Configure PI WEB where your agents work.

Config files

PI WEB uses a global config file for machine-local settings and a project-local file for repository settings.

• Global config: $PI_WEB_CONFIG, or $XDG_CONFIG_HOME/pi-web/config.json, or ~/.config/pi-web/config.json.
• Project config: <project>/.pi-web/config.json for commit-able project settings.

Each PI WEB machine has its own config. When using Fleet/machine federation, Settings uses the selected machine for config that affects work running there: the Pi-compatible agent profile and companion CLI, session daemon tools, PI WEB plugin enablement, external path access, and upload defaults. Gateway/browser-only settings stay local to the gateway: keyboard shortcuts, remote machine registry/tokens, and gateway host/port/allowed-hosts. Remote servers that do not advertise selected-machine settings support report those settings as unavailable instead of silently falling back to the gateway.

Pi package settings are separate from PI WEB config. They live in Pi's package-manager settings on the target machine and are managed by Pi (pi install, pi remove, pi update) or Settings → Pi packages. In a federated setup, Settings → Pi packages targets the currently selected machine. The PI WEB plugins config key only enables or disables discovered PI WEB browser plugins on the machine whose config you are editing; it does not install, remove, or update Pi packages.

If you installed services with a custom config path, rerun pi-web install --config /path/to/config.json after changing that path or after upgrading from a version that only applied the custom path to the web service. This regenerates service files so the web/API and session daemon use the same PI_WEB_CONFIG.

Reverse-proxy deployment paths

The deployment path is not a PI WEB config-file key or environment setting. The published client is portable: one build works at / and at canonical trailing-slash prefixes such as /ai/ or /test/ai/.

For a nested deployment, redirect the slashless prefix to the trailing-slash URL, strip the prefix before forwarding to PI WEB, and proxy authenticated HTTP and WebSocket traffic through the same location. Relative browser and PWA URLs then stay within that prefix. See the reverse proxy deployment example for complete Nginx configuration.

Precedence and reloads

Machine-global runtime values are resolved in this order:

defaults → global config file → environment overrides

Supported project-local settings are then applied for that project's workspaces. For upload defaults, <project>/.pi-web/config.json overrides the global value.

Environment overrides include PI_WEB_HOST, PI_WEB_PORT / PORT, PI_WEB_ALLOWED_HOSTS, PI_WEB_MAX_UPLOAD_BYTES, PI_WEB_AGENT_COMMAND, PI_WEB_AGENT_DIR, PI_WEB_AGENT_SESSION_DIR, PI_CODING_AGENT_DIR / PI_CODING_AGENT_SESSION_DIR for Pi compatibility, PI_WEB_SPAWN_SESSIONS, PI_WEB_SUBSESSIONS, and PI_WEB_ASK_USER.

• host / port: restart the gateway web/API service or process.
• maxUploadBytes: restart both the web/API process and the session daemon on that machine.
• agent.command / agent.dir / spawnSessions / subsessions / askUser / extensionDialogsTimeoutMs: restart the session daemon on that machine.
• pathAccess: applies on the next request; existing file views may need a browser refresh.
• uploads.defaultFolder: applies to newly opened Files upload dialogs and new direct drag/drop batches after config/workspace refresh.
• plugins: reload the browser tab after changing PI WEB plugin enablement.
• Pi package install/remove/update: not a PI WEB config key; after a mutation, type /reload in each idle PI WEB session on the target machine to refresh ordinary Pi resources such as extensions, skills, prompt templates, themes, and context/system prompt files. Reload the browser page separately for PI WEB browser plugin changes. If a global Pi extension adds or removes a provider, or changes a provider's connection settings, manually restart pi-web-sessiond.service; /reload cannot change the startup provider baseline. A known provider refreshing only its own model list is applied without a restart. See Pi extension provider baseline.
• shortcuts: saved settings apply in the browser after config refresh/save.

Global config example

pi-web install creates the initial file. You can also save PI WEB config settings from Settings → General, Settings → PI WEB plugins, Settings → Keyboard, and Settings → Session daemon. Machine-affecting Settings fields target the selected machine; gateway host/port/allowed-hosts and keyboard shortcuts stay local. Pi package operations live separately under Settings → Pi packages.

