fleet-config.yaml reference

fleet launch reads a single YAML file describing a whole fleet — an optional bridge, an optional gui, and an optional map of ships — and starts everything in one process. The file defaults to ./fleet-config.yaml; override it with fleet launch --config-path <path>.

Every top-level section is optional. Only the sections present are started: fleet launch on an empty file ({}) starts nothing and exits.

Scaffold

fleet launch init writes this file verbatim. It is a valid config as written.

# fleet-config.yaml — configuration for `fleet launch`.
# Every section is optional; only the sections present are started.

# The fleet-wide bridge that coordinates ships and serves the fleet API.
bridge:
  dataDirectory: ./.fleet/bridge
  port: 4800
  name: my-fleet-bridge
  # publicUrl: http://this-host:4800  # how ships reach this bridge; required if any ship is on another host
  # insecureNoAuth: true  # DEV ONLY: serve every route unauthenticated and skip creating the first admin

# The web gui. Proxies to the bridge above by default.
gui:
  port: 3000
  # bridgeUrl: http://localhost:4800  # defaults to the local bridge

# Ships that host workspaces. Each key is the ship's default name.
# A ship's shipToken/bridgeToken are set together or not at all, and ${VAR} reads
# the value from the environment so the secret never has to live in this file.
ships:
  ship-a:
    # source: local (the default) spawns the ship in this process.
    source: local
    fleetDirectory: ./fleet/ship-a
    port: 4700
    # name: ship-a  # defaults to the key above
    # shipToken: ${SHIP_A_SHIP_TOKEN}      # pin this ship's credentials instead of generating a fresh pair
    # bridgeToken: ${SHIP_A_BRIDGE_TOKEN}  # launch fails if a referenced variable is unset

  # source: remote registers an already-running ship by URL instead of spawning it.
  # ship-b:
  #   source: remote
  #   url: http://another-host:4700
  #   shipToken: ${SHIP_B_SHIP_TOKEN}      # the credentials that ship was already started with
  #   bridgeToken: ${SHIP_B_BRIDGE_TOKEN}  # must equal the ship's own FLEET_BRIDGE_TOKEN

Top level

Key	Type	Required	Default	Meaning
bridge	object	no	absent	Start a bridge in this process.
gui	object	no	absent	Serve the web GUI in this process.
ships	map of string → object	no	{}	Ships to start and/or register.

bridge: and gui: written with no body parse to null in YAML; both are treated as “enabled with defaults” rather than an error. ships: with no body is not given this treatment.

bridge

Every field has a default, so bridge: {} is valid.

Field	Type	Required	Default	Meaning
dataDirectory	string (non-empty)	no	./.fleet/bridge	Where the bridge persists ships.json, repos.json and ephemeral.json, and where its armory/ directory lives. Resolved to an absolute path.
port	integer	no	4800	Port the bridge’s HTTP + WebSocket API listens on.
name	string (non-empty)	no	bridge	Human-facing name of the bridge.
publicUrl	string (non-empty)	no	http://localhost:<port>	URL ships use to reach this bridge.
insecureNoAuth	boolean	no	absent (authentication required)	Development only: serve every route unauthenticated, and skip creating the first admin.
sweepIntervalMs	integer ≥ 0	no	300000 (5 minutes)	How often to check ephemeral workspaces for a closed pull request. 0 turns the sweep off, leaving POST /workspaces/sweep as the only way to run one.

Note

The dataDirectory default here (./.fleet/bridge) is not the same as the fleet bridge CLI default (./.fleet-bridge). They are separate defaults in separate code paths.

publicUrl

publicUrl is handed to each ship so it can pull the armory, so it has to resolve from the ships’ hosts, not from the machine running the launch. The default, http://localhost:<bridge.port>, is correct for a single-host fleet and wrong the moment a source: remote ship is on another machine — there, localhost is that machine.

Getting it wrong fails quietly: the ship registers, its workspaces work, and only the armory never arrives. So fleet launch warns when a config declares one or more source: remote ships and sets no publicUrl:

fleet launch: bridge.publicUrl is not set, so remote ships "build-box", "gpu-box" will be told this bridge is at http://localhost:4800, which on their hosts is themselves; set bridge.publicUrl to a URL those hosts can reach

It is a warning on stderr, not an error — a source: remote ship can legitimately be on this same host behind a tunnel or a published container port, where localhost still resolves. Local ships never trigger it.

The value is used verbatim; it is validated as a non-empty string, not parsed or normalized like gui.bridgeUrl, so write a full URL with its scheme. The equivalent flag on a standalone bridge is fleet bridge --public-url.

