Lodestar Control Plane

Administration for your tenant: who may use Lodestar, which workstations are authorised, what they may run and what it costs.

Overview

Grant access

Find a person in your directory, then grant them one capability at a time with a ceiling on how much risk it may take. Only the pseudonym is stored: the display name is transient and is kept nowhere.

Entitled users

Session lifetime

How long a practitioner stays signed in to the Lodestar runtime before re-authenticating. A Lodestar session is always cut short by the Microsoft Entra token behind it, so this is a ceiling, not a guarantee. Administrators are not affected: privileged sessions keep the plane's own shorter lifetime.

Usage and allowances

Tokens consumed this calendar month (UTC), per person, against the allowance that applies to them. Cached prompt tokens are counted at a tenth of the rate, because that is roughly what they cost — a long conversation that reuses one context is not charged as if it were new every turn. These are counts only: no prompt or response text is stored here or shown here.

Practitioner	Tokens used	Monthly allowance	Used	Limit set by

Devices

Workstations register themselves when someone signs in to the Lodestar runtime. A device is served no configuration and no model credential until it is authorised here.

Device	State	Practitioner	Registered (UTC)	Last seen (UTC)	Actions

Capability publication

A publication is the signed set of capabilities this tenant's devices may run. It is produced and signed offline, promoted here, and verified again on every workstation before anything installs.

Published model configuration

What every authorised workstation in this tenant is told to use. The model pickers below draw their options from exactly this document.

Author this tenant's configuration

Publishing replaces the whole document. Nothing restarts: a device picks the new configuration up the next time it signs in. No secret belongs here — the model credential is the form below.

Provider

Wire protocol

What devices are told to speak. Every provider in the list above is one of these three: the union is fixed by the device contract, so a new vendor rides an existing protocol rather than adding one.

Provider endpoint

Allowed models

The set a device will accept. Everything else — including the model credential form — is offered from exactly this list. Choosing a provider above seeds it, but the seed is a suggestion and not a limit: model identifiers date faster than this console ships, so anything the operator hosts can be added below.

Add a model identifier

Default model

Chosen from the allowed models above — a default outside that set is a configuration every workstation would refuse.

Additional allowed endpoints

Comma- or space-separated. The provider endpoint above is always allowed and does not need repeating here.

Data terms

Confirm what your agreement with this provider guarantees. Lodestar runs no inference until both are ticked: the control plane refuses the model credential and devices refuse to send a prompt. See PROVIDER-DATA-TERMS.md for how each provider grants zero retention.

Prompts and completions are not retained (zero data retention) Prompts and completions are not used for training Basis

A short reference to the contract clause or ZDR agreement, up to 200 characters. Never paste a key or a secret here: it is sent to every device.

MCP endpoints

Source control hosts

The hosts a practitioner may connect a personal access token to. "Read and write" is the most a device may do on that host. A fine-grained token with an expiry is recommended. A host that is not listed here cannot be connected. Enter only the host name, in lower case, for example github.com.

ServiceNow instances

Leave empty to let practitioners register any instance. Add instances to allow only those.

Conversation retention (days)

Each device deletes a conversation this many days after its last activity, from 1 to 3650. Leave empty to keep conversations. A conversation with a run in progress is never deleted.

Work-folder policy

Fixed by the device contract: a work folder is always chosen explicitly, and never outside its root. Shown because it is part of what is published, not because it is a choice.

Model credential

The key is write-only. It is never returned, never rendered and never left in this form: the field is cleared before the request is sent.

TypeSafe / Jev decision plane

Optional. When enabled, workstations ask TypeSafe's Jev model a few narrow questions — which tools a request needs, which Lighthouse items come first, whether a step needs a closer look — so fewer of them reach the frontier model.

TypeSafe receives minimised customer content as an inference (TB-4) processor, called directly by each workstation. Enabling it, and any change made while it is enabled, requires step-up authentication.

Settings

Enabled for this tenant

Jev model

Pin a versioned model (for example jev-1.13.0) rather than an alias. Workstations calibrate their confidence thresholds per model version, so an alias that moves underneath them changes behaviour without anyone deciding it should.

Endpoint

Capabilities

Agent tool routing

Lighthouse prioritisation

Decision guardrails

Lighthouse data policy

What a Lighthouse prioritisation decision may send to TypeSafe: the item's minimised title, its summary of at most 500 characters (with "Title and summary" only), and deterministic facts such as dates and counts. Never message bodies, attachments or credentials.

Project memory strategy

Settings for how workstations remember project knowledge and shorten very long conversations: thresholds, what is saved, how much comes back, and which stages are switched on, per agent. It is configuration only. Nothing a workstation remembers is ever sent here, and this console cannot read it.

A bundle is signed offline by Pulsar's publication key, which this console never holds. Step 1 composes a request and checks it against the limits workstations enforce; the release engineer signs it; step 2 publishes the signed bundle. Publishing needs step-up authentication.

1. Compose a strategy request

Bundle id

Version

Agents

Comma-separated, or * for all. Start with kepler.

Valid for (days)

Rollout stage

Each stage includes the ones before it.

Kill switch: stop new automatic shortening now (stored data is kept)

Advanced limits

Shorten when this full (0.5–0.95)

Notes brought back per request (0–20)

Recent messages never shortened (4–40)

Model-summary allowance, tokens per month (0 turns model summaries off)

Its own allowance, separate from every conversation's. When spent, summaries are written without the model.

Model time per run, minutes (5–240)

Only time the model spends answering counts, not tools or approvals. With per-turn allowances on, it refills at each message.

Tool calls per run (50–3000)

Versions this bundle revokes (comma-separated)

Security findings

Patterns the control plane flagged in this tenant's activity: who, which rule, and how often. No content is ever recorded. Acknowledge a finding once you have looked into it; if the pattern comes back, a new finding opens.

Admin audit

Every privileged change in this tenant, by pseudonym. Written in the same transaction as the change itself, so "changed but unaudited" cannot happen. Timestamps are UTC, like every other time in this console.

Lodestar Control Plane

Overview

Tenant

Grant access

Entitled users

Session lifetime

Usage and allowances

Devices

Workstation detail

Assigned capabilities

Capability publication

History

Published model configuration

Author this tenant's configuration

Model credential

What this tenant's configuration allows

TypeSafe / Jev decision plane

Settings

Project memory strategy

1. Compose a strategy request

2. Publish a signed bundle

Security findings

Admin audit

Narrow this log