{
  "host": "127.0.0.1",
  "port": 8504,
  "pathAccess": {
    "allowedPaths": ["~/SDKs", "/opt/reference"]
  },
  "uploads": {
    "defaultFolder": ".pi-web/uploads"
  },
  "maxUploadBytes": 67108864,
  "agent": {
    "command": "pi",
    "dir": "~/agent-profiles/research"
  },
  "spawnSessions": true,
  "subsessions": false,
  "askUser": true,
  "extensionDialogsTimeoutMs": 300000,
  "plugins": {
    "workspace-tasks": { "enabled": true },
    "updates": { "enabled": true },
    "info": { "enabled": false }
  },
  "shortcuts": {
    "core:view.chat": "mod+1",
    "core:session.stop": null
  }
}

Project-local config

Project-local config lives at <project>/.pi-web/config.json. Use it for settings that should follow a repository.

{
  "version": 1,
  "pathAccess": {
    "allowedPaths": ["~/SDKs", "/opt/reference"]
  },
  "uploads": {
    "defaultFolder": "manual/uploads"
  }
}

Project-local pathAccess.allowedPaths entries are merged after the global list and deduplicated. Paths must still be host-absolute or ~-prefixed; relative roots are not supported.

Project-local uploads.defaultFolder overrides the global upload destination for workspaces in that project. Current PI WEB servers include this workspace-effective value on local and federated workspace responses; older remote servers may omit it and the browser falls back to the global/default upload folder.

Plugins may own separate project files, such as .pi-web/tasks.json for the built-in Workspace Tasks plugin.

Config matrix

Use this table as the quick reference for where a setting can live, which environment variable overrides it, and whether project-local config overrides or merges with global config. Rows with JSON key — are runtime-only environment variables, not config-file keys. Global means machine-global. In Settings, selected-machine-safe global keys (pathAccess, uploads, maxUploadBytes, agent, spawnSessions, subsessions, askUser, and plugins) are edited for the selected machine; gateway host/port/allowed-hosts, keyboard shortcuts, and machine registry/tokens stay local.

External path access

pathAccess.allowedPaths grants PI WEB's file explorer and absolute @ path completions access to specific filesystem roots outside the current workspace. By default, workspace-relative file reads stay inside the workspace and absolute paths are denied.

Accepted root forms:

• Unix absolute paths, for example /opt/reference.
• Home-relative paths, for example ~/SDKs.
• Windows absolute paths on Windows hosts, for example C:\Users\dev\SDKs.

When an absolute request is served, PI WEB expands ~, canonicalizes configured roots with realpath, requires roots to be existing directories, and rejects symlink escapes outside the allowed roots.

In Settings → General, external filesystem roots are saved on the selected machine. Gateway host, port, and allowed-hosts fields stay on the gateway config.

Manual upload defaults

The Files panel can upload files by dropping them onto the panel or by using the toolbar Upload button. uploads.defaultFolder sets the workspace-effective default destination. The built-in default is .pi-web/uploads; a project-local value overrides the global value for workspaces in that project.

The value must be a non-empty workspace-relative folder. PI WEB normalizes repeated separators and backslashes to /, and rejects absolute paths or .. traversal. In the upload dialog only, clearing the destination field uploads that batch to the workspace root.

Manual uploads use the workspace file-write path: paths stay workspace-relative, parent folder creation is enabled by default, and overwrite is disabled by default. Browser-owned XHR progress is shown per batch/file, and conflicts or errors stay visible in the upload progress UI.

For machine federation, Settings saves the global upload default on the selected machine. Current remote PI WEB servers also return workspace-effective upload defaults on workspace responses; older remote servers may omit them and the browser falls back to the global/default upload folder.

