API & Route Matrix

Primary surfaces of the FastAPI app (kazma_ui). Auth is default-deny for /api/* unless listed open. HITL danger applies to tool execution, not every HTTP route.

Always open (health)

Method	Path	Auth	Description
GET	/health	Open	Liveness
GET	/health/live	Open	Process live
GET	/health/ready	Open	Readiness (DB ping when configured)

Auth & account

Method	Path	Auth scope	Description
GET/POST	/login	Public page	Multi-mode login (user / secret / OIDC)
POST	/api/auth/*	Varies	Login, logout, session (see auth.py, saas_api.py)
GET	/api/saas/*	Admin / operator	Tenants, platform users (RBAC)

Chat & sessions

Method	Path	Auth	HITL	Description
POST	/api/chat/stream	Session/secret	Graph interrupt on danger tools	SSE chat stream
POST	/api/approve/{thread_id}	Session	Resumes graph HITL	Approve danger tool
*	Session CRUD under /api/sessions*	Session	—	Thread list / history (session_manager)

IDE

Method	Path	Auth	HITL	Description
*	/api/ide/*	Session	Via tool registry	Files, apply_patch, codebase search, LSP (GET/POST /lsp), run, git, swarm send (ide_api.py)
GET	/ide	Session	—	IDE page

Swarm / Command Center

Method	Path	Auth	Description
*	/api/swarm/*	Session (GET) / Admin (everything else)	Workers, dispatch, tasks, metrics. One table, kazma_ui/swarm_panel/route_roles.py; refused with 409 swarm_single_operator on an install with platform users
POST	/api/swarm/workers/{name}/circuit-breaker/reset	Admin	Close a worker’s breaker now (the worker card’s Reset, shown while the breaker is open or half-open)
GET	/api/swarm/approvals/{id}	Session	One swarm-bus approval request, for the Swarm page’s approval panel (the link on every chat approval card): the tool, its state, the whole call the gate record keeps, the hash the card shows, and who decided. A chat approval’s gate is not found here
POST	/api/swarm/approvals/{id}/decide	Admin	{"decision": "approve" | "deny"}: answers the request from the web, the same shared decision as the chat buttons (platform: web). 409 not_pending once it was answered or ended
GET	/api/swarm/tasks/export	Session	Task history as JSON, or CSV with ?format=csv (newest 1,000; filters type, status, q as the history list). Task History’s Export CSV
GET	/api/swarm/tasks/{id}/stream	Session	Live task stream (swarm_sse): the task’s events so far, then live ones until it ends. A finished task’s stream always ends with task_completed and the task’s result, rebuilt from the stored task when the server no longer holds its events (after a restart).
GET	/swarm	Session	Swarm panel page
*	/api/replay/*	Session	Time travel: threads, snapshots, restore, fork, compare, clear (replay_routes.py)
GET	/replay	Session	Time Travel panel page
*	/api/research/*	Session	Research: list, detail, compare, export, archive, unarchive (research_panel/routes.py). A report a request names is a path inside a research/reports folder; nothing else is read
POST	/api/research/sessions	Session	Start deep research session (background pipeline)
GET	/api/research/sessions	Session	List durable research sessions
GET	/api/research/sessions/{id}	Session	Session status / log / report path
GET	/api/research/sessions/{id}/stream	Session	SSE progress (snapshot / progress / done)
GET	/api/research/eval	Session	Structural rubric for ?path=research/reports/… or ?session_id=
POST	/api/memory/v2/eval/golden	Session	Run golden memory recall cases (pass rate)
GET	/research	Session	Research panel page (start form + live progress)

Memory / RAG

Method	Path	Auth	Description
GET	/api/system/status	Session	Memory health board + embedder/layer probes (build_memory_health)
GET	/api/memory/graph	Session	Property graph JSON (nodes/edges); optional ?q= filter
GET	/api/memory/graph/stats	Session	Node/edge counts + backend path
GET	/api/memory/graph/search	Session	FTS search over graph nodes (?q=&limit=)
POST	/api/memory/graph/clear	Session	Bi-temporal invalidate of active V2 beliefs for one tenant (?tenant= defaults to default; no all-tenants mode). Tombstones the PG mirror, deletes Neo4j edges, writes a graph_clear audit row. UI confirms.
*	/api/system/memory/*	Session	Backup / restore / maintenance of memory stores

V2 cognitive engine (/api/memory/v2/*)

Active when memory.v2.use_new_stack is true. All routes return shaped JSON on error (never a bare 500); non-numeric params yield a FastAPI 422.