It is also what every source: local ship is pinned to (fleet ship --bridge-url), so those ships refuse an armory push from anywhere else. A value that is not an http(s) URL cannot be a pin; rather than fail the launch, it warns and starts the ships unpinned:

fleet launch: bridge.publicUrl "bridge:4800" is not an http(s) URL, so ships are started unpinned and will accept the first armory push they receive

Ships registered with source: remote are pinned by whatever they were started with — fleet launch does not configure a ship it did not spawn.

insecureNoAuth

A bridge with no users creates the first admin before it serves anything: from FLEET_BRIDGE_ADMIN_USER, FLEET_BRIDGE_ADMIN_EMAIL and FLEET_BRIDGE_ADMIN_PASSWORD if all three are set, otherwise by prompting on stdin. On a headless box with none of them set there is no terminal to answer the prompt, and the launch fails:

fleet launch: fleet-bridge has no users and stdin is not a terminal — set FLEET_BRIDGE_ADMIN_USER, FLEET_BRIDGE_ADMIN_EMAIL, FLEET_BRIDGE_ADMIN_PASSWORD to create the first admin, or start with --insecure-no-auth

insecureNoAuth: true is how a fleet launch config says “this is a dev fleet”: it skips that bootstrap and serves every route as an admin, with a banner on stderr for as long as the bridge runs. It gives up authentication and authorization — see authentication for exactly what.

The FLEET_INSECURE_NO_AUTH=1 environment variable is not consulted by fleet launch; it only affects a standalone fleet bridge. Under fleet launch the setting has to be this key.

gui

Both fields are optional, so gui: {} is valid — as long as a bridge exists to proxy to.

Field	Type	Required	Default	Meaning
port	integer	no	none — Bun picks the port	Port the GUI listens on.
bridgeUrl	string (non-empty)	no	http://localhost:<bridge.port>	Bridge the GUI reverse-proxies /bridge/* to.

When bridgeUrl is omitted, the GUI targets the launched local bridge. That is why a gui section with neither a bridge section nor a bridgeUrl is a validation error — there would be nothing to point it at.

The value is normalized like every other Fleet URL: a bare port becomes http://localhost:<port>, a bare host:port gains an http:// scheme, and a full URL is used as-is.

ships

A map, not a list. Each key names a ship and supplies the defaults for that entry: the key becomes the ship’s name and, for local ships, its fleetDirectory (./fleet/<key>).

Each value is one of two shapes, discriminated by source. When source is omitted the entry is treated as local, so ship-a: {} is a complete, valid ship.

source: local

The ship is spawned inside the fleet launch process.

Field	Type	Required	Default	Meaning
source	"local"	no	local	Discriminator.
fleetDirectory	string (non-empty)	no	./fleet/<key>	Directory holding this ship’s workspaces (<dir>/<repo>/<name>). Resolved to an absolute path.
port	integer	no	4700	Port this ship listens on.
name	fleet identifier	no	the map key	Human-facing name of this ship.
shipToken	string (non-empty)	no	freshly minted	The token this ship presents to the bridge. See ship credentials.
bridgeToken	string (non-empty)	no	freshly minted	The token the bridge presents to this ship.

Because port defaults to 4700 for every local ship, two or more local ships must each set a distinct port.

source: remote

The ship is already running elsewhere; fleet launch only registers it with the bridge.

Field	Type	Required	Default	Meaning
source	"remote"	yes	—	Discriminator. Must be written explicitly; omitting it makes the entry local.
url	string (non-empty)	yes	—	Base URL of the running ship.
shipToken	string (non-empty)	no	none	The token this ship presents to the bridge. See ship credentials.
bridgeToken	string (non-empty)	no	none	The token the bridge presents to this ship — the ship’s own FLEET_BRIDGE_TOKEN.

Remote entries accept no other fields — no port, no name, no fleetDirectory. The ship’s name is discovered by the bridge from the ship’s own sync event, not taken from the map key.

Ship credentials

shipToken and bridgeToken are set together or not at all, on either kind of ship. One without the other fails the launch before anything starts:

fleet launch: ship "gpu-box" sets shipToken but not bridgeToken; a ship is registered with both a shipToken and a bridgeToken, or neither

What each token does, and which end holds it, is covered in authentication. What matters here is the default:

• A source: local ship with neither key set gets a freshly minted pair — fleet launch generates it, hands it to the ship it spawns, and registers the ship with it. That is the usual case. A pair is only minted for a ship the launch is about to register: one the bridge’s roster already holds keeps the credentials that roster entry was registered with, and is spawned without a pair of its own.
• A source: local ship with both keys set uses those instead of minting.
• A source: remote ship with neither key set is registered with no credentials, and the bridge talks to it unauthenticated.
• A source: remote ship with both keys set is registered with them. The ship itself must already have been started with FLEET_BRIDGE_TOKEN set to the same bridgeToken, or the bridge’s very first call to it is rejected and the registration fails.

${VAR} interpolation

Both fields accept ${VAR} in place of a literal, so the secret never has to be written into the config file:

ships:
  gpu-box:
    source: remote
    url: http://10.0.0.7:4700
    shipToken: ${GPU_BOX_SHIP_TOKEN}
    bridgeToken: ${GPU_BOX_BRIDGE_TOKEN}

The rules are deliberately narrow:

Value	Result
${NAME} — the whole value, optionally with surrounding whitespace	Replaced by the environment variable NAME, trimmed.
any string with no ${ in it	Used literally.
anything else containing ${ — tok-${NAME}, ${A}${B}	Rejected.
${NAME} where NAME is unset, empty, or only whitespace	Rejected.

