History

Changelog

What changed in Azoth OS, newest first.

Stricter governance, chosen per workspace

Added

  • A workspace can be governed more strictly than its deployment: an owner or admin chooses observe, govern or enforce in Settings › Organization. A workspace only ever tightens — looser modes are shown greyed out — and each change is audited and sealed.
  • A use case classified as a prohibited practice under the AI Act is never served, in any governance mode: it can only be rejected, and approving a use case now requires its AI Act class.

Changed

  • Each decision a person takes on a four-eyes request — an approval, an exception, a waiver, a litigation hold released — is sealed with who asked and who decided, as pseudonyms. A request made with an API key names the key’s owner or its creator; a request whose author cannot be told is rejected, never approved.
  • When a workspace’s four-eyes setting cannot be read, approval is required.

Fixed

  • When the emergency stop of a use case cannot be read, its calls are refused in every governance mode, and the refusal is sealed. A use case whose record cannot be read keeps its rules and budgets, and is refused under enforce.

Dossiers that hold what they sign

Added

  • Each decision now seals when it was taken, by whom — a person or a key as a pseudonym, or the system — and under which trace, inside the hash of the chain.
  • The verifier 2.1.0, in Proof › Verification and in the browser page, checks that the time, the actor type and the trace of each entry of a dossier are the ones its seal holds.

Changed

  • A dossier verifies only when it is signed and its signature checks against the keys you pin with --keys: since the verifier 2.0.0, an unsigned dossier, or one whose signature only its own keys vouch for, fails. The signature also covers the journal’s head when the dossier was exported, the entries it lists and its filters.
  • Verify now checks again every entry from the first one kept, not only the new ones, and the chain of every workspace is checked again in full each week.

Security

  • The semantic cache follows the workspace’s data policy on every write and every lookup: nothing is stored for a workspace whose policy keeps no answers, and personal data is redacted before the words of a prompt are kept.
  • The content kept beside a decision — the inputs and outputs of agent tools included — is redacted at every depth and follows the data policy before it is queued. The sealed decision does not change.

The gateway’s time measured, and a status page with its history

Added

  • Answers of /api/v1/chat and /api/v1/chat/completions carry Server-Timing: the time the provider took, the time Azoth added and the total.
  • The status page keeps 90 days of history: what the platform’s own check saw every 5 minutes, the degraded and unavailable periods, and the windows nobody observed — never shown as operational.

Changed

  • A streamed answer must start, keep flowing and end within set limits. A stream cut midway is recorded as a failed call — no tokens, no cost, nothing sealed — and its budget reservation is given back.
  • A streamed answer carries no receipt — its headers leave before its decision is sealed — but carries X-Azoth-Trace-Id, which leads to its decision.

Fixed

  • A served call whose seal cannot be queued is now a visible proof gap, raised as an alert, instead of a line in a log.

A clearer workspace

Added

  • The home page shows the month’s calls, sealed refusals and spend, each figure a link to the screen it comes from.
  • The pending invitations of an organisation are listed in Settings › Organization, where they can be sent again or taken back.
  • Name your workspace’s security contact in Settings › Organization: Azoth’s notice of a breach of your data reaches them.

Changed

  • When role changes need four eyes, an invitation that grants a privileged role waits for another owner or admin too.
  • The person who asked for a decision is told who decides it, and how to invite someone when nobody else can.
  • Dates, numbers and amounts follow your language on every screen, the demos speak it, and Arabic gets its own typeface.
  • The sign-up page states its one password rule, in your language; a page that exists in English only says so in the language you chose.

Fixed

  • A dashboard section that fails keeps its header and shows one error state, with Try again and Contact support.

Capabilities you choose, and human attribution

Added

  • Human attribution: a personal API key makes every call its owner’s, and an owner or admin can allow a service key to name the person a call is for in X-Azoth-On-Behalf-Of. Each decision is sealed with that person’s pseudonym, never a name.
  • Settings › Access can require every call to name its person; a call that names nobody is then refused, sealed and explained. Attribution applies only to a workspace that uses it.
  • Where Azoth runs them, usage discovery, model change control, red-teaming and the grounding check stay off for a workspace until an owner or admin turns them on in Settings › Modules and retention; exception requests and the portfolio section of the monthly report can be turned off there.

