key (secret tailscale-auth-key) reads from AW_SECRET_TAILSCALE_AUTH_KEY; a git credential's PAT (secret git-token-<name>) reads from AW_SECRET_GIT_TOKEN_<NAME>; and so on. Override the convention per secret via the secret's backend_mappings (env-var: CUSTOM_NAME).

Use agw secret list to see the exact env var name for each declared or auto-declared secret, and agw secret describe <name> for the full per-secret view (origin, usages, backend mappings, resolution preview).