Method	Path	Auth	Description
GET	/api/memory/v2/health	Session	V2 health snapshot — active/superseded/archived belief counts, episode/entity/procedural stats, queue depth. Drives the dashboard KPI grid (pollV2Health, 5s cadence).
GET	/api/memory/v2/beliefs	Session	Active beliefs list. ?q= FTS filter, ?limit= (default 50, clamped 1–200).
GET	/api/memory/v2/beliefs/{id}	Session	Belief detail + supersede chain.
POST	/api/memory/v2/beliefs/{id}/invalidate	Session	Soft-invalidate one belief (+ best-effort Neo4j edge delete).
POST	/api/memory/v2/beliefs/invalidate-batch	Session	Soft-invalidate many ({ "ids": [...] }).
PATCH	/api/memory/v2/beliefs/{id}	Session	Operator edit of active triple: optional subject, predicate, object, predicate_type. Sets extraction_method=user_explicit; clears embedding if object changes.
GET	/api/memory/v2/graph	Session	Belief graph {nodes, links, stats, groups} for the canvas. Bi-temporal + filter params: ?at=<unix_ts> (point-in-time scrub; superseded beliefs marked superseded=true), ?type= (functional/set/state predicate_type), ?entity_type= (person/tool/concept/…), ?limit= (default 200), ?source=neo4j (optional probe). stats.total_links vs stats.links is the slicing delta shown on the truncation banner. Invariants: unique node ids; no virtual fact node when object text equals an entity id; no dangling links; hub node id=user with display name from entities.user (self person shells collapsed onto hub); payload-object subjects carry a hub related_to anchor.
GET	/api/memory/v2/entities	Session	Entity list for /memory ops. Flags: empty, isolated, protected, core (always protected, never unprotectable), is_self, graph_id (self shells → "user"). Query: ?q=, ?empty_only=, ?isolated_only=, ?limit=.
POST	/api/memory/v2/entities/{id}/protect	Session	{ "protected": true|false }: a protected entity cannot be deleted or merged away (the Memory page’s Protect / Unprotect). Core entities refuse false.
POST	/api/memory/v2/entities/{id}/rename	Session	Display rename only ({ "name": "…" }). Id stable; aliases preserved. Self/person User shells also upsert hub entities.user. Returns hub_synced, graph_id.
POST	/api/memory/v2/entities/merge	Session	Merge source into target (beliefs rewired, aliases union).
POST	/api/memory/v2/entities/link	Session	Create belief edge (subject, predicate, object).
DELETE	/api/memory/v2/entities/{id}	Session	Delete entity shell (blocked for protected ids: user, assistant, …). Copies matching entity_merges rows to entity_merges_archive before dropping live ledger rows (FK).
GET	/api/memory/v2/admin/summary	Session	Counts for ops chips (live/invalidated beliefs, empty/isolated entities).
GET/POST	/api/memory/v2/hygiene/*	Session	Preview + run empty purge / near-dup invalidate / archive.
GET/POST	/api/memory/v2/entity-merges*	Session	Quarantine merge list + approve/reject.
GET/POST/DELETE	/api/memory/v2/graph/groups*	Session	View-only groupings (list/create/delete/move/tier). Never mutates beliefs. Create refuses a node that already has a parent (the answer names the move route); Group under on a grouped node moves it, and Tier… sets a grouped node’s tier. Canvas poll uses groups on GET /graph; Ungroup is DELETE …/groups/{id}.
GET	/api/memory/v2/graph/export	Session	On-demand JSON or GraphML (?format=json|graphml).
POST	/api/memory/v2/probe	Session	Recall dry-run (explain chips).
POST	/api/memory/v2/federated-search	Session	Memory + KB labeled search.
POST	/api/memory/v2/eval/golden	Session	Golden recall suite.

Page: GET /memory (HTML admin). Guide: Memory & RAG · Memory best path.