Pi-compatible agent profile and companion CLI

agent.command selects the Pi-compatible companion CLI used by pi-web doctor and, when it can be generated safely, package-managed update commands. It defaults to pi. This setting does not replace the embedded runtime: every session continues to use PI WEB's bundled Pi SDK.

agent.dir selects the Pi-compatible state profile used for auth providers, models, settings, sessions, Pi packages, and Pi-package-backed PI WEB plugin discovery. It defaults to ~/.pi/agent only for a canonical Pi companion command. The directory must use the data layout supported by the bundled Pi SDK; PI WEB does not load or convert incompatible fork formats, migrate profile data, or repartition PI WEB-managed archives when the profile changes.

{
  "agent": {
    "command": "pi-lab",
    "dir": "/opt/pi-profiles/lab"
  }
}

An alternate command always requires an explicit state directory. The command must be a safe bare executable name such as pi-lab or a host-absolute executable path such as /opt/pi/bin/pi; relative paths, shell expressions, and launcher strings are rejected. The state directory must be host-absolute or start with ~. In a federated save, the gateway transports Unix and Windows absolute paths without reinterpreting them, and the target machine validates and returns the persisted profile.

Environment variables take precedence over the config file. PI_WEB_AGENT_COMMAND selects the companion CLI, PI_WEB_AGENT_DIR sets the profile state directory, and PI_WEB_AGENT_SESSION_DIR overrides session storage separately from agent.dir. The legacy PI_CODING_AGENT_DIR and PI_CODING_AGENT_SESSION_DIR names apply only to a canonical Pi companion command; PI WEB never derives ambient environment-variable names from an arbitrary command. Use the explicit PI_WEB_AGENT_* names for alternate commands. PI_WEB_AGENT_DIR is an unconditional override, while a legacy PI_CODING_AGENT_DIR override stops applying when Settings selects an alternate command so the command and directory can transition together.

The session daemon resolves the persisted desired values plus its environment once at startup. That secret-free active profile stays fixed for the daemon lifetime. Settings → Session daemon saves command and directory together as desired configuration and shows whether the profile is active, needs a restart, or cannot be compared. Until the daemon restarts, sessions, Pi package operations, Pi-package-backed PI WEB plugin discovery, status/install detection, and update planning continue to use the daemon-owned active profile; a web/API restart recovers that same active profile instead of applying the newly saved values.

Pi extension provider baseline

This policy applies to Pi runtime extensions, not PI WEB browser plugins. Pi extensions are runtime modules loaded by the session daemon and can call pi.registerProvider(...); PI WEB plugins are browser-side UI modules and never run in the session daemon.

PI WEB shares one model runtime across all sessions. When the session daemon starts, before any project resources load, it initializes global Pi extensions from the active agent profile (agent.dir), including extensions supplied by globally configured Pi packages. Provider registrations made by synchronous or awaited asynchronous extension factories during this bootstrap join the shared baseline. PI WEB captures both config-form registrations (pi.registerProvider("id", config)) and native-provider registrations (pi.registerProvider(provider)), alongside Pi built-ins, environment credentials, and providers from the active agent directory's models.json.

After startup capture, a provider's connection settings are fixed for the daemon lifetime. Later attempts to add a provider, replace an existing provider's configuration, register a native provider, or unregister a provider are no-ops, regardless of source or provider ID. This includes project extensions attempting to add or replace a provider, lifecycle callbacks such as session_start, and /reload. Non-provider Pi extension features continue to load and reload normally.

Model list refresh for a known provider

One narrow update is applied after startup: a provider captured in the baseline may refresh its own model list. Extensions that fetch an updated catalog typically re-send their complete provider configuration, so PI WEB compares the incoming registration against the recorded baseline and applies it only when both hold:

• the provider ID is already in the startup baseline, and
• every field except the model list is unchanged — name, baseUrl, apiKey, api, streamSimple, headers, authHeader, oauth, and refreshModels.

Anything else stays a no-op, including a provider that was not in the baseline and a known provider whose credentials, base URL, or API surface differ from startup. Function-valued fields cannot be compared by value, so a registration that supplies a new streamSimple, refreshModels, or oauth implementation is treated as a change and ignored.

An applied refresh becomes the new comparison point, so a provider can refresh repeatedly. Re-sending an unchanged model list is a replay rather than an update and is ignored. Refreshed models are visible to sessions immediately; no restart and no network request is involved, because the extension has already produced the catalog.

