Skip to content

Roadmap

What's built, and what isn't.

Candor is a project invariant: everything here is labelled by honest status, not by ambition. As of July 2026, Machinaut v0.3.0 is live at console.machinaut.ai on Lodestar v0.5.x — the explorer and platform (P0) are shipped; the governance console (approvals, calibration, SIEM, audit, team & roles) is live in beta; and v0.3.0 adds enterprise identity (OIDC SSO, SCIM, verified domains) and the Fleet health map, both in beta. SAML, managed key custody, one-click local signing, and the CLI / Desktop are still planned.

Status key:ShippedLive & usableBetaUsable, hardeningIn developmentCode built; not yet GAIn designDesigned; awaiting a dependencyPlannedSequenced, not started
The public free tier is P0 + P0.5 together, and both are live — P0 shipped, P0.5 in beta. The OSS dependencies cleared: the shipper (ADR-0014) landed in v0.4.0 and the approval channel (ADR-0015) merged, with the customer-proxy loop integration-tested end-to-end. The v0.3.0 enterprise surfaces (E1 / E2) sit on the paid tiers — you pay for the fleet lens and the identity controls, never for reach into your own data. What remains is beta-hardening and the planned depth below.
  1. P0

    Ingest · Explorer · Share

    Shipped

    The funnel and infra. Get local NDJSON off the box, store it tenant-scoped, and render the existing Lodestar chain/report behind a hosted, multi-tenant explorer with shareable links — plus the platform to run it.

    • session_ship@1 ingest endpoint (POST /v1/events)
    • Tenant-scoped storage + email/password & API-token auth
    • Hosted explorer over projectChain / renderReport; shareable session links
    • Retention & free-tier quotas, telemetry / metrics, Kubernetes deploy

    Gating dependency. The OSS shipper (ADR-0014) landed in Lodestar v0.4.0 and the dependency is pinned to v0.5.x. Shipped and running at console.machinaut.ai.

    Done when A solo dev runs lodestar ship and gets a hosted, shareable chain-report link in minutes; storage is tenant-scoped; auth works. Done.
  2. P0.5

    Approvals · Calibration · SIEM

    Beta

    The differentiator. A read-only explorer alone loses to Langfuse — the approval inbox is what makes this a trust layer. All three halves are live in beta; the OSS approval channel merged and the loop is integration-tested end-to-end.

    • Hosted approval inbox + customer-side verification (live loop, integration-tested)
    • Calibration dashboard, per-class ECE / Brier
    • Sentinel SIEM + Slack / webhook alerting (email alerting planned)

    Gating dependency. Cleared. The OSS ApprovalChannel HTTP transport (ADR-0015) merged and pendingApprovals graduated to -trace; the customer-proxy loop is integration-tested here (announce → resolve → fetch → customer-side verify → consume).

    Done when A held push-to-main is approved from a phone, the signature verifies customer-side, and the action proceeds. Working in beta.
  3. Console

    Team governance console (v0.2.0)

    Beta

    Make it a team product. The governance surface a team needs the day a second person is accountable for what the agents do — a second signer, a cross-session record, and exportable proof.

    • Overview dashboard across projects & teammates
    • Org-scoped audit log with CSV / NDJSON export
    • Team & roles — owner / member RBAC, email-bound invites
    • Policies — draft + impact-simulation (customer-signed; the cloud never signs)

    Gating dependency. Shipped in the v0.2.0 console as beta. Policies are unsigned-draft-only by design (the forgery boundary extends from approvals to policies — you sign the exported artifact locally). Drafting and impact-simulation ship as beta API endpoints today; the policies console surface is still planned, and the simulation is indicative — the authoritative dry-run runs locally against your kernel.

    Done when A team runs the loop with a second signer, an audit trail across everyone’s sessions, and roles — without self-approval. Live in beta.
  4. E2

    Enterprise identity (v0.3.0)

    Beta

    Who is allowed to sign in, and how fast they can be revoked. SSO is not “log in with Okta” — it is how an organisation governs who may become an approver, and how quickly that access disappears when someone leaves.

    • OIDC SSO — discovery-based config, JIT provisioning, IdP group → role mapping
    • Enforce-SSO with break-glass owners (never gates ingest or machine tokens)
    • SCIM 2.0 — automated provisioning; deprovision kills sessions + personal API tokens
    • Verified domains — DNS-TXT proof, optional auto-join on first SSO sign-in

    Gating dependency. Shipped in v0.3.0 as beta and live on the hosted console — SSO, SCIM, verified domains and Fleet are all switched on at console.machinaut.ai. Scope is deliberate: SSO governs console access, never signing authority — an SSO-granted owner still cannot forge an approval, because approver keys stay customer-held. SAML 2.0 is planned, not built (OIDC first). Auto-join is exact-domain-match only and requires your IdP to assert email_verified; SCIM deprovisioning revokes console access and personal tokens, but cannot rotate a customer-held approver key — that stays a customer-side operation.

    Done when An admin connects an IdP, a leaver is deprovisioned from the directory, and their console access and tokens die with it — without anyone touching an approver key. Working in beta.
  5. E1

    Fleet — the multi-org health map (v0.3.0)

    Beta

    One pane for the governor who runs agents across several orgs. Fleet unifies what you can already reach one org at a time — a worst-first health map of every agent you govern.

    • Cross-org roll-up: orgs, agents, sessions, held-rate, open alerts
    • Worst-first agent table — held-rate, calibration, status, last seen
    • 14-day activity trend; unattributed sessions surfaced, never silently merged
    • Adds zero reach beyond membership — a strict subset of what switch-org already shows you

    Gating dependency. Shipped in v0.3.0 as beta, on Team and Enterprise orgs. Health only: spend and drift are labelled planned in the product because they are not built — there is no usage/$ metering yet (it is gated on billing) and no drift time-series. We would rather ship the health map honestly than dress two empty columns up as features.

    Done when A platform lead owning five orgs opens one page and sees the worst-behaving agent across all of them, first. Working in beta.
  6. P1

    Identity & policy depth

    Planned

    Round out the policy write-path and the long tail of identity.

    • SAML 2.0; magic-link, password reset; per-owner MFA on break-glass
    • The policies console surface (draft + impact-simulation ship as beta API today)
    • Managed-keys toggle & one-click local signing
    • The Machinaut CLI / SDK; Machinaut Desktop (in progress — ADR-0001)

    Gating dependency. Managed key custody is surfaced in the console but refused by design — every tier defaults to customer-held keys, and no tier ships defaulted to managed. It stays planned until the per-tenant KMS work lands.

  7. P2

    Moat + breadth

    Planned

    Widen the surface once the core loop is proven.

    • Memory-firewall console (read-only first)
    • Registry hosting — OSS pack-format landed; gated on the pack schemas being declared stable
  8. P3

    Enterprise parity

    Planned

    What large customers require to adopt.

    • Residency & retention
    • SOC2 / ISO
    • On-prem / air-gapped
    • SLAs & support tiers

    Gating dependency. Customer-held approver keys are deliberately NOT here — they already ship on every tier, Free included. Enterprise parity is about the compliance and operational surface (residency, attestations, SLAs), not the forgery boundary, which is not an upsell.

Self-host with Lodestar OSS is unlimited and free, forever. The cloud free tier covers the solo loop and is never gated — consistent with the open-core posture. Pricing meters on agent runs / traces; no per-seat on self-serve.