Settings & config

Method	Path	Auth	Description
*	/api/settings*, config export	Admin/operator	ConfigStore-backed settings UI
PUT	/api/settings/single, /api/settings	Admin	Save one setting, or several in one transaction. Both apply the same value checks (settings_validation): a time zone the scheduler cannot resolve, a retention the sweep cannot read, a server-status event Kazma does not send answer 400 and nothing is written.
GET	/api/settings/system/backup	Admin	The settings backup (YAML): every stored setting, keys as vault references, no credential and none of Kazma’s own state. Configuration.
GET	/api/settings/export?format=yaml|json	Admin	The same backup as YAML or JSON.
POST	/api/settings/system/restore?dry_run=&sections=&expect=	Admin	Body: a backup (YAML or JSON, at most 10 MB). dry_run=true returns the plan (changed, keys_restored, keys_to_reenter, refused, kept, digest) and writes nothing. Without it the plan is written; with expect (the preview’s digest) a plan that changed since is refused with 409 and the plan as it is now. 400 for a file that is not a backup, 413 over 10 MB, 503 while settings are held in memory (database away). sections=agent,providers restores only those.
GET / POST	/api/settings/system/restore/undo	Admin	GET: what undoing the last restore would put back (available). POST (expect= the GET’s digest): put it back; a setting changed again since and a key the restore brought back stay. 409 when there is nothing to undo.
GET/PUT	/api/settings/agent/nonstop	Session / Admin	Non-Stop & Self-Healing watchdog/failover/ledger settings
*	Workspace routes /api/workspaces*	Session	WorkspaceStore CRUD
GET/PUT	/api/workspace/extra-roots	Session	Folders outside the workspace the agent may use without asking (Workspace page). PUT replaces the list; a new root must be a full path to an existing folder, not a whole drive and not one holding Kazma’s own files (400 says why). A failed read is a 500, never an empty list.
GET	/api/notifications/turn-complete	Session	The “notify when a task finishes” switch, the one reader the Web Push sender uses too

Documents / Document Intelligence

Mounted from documents_api.py at /api/documents/*. Thin transport over the shared DocumentIngestionService (no parallel parser path). Tenant, actor and workspace come from the request, never from the body: the tenant from the tenant middleware, the actor from the signed-in principal’s name (operator for the shared secret sent as a header, local for a request without a credential), the workspace from the active workspace row. Clients never supply trusted server paths.

A document belongs to the actor that ingested it; every read, job read and action checks the document’s ACL. A document or job the caller cannot read answers exactly like one that does not exist (404). On a single-operator install these routes, the chat commands and the agent’s tools all act as the install’s operator (documents.principal, since 2026-10-13), so each sees the documents the others ingested; with platform users (multi-user mode) these routes act as each user, and chat and the agent refuse.

Guide: Document Intelligence · Ops: Document processing · Security: Document security.