Model lists are shared daemon-wide state. If extensions in two workspaces register different model lists for the same provider ID, the last registration wins. A model entry may also carry its own baseUrl and headers, which take precedence over the provider-level values for that model, so an accepted refresh can change where requests for those models are sent. Both are accepted trade-offs: a catalog is treated as a property of the provider rather than of the project, and Pi extensions are trusted daemon code.

Provider decisions in the daemon log

Ignored mutations are written to the session-daemon log once per operation and provider ID, so a replaying extension cannot flood the log. Applied model list refreshes are logged every time, with the resulting model count, because each one changes shared runtime state. Neither entry contains provider configuration or credentials, and PI WEB does not show a session warning or notification.

This prevents accidental provider, configuration, or credential contamination between projects; it is not a security boundary because Pi extensions remain trusted daemon code.

Configure providers before the daemon starts: use the active agent directory's models.json, or install the Pi extension globally in that agent profile. Project Pi extensions and project-level models.json files cannot add providers to PI WEB's shared baseline.

Background model catalog refresh

PI WEB shares one model runtime across all sessions, and provider model catalogs are refreshed over the network only on the session daemon's own background schedule. Requests never start a catalog fetch of their own, so a slow or unreachable provider cannot stall opening the model selector, starting a session, or the auth dialogs on its own account.

A refresh that is already in flight can still briefly delay starting or opening a session, because the shared runtime is read while that refresh is running. PI WEB says so while you wait: the session's activity line names the startup step it is on and adds provider model lists are refreshing when a background refresh is running at the same time. That note reports what is happening concurrently, not a proven cause.

The session daemon runs the refresh:

• 15 seconds after the daemon starts, then hourly. Pi treats stored catalogs as fresh for four hours, so most hourly ticks make no network request at all; the shorter tick only makes sure a due refresh is not delayed to the next tick.
• Immediately after a provider login or logout, bypassing that freshness window, because the cached catalog is known to be wrong.

Each run is bounded: it is aborted after 60 seconds, and a run that times out or cannot reach a provider earns one retry after five minutes; a provider that answers with an error status is retried on the next scheduled refresh instead. Failures never clear the stored catalogs — the last successfully fetched models stay in use and the daemon log records what failed. A refresh in flight is also aborted when the daemon shuts down.

Models fetched by a background refresh appear the next time a client asks for the model list, so a model selector left open across a refresh may need to be reopened.

To turn the background refresh off entirely, set PI_WEB_OFFLINE or PI_OFFLINE in the session daemon's environment and restart it. In offline mode PI WEB performs no provider catalog network requests, including after logins, and sessions use the catalogs already stored in the agent profile. The PI_WEB_SKIP_VERSION_CHECK and PI_SKIP_VERSION_CHECK keys do not affect this refresh; they only suppress PI WEB release checks.

Session daemon tools

spawnSessions

Boolean. Controls whether agents receive the spawn_session tool. Defaults to true. Set it to false if you do not want an agent to start independent PI WEB sessions.

Environment override: PI_WEB_SPAWN_SESSIONS=0|1|true|false.

subsessions

Boolean. Beta. Controls whether agents receive the tracked-subsession tools: spawn_subsession, list_subsessions, check_subsession, read_subsession, and yield_to_subsessions. Defaults to false and also requires spawnSessions to be enabled.

Tracked subsessions are join-oriented. Calling spawn_subsession returns immediately, so the parent can continue independent work while the child runs. Work whose result the parent does not need to join belongs in the fire-and-forget spawn_session tool instead.

At a join point, after finishing its independent work, the parent calls yield_to_subsessions alone as the final action in its tool batch. Pi ends a tool batch early only when every result in that batch is terminating. If any tracked child is still working, the action ends the current agent run so the parent becomes idle. If none are working, it does not end the run and clearly reports that there is nothing to wait for.