A variable name must match [A-Za-z_][A-Za-z0-9_]*. Interpolation applies only to these two fields — no other key in the file is expanded.

An unset variable is an error rather than an empty value, and this is the point of the feature: the alternative is a fleet that comes up looking healthy with a ship registered unauthenticated, which nothing would tell you about.

fleet launch: ships."gpu-box".shipToken is ${GPU_BOX_SHIP_TOKEN}, which is unset or empty in the environment; export GPU_BOX_SHIP_TOKEN, or delete the key to register this ship without credentials

A partial interpolation is rejected for the same reason — taking tok-${NAME} literally would register a token that is not the one anybody meant:

fleet launch: ships."gpu-box".shipToken contains "${" but is not exactly one ${VAR} reference; write the whole value as ${VAR}, or as the literal secret

The resolved value is trimmed, so a variable set from a file or a here-doc with a trailing newline still works.

Validation errors

Zod rejects a malformed file before anything starts; the CLI prints fleet launch: <message> and exits 1.

Error	Raised when
config file not found: <path>	The path given by --config-path (or ./fleet-config.yaml) does not exist.
failed to parse config file <path> as YAML: <message>	The file is not valid YAML.
ships "<a>" and "<b>" both use port <port>; give each local ship a distinct port	Two source: local ships share a port. Remote ships are exempt — they have no port field.
gui is configured with no bridge to proxy to; add a bridge section or gui.bridgeUrl	A gui section exists with neither a bridge section nor gui.bridgeUrl.
ships."<key>".<field> is ${VAR}, which is unset or empty in the environment; …	A shipToken/bridgeToken references a variable that is not exported, or is empty.
ships."<key>".<field> contains "${" but is not exactly one ${VAR} reference; …	A shipToken/bridgeToken mixes literal text with an interpolation.
ship "<key>" sets <a> but not <b>; a ship is registered with both a shipToken and a bridgeToken, or neither	Exactly one of the two token keys is set on a ship.
a zod issue list	Any schema violation: source: remote with no url; a non-integer port; a non-boolean insecureNoAuth; an empty dataDirectory, fleetDirectory, name, url, bridgeUrl, shipToken, or bridgeToken; a name that is not a valid fleet identifier; an unknown source value.

Ordering matters when reading a failure: schema parsing runs first, then — per ship, in map order — token interpolation and the both-or-neither check, then the duplicate-port check, then the gui/bridge check.

What launch actually does

1. Loads and normalizes the config, resolving every ${VAR} in a ship’s tokens.
2. If bridge is present, starts the bridge — creating the first admin unless insecureNoAuth is set — and keeps its manager.
3. Plans the registrations against the bridge’s roster: a ship the bridge already holds is skipped, and so is a config entry whose URL an earlier entry claimed.
4. Settles each ship’s credentials in map order: the configured pair if both keys are set, otherwise a freshly minted pair for a source: local ship the bridge is about to register, and none for a source: remote ship or one step 3 skipped.
5. Starts every source: local ship — pinned to the launched bridge’s publicUrl, and handed its pair.
6. Registers the ships step 3 planned, at http://localhost:<port> (local) or their url (remote), printing registered ship "<key>" (<url>) with the bridge for each.
7. If gui is present, serves the GUI against gui.bridgeUrl or the local bridge.

Non-fatal cases to expect in the log:

• With no bridge section, each ship logs no bridge configured; not registering ship "<key>" (<url>) — local ships still start.
• A ship the bridge’s roster already holds logs ship "<key>" (<url>) is already registered with the bridge; one held under the same name at a different URL warns ship "<key>" is already registered with the bridge at <rosterUrl>, not <url>.
• Two config entries on one URL warn ships "<firstKey>" and "<key>" both point at <url>; registering it once.
• A registration that throws logs could not register ship "<key>" (<url>): <message> and the launch continues with the next ship.

Examples

A single machine running everything:

bridge:
gui:
  port: 3000
ships:
  local:

That is a bridge on 4800 with data in ./.fleet/bridge, a GUI on 3000 proxying to it, and one ship named local on 4700 with workspaces under ./fleet/local.

Two local ships plus one already-running remote ship:

bridge:
  port: 4800
  publicUrl: http://10.0.0.2:4800
gui:
  port: 3000
ships:
  ship-a:
    port: 4700
  ship-b:
    port: 4701
    fleetDirectory: /srv/fleet/ship-b
  builder:
    source: remote
    url: http://10.0.0.7:4700

publicUrl is set here because builder is on another host: without it, that ship would be told to pull the armory from http://localhost:4800, which on 10.0.0.7 is 10.0.0.7.

A GUI-only process pointed at a bridge on another host:

gui:
  port: 3000
  bridgeUrl: http://bridge.internal:4800

See configuring a fleet for the task-oriented walkthrough, multi-host for spreading ships across machines, and authentication for what the tokens and insecureNoAuth actually control.