Intake, library, jobs

Method	Path	Auth	Description
POST	/api/documents	Session	Streamed upload. Body: the file’s bytes; filename in X-Document-Filename (URL-encoded) or ?filename=; optional X-Idempotency-Key and ?force_ocr=1. The declared size, writes switched on, the storage floor, the queue, the tenant’s caps and its rate window are answered before the body is read (see the answers below; the tenant byte quota is checked while the file is stored); at most documents.capacity.max_concurrent_uploads (default 4) uploads are received at once per process (503 uploads_busy). The body is spooled under the storage root and cut off at documents.intake.max_bytes or at its declared length (413).
POST	/api/documents/import	Session	Ingest a file of the active workspace: {"path": …} (relative to the workspace, or absolute inside it); 400 path_outside_workspace otherwise
POST	/api/documents/generate	Session	Generate + durably ingest: target_format, payload (an object of at most 1 MiB serialized, 413 above; nesting deeper than 32 is a 400), optional output_name, title
GET	/api/documents	Session	The documents the caller can read, each with its newest job’s state
GET	/api/documents/health	Session	Parser/OCR/renderer/PDF-mutator readiness + worker status (read-only diagnostic). With KAZMA_DOCUMENT_WORKER_ISOLATION=container the engines are those the worker image reports
POST	/api/documents/merge	Session	Merge the PDFs of document_ids (at least two) into a new artifact
GET	/api/documents/{document_id}	Session	Detail (versions + jobs + artifacts)
GET	/api/documents/{document_id}/versions	Session	Version list
GET	/api/documents/{document_id}/content	Session	Normalized content, fenced as untrusted data: ?page= or ?page_start=&page_end=, offset, max_chars (default 20,000, at most 200,000), with a continuation. 409 not_ready before processing finishes
GET	/api/documents/{document_id}/artifacts	Session	Derived artifacts for a document
GET	/api/documents/artifacts/{artifact_id}/download	Session	Download an artifact by opaque ID (404 when the caller cannot read it)
POST	/api/documents/{document_id}/convert	Session	Convert the current version: target_format, optional output_name → a new artifact
GET	/api/documents/{document_id}/pdf-info	Session	Structural PDF report
POST	/api/documents/{document_id}/split	Session	Split a page range: start_page, end_page, optional output_name
POST	/api/documents/{document_id}/fill-form	Session	Fill AcroForm fields: fields (object), optional output_name
POST	/api/documents/{document_id}/redact	Session	Physically redact terms (list) → a new artifact; the original is unchanged. No confirmation dialog on the API (the page asks first)
POST	/api/documents/{document_id}/index	Session	Publish the current version to the Knowledge library library_id
POST	/api/documents/{document_id}/unindex	Session	Remove it from the library library_id
POST	/api/documents/search	Session	Search a library: library_id, query, top_k (default 5, 1–50); hits come back inside an untrusted-data fence
GET	/api/documents/jobs/{job_id}	Session	Job status. 404 Job not found for an unknown job and for one whose document the caller cannot read, identically
GET	/api/documents/jobs/{job_id}/events	Session	Append-only job event history (the same 404 rule)
POST	/api/documents/jobs/{job_id}/cancel	Session	Cooperative cancel (the same 404 rule; 409 for a job that can no longer be cancelled)
POST	/api/documents/jobs/{job_id}/retry	Session	Re-run a dead_letter or rejected job as a fresh job (409 for any other state). Needs write permission on the document and passes the same admission as an upload (429/503/507)
POST	/api/documents/{document_id}/delete	Session	Tombstone: the document leaves every Knowledge library first (a library it cannot leave stops the delete, 500 document_unindex_failed, so a retry can finish it), then its metadata is archived; garbage collection reclaims the bytes later. Optional body {"reason": …}
DELETE	/api/documents/{document_id}	Session	The same tombstone as POST …/delete (one implementation); the reason goes in ?reason=

Answers shared by these routes. The routes with a JSON body (import, generate, merge, convert, split, fill-form, redact, index, unindex, search, and the optional body of POST /{id}/delete) read at most 2 MiB: a declared length past it is refused before a byte is read and an undeclared one is cut off (413 request_too_large); malformed JSON, a body that is not an object or nesting too deep is a 422 invalid_json. A live document policy that cannot be read or holds an invalid saved value refuses the call with 503 and Retry-After: 30 (document_policy_unavailable). One status per error code on every route (documents_api._STATUS_BY_CODE, since 2026-10-13; the route helpers kept three tables that disagreed): a refusal the server’s state causes is a 503 with Retry-After: 30 wherever it happens — durable writes switched off (document_platform_disabled: reads, PDF info and cancel keep working), a required malware scanner that cannot run (malware_scanner_unavailable, answered by the upload before its body is read), no permitted worker profile (worker_isolation_*). A code the table does not name keeps its route’s default: 400 for intake and lifecycle, 422 for the action routes. Refusals of upload, import and retry carry a code; the 429, 503 and 507 ones also a reason and a Retry-After header:

Status	Code	When
413	intake_too_large, quota_exceeded	Past documents.intake.max_bytes (declared or received), or past the tenant byte quota
507	storage_full	Free space minus the incoming file would fall below documents.capacity.storage_free_floor_bytes
503	queue_full, capacity_unavailable, uploads_busy	The durable queue is at documents.capacity.max_queued_jobs; a capacity measurement (storage, queue, tenant load, admission store) could not be made; too many uploads being received
429	tenant_queue_full, tenant_active_full, rate_limited	The tenant’s queued/active caps or its files/bytes-per-minute window
409	idempotency_conflict, intake_in_progress, intake_taken_over	The key was used for another request or by another actor; the same key is still being received; the upload took so long its lease lapsed and another process rolled it back (nothing was saved: send it again)

When a secret is configured, upload, import, generate, merge, convert, split, fill-form, redact, index and search also count against the per-principal HTTP rate limit bucket documents (429 with Retry-After): 10 requests a minute, search allowed up to 30 in the same window; a saved api.rate_limit.documents_per_minute replaces both.