Changed

  • Settings › Modules and retention is shorter: what is on, what a higher plan includes and what is not enabled here, each row opening its details.
  • A feature your plan does not include names the plan that does and links the pricing page, instead of a refusal that blames your role.

Fixed

  • Google sign-in and the password reset work again; a reset link opens in the browser that asked for it.
  • A provider key is saved as pasted — only what surrounds it is trimmed — and refused only when the provider calls it invalid.
  • Calls to Claude that carry system instructions are served again on both chat routes.
  • An API key name already taken is explained, instead of a server error.

Security

  • Deleting a workspace or erasing an account cancels its subscription first; nothing is deleted when the cancellation fails.
  • Only an owner sets up single sign-on, and pointing it at another identity provider withdraws the domain proof.
  • The signing secret of a webhook and the alert channels are read by the server only, and an agent is stopped or resumed only through the emergency stop.
  • A litigation hold blocks the deletion of its workspace, and erasing a person also erases the personal data kept beside their decisions.

Help that answers first, and a site AI assistants can read

Added

  • Every error code of the API and the dashboard is explained on the error codes page — what it means, its usual cause, what to do and who can do it — and an API error answer with a code carries doc_url, its link.
  • A refusal says what to do next: the screen where it is set, or who to ask.
  • A provider outage seen across Azoth shows as one banner in the dashboard, for the providers your workspace connected.
  • The assistant answers how-to questions, what changed, whether something is broken and what an error code means, from the documentation, the changelog, the state of the services and the error catalogue, with their pages.
  • One help page: search the documentation, ask the assistant, see the state of the services and contact support. A ticket joins the context Azoth computes — each item removable, never the content of a call or a key — and is answered by e-mail and on the help page, where your plan’s first-answer target is shown.
  • Azoth’s support reaches your data only with an owner’s consent: an access of 4 hours at most, revocable at any time, each step sealed in your journal.
  • The documentation, the pricing page and the Enterprise page have a URL per language, and each documentation page is also served as Markdown (/docs/<page>.md). /llms.txt and /pricing.json describe the site and the offer to AI assistants.
  • Examples anyone can check offline — a demo dossier, its keys and a receipt — are linked from the verification page. The verifier is licensed under Apache-2.0.

Plans by governed system, and yearly billing

Added

  • The pricing page presents each plan by its governed systems — a use case approved and in force, or an active agent bound to none — with its monthly volume of sealed decisions and its journal retention. Drafts are free and unlimited. These quotas are not enforced on workspaces yet.
  • Yearly billing for Starter and Growth, in the billing settings, where a yearly price is offered: a year is never billed at the monthly price.
  • The Enterprise page sets out its floor, its pilot and the sovereign option.

Changed

  • The pricing page shows the monthly and the yearly price of each plan side by side.

From sign-up to your first sealed decision

Added

  • Pick Free, Starter or Growth when you sign up. Starter starts a trial with no card once your e-mail is confirmed, and ends by itself if no payment method was added.
  • Get started is three steps: connect a provider, send your first call through the gateway, see its sealed decision.
  • Send my first call sends a real call through the gateway, with your rules applied, under a key made for that call alone and revoked as soon as it is over. The step shows the answer, its cost, its receipt and its sealed decision.
  • Where it is offered, the Azoth trial key lets a new workspace send a few short calls before connecting its own provider key — paid by Azoth, governed and sealed like any other call.
  • Test models come from the model catalog in force; GET /api/v1/models/defaults returns them.

Changed

  • Setting up counts as done once the gateway has served a call; until then the home page shows Finish setting up.
  • A decision opens in Proof from its call, and waits there while it is being sealed.

The AI footprint in your code repositories

Added

  • Read, repository by repository, what AI coding tools left in a GitHub repository: co-author trailers (declared), markers and configuration files. Only the hits are kept — an author as a pseudonym, never a name, an address or a message — and turning the reading off deletes everything read.
  • Freeze a repository’s findings into a sealed statement, exported as a signed, timestamped dossier that the offline verifier checks.
  • Repositories with an AI coding tool in them appear among the unmandated uses: writing code with AI is an AI use too.
  • azoth proxy and azoth hook install in @azothos/cli 0.2.0: an AI coding tool’s calls go through the gateway, and each commit names them in an Azoth-Receipt: trailer. Such a commit is shown as attached to the sealed call — never as proof that the code came from it.
  • Drafts for your lawyer: the notice to employees, the information of the works council and the register lines. Drafts, never legal advice.

