Repository navigation
Artifacts
| Artifact | Get it |
|---|---|
| Docker image | docker pull ghcr.io/brtydse100/litegate:2.18.0GHCR |
| Live demo | Current demo |
| Docs | Versioned documentation |
| Source | Source |
| License | MIT license |
What's Changed
- Add optional per-model key budgets and speed up policy editing by @brtydse100 in #47
- Discover Advanced policy settings and lazy-load organization branches by @brtydse100 in #48
Independent model budgets within a key allowance
Administrators can set a model-specific dollar amount and reset period on
selected existing keys, such as $25 for an Opus model alias every two hours.
Model spending also counts toward the key's overall budget; both limits apply
independently. Explicit replacement supports editing or clearing the complete
per-model mapping. Zero-dollar model limits are rejected because LiteLLM treats
zero as unlimited. Limits must be positive; use allowed-model restrictions to
deny access instead. LiteLLM performs accounting, resets and enforcement.
Per-model budgeting is off by default. Activate a valid LiteLLM Enterprise
license on the upstream proxy, then set LITELLM_ENTERPRISE_ENABLED=true in
LiteGate and restart it. Docker Compose YAML uses
litellm_enterprise_enabled: true; Helm uses
config.litellmEnterpriseEnabled=true. This is a manual feature switch, not
automatic license detection. Turning it off hides editing and rejects new
per-model API writes without removing existing upstream limits.
Keys with per-model budgets require administrator-assisted rotation in LiteLLM.
LiteGate rejects self-service regeneration before creating, renaming or revoking
keys, so replacement cannot silently drop limits or restart model usage windows.
This safeguard also applies when the feature switch is off. See
per-model budget configuration.
Advanced settings discovered from the installed LiteLLM
Key and team editors retain familiar controls and discover additional fields
under Advanced settings from the connected proxy's OpenAPI request schemas.
Supported newly advertised fields can appear without adding matching LiteGate
form definitions. Typed controls handle simple values; JSON inputs handle nested
settings. Only opted-in edits are sent, and changing targets discards drafts.
Missing schema discovery leaves familiar controls usable.
Discovery and writes remain administrator-only. Bounded caching, local-reference
resolution and server-side validation exclude credentials, identity/member
bindings, spend counters and dedicated workflows. Existing Enterprise and
organization-owned allowance protections still apply. Unknown schema constraints
are omitted safely. The approved jsonschema dependency is included in the image;
source installations should install the updated backend requirements. No new
service or persistent state is required.
The pinned LiteLLM 1.98 integration supports discovered MCP rate/permission
settings and team-member allowance defaults. It does not advertise per-model
team-member budgets; LiteGate does not invent unsupported features. An unrelated
Advanced MCP team update now preserves existing metadata and member-budget
associations instead of dropping the member allowance. License entitlement and
breaking upstream semantics can still require compatibility work. See
discovered policy settings.
Progressive organization navigation and responsive editing
The Organization page shows root-group metadata first. Selecting or expanding a
group requests its direct children and live team details; deeper branches load
as opened and revisits use cached results. Users and usage reports wait for an
explicit selection. Loading and failures are independent so an unresolved inner
team does not block navigation. Team spend is labelled separately from subtree
usage aggregates, and the legacy full-overview API remains compatible.
Key policies use the page scrollbar for key/model lists and cache selection
lookups, avoiding repeated full-selection searches while editing large key sets.
The synthetic public demo and operator/API guides cover the new workflows.
Upgrade and validation
Pin image 2.18.0 and restart the deployment. Existing settings remain valid;
the Enterprise switch defaults to false. Review the regeneration restriction
above if model budgets were previously configured directly in LiteLLM.
Validation included 696 backend tests, 115 frontend tests, 13 browser tests and
two demo tests, plus pinned live LiteLLM compatibility checks and a Docker build.
Local simulated Enterprise activation verified actual LiteLLM model-budget
enforcement and key/team MCP-setting persistence with PostgreSQL and synthetic
model responses. Genuine license verification and external MCP tool invocation
were not tested.
Full Changelog: v2.17.0...v2.18.0
Contributors
Thanks to @brtydse100 for the changes in this release.