Operations (/api/documents/ops/*)

Method	Path	Auth	Description
GET	/api/documents/ops/metrics	Session	Content-free metrics snapshot
GET	/api/documents/ops/capacity	Session	Backpressure snapshot + degraded_reasons
GET	/api/documents/ops/readiness	Session	Read-only: storage backends and multi-replica honesty (jobs_multi_replica, metadata_multi_replica; a SQLite backend chosen on purpose is not degraded), policy (configured vs effective, restart_required), malware, worker_isolation (hostile_document_profile, image_code_matches), stranded_intake_reservations, rollout, accepting_durable_writes. status is degraded when a configured backend fell back or degraded_reasons names a problem (worker isolation unavailable, stranded intake reservations, an outdated worker image, malware scanning blocking intake), disabled while durable writes are off
GET	/api/documents/ops/retention	Session	Live retention/GC policy view
GET	/api/documents/ops/audit	Admin	Keyset-paged operational audit of the tenant (?document_id=&event_type=&limit=&before_id=). Admin-only: the trail names documents and actors the caller may not be able to read
POST	/api/documents/ops/maintenance/dry-run	Admin	GC dry-run report
POST	/api/documents/ops/maintenance/run	Admin	GC run (the page always dry-runs and confirms first; the route does not require it)

HTML page: GET /documents (session). The TUI Documents tab reads GET /api/documents and GET /api/documents/{document_id}/content from the running server.

Email (email_api.py)

Method	Path	Auth	Description
GET	/api/email/status	Session	Active provider, auth modes, presets summary
GET	/api/email/accounts	Session	Every mail account: source main / settings / env, name, type, address, auth (oauth / password / incomplete), calendar, removable
POST	/api/email/accounts	Session	Add another account that signs in with a password (alias, type, address, password, hosts); the login is tried first and a failing one is refused (400)
POST	/api/email/accounts/{alias}/remove	Session	Forget an account added in Settings (its tokens and password)
GET	/api/email/presets	Session	Gmail/Microsoft IMAP/POP host presets
POST	/api/email/protocol/connect	Session	Save IMAP/POP for gmail|microsoft|generic
POST	/api/email/protocol/disconnect	Session	Clear protocol + OAuth tokens for provider
POST	/api/email/gmail/connect	Session	Gmail app-password → IMAP
POST	/api/email/gmail/disconnect	Session	Clear Gmail creds
POST	/api/email/oauth/gmail/client	Session	Save Google OAuth client id/secret
GET	/api/email/oauth/gmail/start · start.json	Session	Browser OAuth redirect / JSON; ?account=<name> signs in another account under that name
GET	/api/email/oauth/gmail/callback	Open (OAuth)	Token exchange; redirects to Settings
POST	/api/email/oauth/microsoft/client	Session	Save Azure app id/secret
GET	/api/email/oauth/microsoft/start · start.json	Session	Browser OAuth; ?account=<name> for another account
GET	/api/email/oauth/microsoft/callback	Open (OAuth)	Token exchange
POST	/api/email/oauth/microsoft/device/start · …/poll	Session	Device-code fallback; body {"account": "<name>"} for another account
POST	/api/email/oauth/microsoft/disconnect	Session	Clear Microsoft mail: tokens, the sign-in’s address and scopes (Outlook Calendar keeps its own switch)
GET	/api/calendar/status	Session	Each calendar’s state and address, and the active provider
GET	/api/calendar/oauth/google/start · start.json	Session	Google Calendar sign-in (callback: the Gmail one)
GET	/api/calendar/oauth/microsoft/start.json	Session	Outlook Calendar sign-in for the calendar only (callback: the Microsoft one, purpose=calendar); mail is left as it is
POST	/api/calendar/oauth/microsoft/device/start	Session	The same by device code (poll with /api/email/oauth/microsoft/device/poll)
POST	/api/calendar/oauth/google/disconnect · …/microsoft/disconnect	Session	Turn that calendar off until it is connected from the card again; mail stays connected

Agent mail ops use tools (email_list, …), not these HTTP routes. Guide: Email integration.