Changed

  • Streamed answers now carry X-Azoth-Trace-Id.

A living, signed model catalog

Added

  • Prices come from a versioned model catalog — input, output, cached and batch prices, context window, capabilities, hosting zones, lifecycle — and each sealed cost carries the price that made it. A new price is a new version; a published version never changes.
  • A person publishes the prices: every week the providers’ listings propose new models, retirements and end dates, and Azoth’s team checks each price against its source before publishing it.
  • The catalog is signed outside the app. Self-hosted and air-gapped deployments import it signed, and the verifier checks the signature and computes every sealed cost again.
  • A mandate’s models: those ending within 30 days, a cheaper alternative from the same provider, a comparison on the mandate’s reference set, and a rule that holds a mandate to its models.
  • A model served without a price is sealed at zero — never a guess — and listed among the items waiting for a person.

Changed

  • Routing weighs the quality your workspace measured on its own reference cases, or stays neutral when nothing was measured; the quality floor now works on its real scale.

Public systems protected, and Azoth itself

Added

  • A system your site’s visitors reach is protected against cost exhaustion as soon as it is marked public: a visitor is slowed, challenged, then paused; an unusual volume caps the system’s answers, then pauses it. A refusal spends no model call.
  • An exhaustion incident is sealed at each change of tier and exported as a dossier the verifier checks; the cost it avoided is labelled as an estimate.
  • The SDK sends the visitor’s opaque id and passes a challenge: endUser, challengePassed, exhaustionChallengeOf.

Changed

  • A flood of refused calls no longer reaches the database: known keys and refused counters are kept in memory for a moment, and a revoked key is refused everywhere within 30 seconds.
  • Azoth’s own AI has a daily spend cap; past it, the assistant keeps answering its suggested questions without a model. Sign-up refuses disposable e-mail domains.
  • One story across the site and the documentation: mandate, rule, proof. Cost control follows from governed AI; it is not the promise.

Receipts, the applicable law and five evidence dossiers

Added

  • Every gateway answer carries a signed receipt (X-Azoth-Receipt), and each workspace’s public keys are published.
  • Three timestamp levels per workspace — standard, qualified, local — every 5 to 60 minutes.
  • The offline verifier also runs as one browser page, with no network access: Verify it yourself.
  • The applicable law at each call: EU, UK and US packs, the strictest rule applied, its reasoning sealed and decided again by the verifier; your own readings of the law, decided four eyes; the retention the law asks.
  • Five signed evidence dossiers — audit, incident, risk profile, litigation, regulator — and the litigation hold.
  • Compliance bundles export OSCAL Assessment Results, and governance webhooks push blocked decisions, derogations and proof breaks to your tools.
  • Regulatory changes: a new version of a law decided again on your last 30 days; a weekly summary composed without any model.
  • The assistant reads spend, footprint, model choice, compliance coverage and regulatory changes.

Mandates: who answers for each AI use

Added

  • A mandate names who answers for a system — its accountable person, a member of the workspace — and until when. A suspended or ended mandate is not served under enforce, and every transition is sealed.
  • One table of systems and agents, and one page per mandate with its activity, its proof and its own stop. A sub-agent never receives more than its parent.
  • The home page and the queue of what waits for a decision, shown to whoever may act on it.
  • One list of sealed decisions, each with its detail: rules, security verdicts, grounding check, routing, law.
  • Spend per mandate: sealed cost and footprint, the tightest budget, the realized savings.

Changed

  • Five entries in the navigation — Home, Mandates, Rules, Proof, Settings — and Settings in five sections; former addresses redirect.
  • The assistant is on by default, can be turned off per workspace, and answers its suggested questions without a model when none may answer.

The assistant, in the app, Slack and Teams

Added

  • Ask Azoth about your workspace from a side panel (⌘K): short answers that link to the screens their figures come from, every figure checked against what the tools returned. Read-only.
  • Drafts an owner or admin confirms on the act’s own screen — a rule, a stop, a mandate, a dossier. The assistant never acts by itself.
  • The same read tools over an API key and the MCP server.
  • Slack and Microsoft Teams: questions answered to you alone, and approvals and exceptions decided with a click by owners and admins linked with MFA — sealed.
  • Get started does the first steps in place: a key, a provider, a test call.