A completion notice wakes an idle parent or queues behind in-flight work. Each notice lists any other tracked children still working, so the parent can continue work or call yield_to_subsessions again at the next join point. Further notices arrive automatically; do not poll. The notice includes the child's final output when it fits. If that output is too long, PI WEB omits it entirely instead of adding a truncated duplicate to the parent's context and directs the parent to retrieve it with check_subsession.

list_subsessions, check_subsession, and read_subsession never yield or change control flow. They are for deliberate inspection or recovery, not completion polling. While a child works, agent-facing check_subsession and read_subsession withhold partial output and direct the parent to continue independent work or yield at the join point. Output becomes available when the child stops. Included output and transcripts follow a labeled marker and come last, after PI WEB guidance.

Both spawn_session and spawn_subsession accept an optional model parameter, given as an exact provider/model-id such as anthropic/claude-sonnet-4-5. When set, the new session starts on that model instead of inheriting the dispatching session's model. The match is strict: an unknown or malformed value is rejected with an error. A #provider/model-id reference in the prompt (see Prompt completions) is how users ask for a specific model; agents forward that reference as this parameter. The new session also inherits the dispatching session's thinking level, clamped to its model's capabilities.

In Settings → Session daemon, these keys are saved on the selected machine. Restart the session daemon on that machine after changing them.

Environment override: PI_WEB_SUBSESSIONS=0|1|true|false.

askUser and ask_user

askUser controls whether agents receive the core ask_user tool. It defaults to true; set it to false, or set PI_WEB_ASK_USER=false, to remove the tool. The environment override accepts 0|1|true|false and takes precedence over the config file.

Use Settings → Session daemon → Allow agents to ask questions to change askUser on the selected machine. An environment override makes the toggle read-only.

The tool accepts one set of 1–20 questions. Each question has a unique id, its question text, optional supporting detail, up to 12 options with stable values and user-facing labels, and an optional multiple flag. The browser always adds a Custom free-text answer, including when the model supplies no options. No question is required: the user may leave any of them unanswered.

Calling ask_user posts the whole set as one browser form and ends the current agent run instead of waiting for the user. The open form is owned by the session daemon, so it survives a browser disconnect, browser reload, or web/API restart while that daemon keeps running. When the user submits, the answers arrive as a follow-up that wakes the session; each question is reported with its selected option values or free text, or explicitly as unanswered.

PI WEB confirms a partial submission before sending it and names the unanswered questions. Only one ask can be open per session: a later ask_user call supersedes the earlier one, reports that fact and its unanswered questions to the model, and turns the earlier card into a read-only transcript record. Submitted and cancelled asks likewise remain readable in the transcript.

Sending an ordinary chat message while a form is open voids the form: the card closes as cancelled and the model is told its questions went unanswered as part of the turn the message itself starts.

Extension dialogs

Pi extensions can ask the user questions from ctx.ui.confirm(), ctx.ui.select(), and ctx.ui.input() — including from session_start hooks and in-flight tool_call hooks. PI WEB renders these dialogs inline in the session transcript and answers them through a dedicated session-daemon channel, never the prompt queue, so a dialog parked inside a tool_call hook cannot deadlock the run. Dialog support is always on; there is no enable flag. See Pi extension dialogs in PI WEB for behavior details and author guidance.

extensionDialogsTimeoutMs is the unattended-dialog safety valve: how long the session daemon waits for an answer before settling the dialog with its kind's cancel value (false for confirm, undefined for select and input). It defaults to 300000 (5 minutes); set it to 0 to wait forever. An extension's own timeout option still applies, and the effective deadline is the sooner of the two.

Prompt completions

The chat composer opens completion menus on three trigger characters:

• / at the very start of the draft completes session commands.
• @ completes file paths: @ for tracked files, @ (at, then space) or !@ for all files. Picking one inserts an @path reference into the draft, quoted automatically when the path contains spaces.
• # completes the models available to the session, filtered case-insensitively as you type (at most 12 entries). Picking one inserts a #provider/model-id reference into the draft, which tells agents the request should run on that model — for example as the model parameter of spawn_session.

Optional completion tools

File and path @ completions work without extra tools. If fzf is available on the PI WEB server's PATH, PI WEB uses it to improve completion filtering and ranking; otherwise it falls back to built-in ranking.