X publisher

Method	Path	Auth	Description
GET	/x	Session	X Studio page (composer + X-only planner).
GET	/api/x/status	Session	Configured?, handle, caps. Never returns secrets.
POST	/api/x/preview	Session	Dry-run ToU policy (evaluate_post). No network, no ledger.
GET	/api/x/drafts	Session	Flattened save_proposal items for the Studio inbox.
GET	/api/x/audit	Session	Recent x_audit.db rows (?action=&limit=).
POST	/api/x/post	Session + CSRF	Immediate post (publish_x_post). Operator click is the approval. Optional reply_to_id, proposal_id (stored text wins).
POST	/api/x/delete	Session + CSRF	Delete a live tweet (delete_x_post). Operator click is the approval.
POST	/api/x/credentials	Session + CSRF	Save four OAuth 1.0a keys (vaulted) + handle + caps.
POST	/api/x/test	Session + CSRF	GET /2/users/me with stored keys.
POST	/api/x/disconnect	Session + CSRF	Delete keys, disable posting.
GET	/api/x/reply	Session	Auto-reply config + live poller status.
POST	/api/x/reply	Session + CSRF	Save auto-reply settings; starts/stops the mentions poller.
POST	/api/x/reply/preview	Session + CSRF	Dry-run a draft. Nothing published or recorded.
GET	/api/x/reply/conversations	Session	Summon log (parent + mention + draft/reason).
POST	/api/x/reply/poll	Session + CSRF	Fetch mentions from X now (ignores stuck cursor).
POST	/api/x/reply/approve	Session + CSRF	Publish a held (or wire-failed) draft.
POST	/api/x/reply/deny	Session + CSRF	Park a held draft.
POST	/api/x/reply/retry	Session + CSRF	Re-run a skipped/failed/drafting summon.
POST	/api/x/reply/delete	Session + CSRF	Delete the posted reply on X (if any) and drop the log row.

Chat tweets still go through x_post (always HITL + proposal_id). Guide: X publisher. Mentions: X auto-reply.

Scheduled (cron + X)

Method	Path	Auth	Description
GET	/scheduled	Session	Mixed clock page (cron jobs and X posts).
GET	/api/scheduled/tasks	Session	Every cron job (upcoming and finished, each with its status class and the actions the server allows) and the X publication store’s operations. Prompts are display summaries; the address is a destination label.
GET	/api/scheduled/cron/{job_id}	Session	One job with its full prompt and version, for the edit dialog.
POST	/api/scheduled/cron	Session + CSRF	Book a job: timing, prompt, optional kind (reminder/task) and capabilities (read/read_write). Caps, length limits and the scheduled rate bucket apply.
PUT	/api/scheduled/cron/{job_id}	Session + CSRF	Edit timing and/or prompt with the expected_version the form read: 200 rescheduled, 202 edit_pending (during a fire), 409 conflict / not_editable.
DELETE	/api/scheduled/cron/{job_id}	Session + CSRF	Cancel: 200 cancelled, 202 cancel_pending (during a fire), 409 for a finished job.
POST	/api/scheduled/cron/{job_id}/restore	Session + CSRF	Run again: book a finished job anew at timing.
POST	/api/scheduled/x	Session + CSRF	Book an X post (book_x_post). Optional reply_to_id, proposal_id. Answers the stored time and state.

X posts on this page are rescheduled and cancelled through POST /api/x/queue/{operation_id}/reschedule and .../cancel (with the revision the page read), the routes X Studio uses. Every mutating route checks the same origin (kazma_ui/same_origin.py). On an install with platform users every route answers 409 scheduled_single_operator. Behaviour: Scheduled reminders and tasks.

X Studio’s All clocks link is this page. The Studio planner itself is X-only.

Gateways & platforms

Method	Path	Auth	Description
Webhooks	Telegram/Discord/Slack callbacks	Adapter secrets	Inbound messages + HITL buttons
*	/api/gateway/*	Session	Gateway status/control from CLI

Other

Method	Path	Auth	Description
WS	Voice routes	Session	routes_voice_ws.py
*	Chaos routes	KAZMA_CHAOS_ENABLED	Dev only
Static	/, /chat, /settings, /x, /scheduled, …	Cookie/session	HTML pages

Exact route lists evolve with routers mounted in app.py. For extension points see API & Extension Points.

Auth model summary

Mode	When
Shared secret cookie / header	Single-operator
Opaque web session	Multi-user default
Platform RBAC (viewer/operator/admin)	SaaS APIs
OIDC PKCE	SSO login
API token / JWT	Programmatic (where enabled)

See Multi-user SaaS and Environment variables.