Tier Matrix Open-Core
Four separate pages each describe part of the boundary between what Intutic does with no account at all and what it does once a control plane is involved: Getting Started has the standalone architecture, Security has the Standalone-vs-Connected data-flow diagrams and the Enterprise-tier identity/compliance summary, Security & Identity has the full team-role table, and the FAQ has the standalone works/doesn't-work split. This page is the one place that states the whole boundary at once — everything below is drawn from those pages, not new content, so treat this as the index and follow the links for the full detail on any row.
There are three tiers, not two:
- Standalone (Open-Core) — no account, no control plane, runs entirely on your machine.
- Connected, Free — connected to a control plane, in a workspace with no paid plan.
- Connected, Paid — a workspace on a paid plan: Self-serve, Biz Org or Enterprise.
What runs 100% locally, with no account
Everything here works fully offline, today, in open core — nothing on this list is a trial or a degraded preview of a paid feature:
- Local rules and gates —
intutic gates installcompiles your.intutic/sops/markdown into each harness's rules file, such as.cursor/rules/intutic-governance.mdc,.claude/rules/intutic-governance.mdor the rules section ofAGENTS.md, and installs each harness's blocking gate;intutic startrewrites them as files change. - Local policy enforcement — the Rust proxy intercepts every LLM request/response, evaluates it against local WASM rules, blocks prohibited tool calls, and injects steering warnings into the stream in real time.
- Local cost and token ledger — session traces and daily spend land in local JSONL logs (
~/.intutic/logs/), whichintutic traces listreads with no account. Pre-flight estimates from recorded usage need a connected workspace; standalone,@intutic predictuses static per-model multipliers. - Local spend caps — a daily budget ceiling (
~/.intutic/config.json), enforced natively inside the proxy on every request. - Intelligent routing and the bandit — the routing engine and its cache-honest guard (see Intelligent Routing) run locally; arm state persists to
~/.intutic/bandit-state.json. - DLP scanning and prompt-injection detection — pattern-based, no LLM call, no network egress.
What a control plane adds
| Standalone | Connected | |
|---|---|---|
| Bandit arm state | ~/.intutic/bandit-state.json | bandit:{workspaceId} in Valkey — shared and durable |
| Who updates routing arms | this proxy, deterministically | the control plane's daily reward job (deterministic, no model call), once it claims the workspace |
| Feature flags | config.yaml | control plane, authoritative |
| Auth & budgets | local daily spend cap | workspace virtual keys; spend budgets per workspace, per virtual key and (Biz Org and up) per member, by UTC day and month, each one refusing requests or alerting only; requests-per-minute and tokens-per-minute limits per key |
| Response cache | per-process | shared across every proxy in the workspace |
| Team dashboard | — | web UI: spend, audits, compliance scores |
| Ticket board integration (Jira/Linear) | — | cost attribution to a ticket key |
| Corporate SOP distribution | local files only | centralized, tamper-proof, pushed to every developer |
| Guardrails from policy documents | hand-written SOP files and rules | Notion, Confluence, GitHub, Google Docs and uploads become cited, shadow-first guardrails a person promotes |
| LLM-as-judge (L2/L3 semantic checks) | — | on a paid plan or a trial, not Free: Free stays deterministic |
| MCP server governance | the proxy's local checks: DLP, injection scanning, anomaly detectors, WASM rules, TOFU pinning | plus the server registry with approvals and default-deny, approval holds, allowlists, the calling member on every event, call budgets and tool-change risk scoring, on every plan; the SSO group policy on plans with single sign-on |
| Decide rules | decide: lines in local SOPs, asked of your own self-hosted engine and pinned by intutic decide eval; enforced by the hook gates, the gate SDKs and the proxy | plus rules served from the control plane with their pins, and its shadow stamps |
| Decision traffic | the proxy governs decision-model calls: DLP, pinned model versions, abstention and human-review headers, the local decision-traffic log, its own daily cap | plus workspace and virtual-key budgets |
Qualified requires_before, scope_paths_mode: enforce | enforced by the hook gates, the gate SDKs and the proxy | the same |
plan_approval: | no plan can be approved, so every call but a read is refused (PLAN_NOT_APPROVED) | enforced against the session's approved plan |
review_before: and closed_lanes: holds | refused, not held: approval needs a control plane | held until a person approves the call |
Arm state carries over rather than resetting when you connect: a workspace that learned standalone keeps that learning, and the local loop stands down within 60 seconds of the control plane taking over. See Getting Started for the mechanics.
Nothing that requires a control plane was ever available in open core — so "what standalone gives up" is exactly the right-hand column above, never something silently downgraded. Provider credentials are held in memory only and never written to disk in either mode.
Upgrading
No CLI or proxy binary changes at any step — the progression is entirely config and billing state:
- Standalone —
intutic start, budget enforced against~/.intutic/config.json. - Connected, Free — register on the dashboard, then connect to it. The sync daemon uploads buffered local traces and starts syncing workspace rules immediately; no reinstall.
- Connected, Paid — an administrator sets limits and unlocks tier features from the dashboard; the next sync-daemon handshake picks them up.
Related
- Getting Started — standalone architecture, what Valkey adds, and the standalone/connected mode switch in detail
- Security — Standalone-vs-Connected data-flow diagrams, encryption, and the Enterprise-tier identity/compliance summary
- Security & Identity — the full team-role table, SSO setup, and API key management
- FAQ — the standalone works/doesn't-work split this page draws from