A clearer product

Added

  • A call the provider failed is a trace: Live calls says why, and the error rate counts it.
  • A budget for one mandate, which its calls pass on top of the workspace’s; an agent run counts toward the budgets of the systems it acts for.
  • PDF versions of the evidence, decision, incident and compliance dossiers.
  • A second workspace, within your plan, from where the workspace is shown.

Changed

  • A new visual charter, one page template and one type scale across the product and the site.
  • A page without data shows one clear empty state, and a failed load says so with a retry — never a wall of zeros.
  • Fewer tabs, one-line introductions and business words; on a phone, a labelled Menu and tables shown as cards.

Azoth’s own controls, and OIDC sign-in

Added

  • Every access by Azoth’s staff is sealed in a platform journal, with MFA required and any break-glass declared before it is used.
  • Production is deployed only from a green CI, and each deployment is sealed; every change is scanned for secrets, code issues, vulnerable dependencies and container flaws.
  • Backups are restored every month as a drill, with a report kept and sealed; the bundled database of a self-hosted deployment is backed up every night.
  • Quarterly access reviews, for a workspace and for Azoth’s team, and a monthly evidence package for auditors that the verifier checks.
  • Azoth’s own model calls are pinned, gated and sealed like any governed call.
  • OIDC sign-in per workspace, limited to its verified domain, and through the organisation’s OIDC provider on a self-hosted deployment.

Governed MCP tools, compliance bundles, red-teaming

Added

  • A governed MCP gateway: approved tools only, every tool call through the agent’s mandate and Azoth Guard, held for a person when it cannot be undone, sealed without content.
  • Every governance act — a rule published, a mandate approved, an exception decided — is sealed in the journal.
  • Compliance bundles for the EU AI Act, the NIST AI RMF and ISO/IEC 42001: each requirement covered from sealed evidence, with what is missing, in a bundle the verifier checks. Technical evidence, never a certification.
  • Red-teaming before deployment: versioned probes through the real gateway, deterministic oracles, worst-case cost caps. A high-risk system is approved only once its latest campaign held, or with a waiver decided four eyes.

Self-hosted and air-gapped

Added

  • One image for SaaS, self-hosted and air-gapped deployments; outside SaaS, the plan comes from a licence.
  • A complete Helm chart, with a bundled or external Supabase and Redis and secrets generated offline. Migrations stay a person’s act, with the runner inside the image.
  • E-mail through your own SMTP relay; SAML with Entra ID or ADFS, and a single logout that ends the sessions.
  • Air-gapped: the server reaches only the hosts allowed, and a model call to any other is refused and sealed. Models hosted inside the deployment go through the same gateway.
  • A signed image on a private registry and a signed offline bundle that verifies with openssl alone.

Ten features for the proof

Added

  • The sealed decision journal is on by default, and its chain is verified every hour as it grows.
  • Agents carry an accountable owner and a versioned mandate; every tool call is checked against it, and a step that needs a person pauses the run.
  • Trusted RFC 3161 timestamps on the dossiers and on each workspace’s chain head.
  • Decision dossiers gather every call of one business decision by an opaque reference, and a client vault keeps your content on your side.
  • Usage discovery: the AI uses no mandate covers, grouped from metadata only and validated by a person.
  • Refusals that explain themselves in five languages, with compliant options and time-limited exceptions decided by the rule’s owner.
  • A grounding check on answers, proportionate to the stakes — always an estimate, shown with its judge’s measured reliability.
  • Incident mode: declare a faulty scope, find exactly the decisions it touched, contain it in one click, prepare a report draft that is never sent.
  • Model change control: when a model version changes, your reference set runs again on both versions within worst-case cost caps; switching stays a person’s decision.
  • A value and risk portfolio per mandate: measured cost beside the declared value, never summed.

A first run that works

Added

  • A guided first run on the gateway path, and a governed test call sent from the dashboard, with no code.
  • Rule templates with plain-language help; a rule can target a declared system.
  • An approvals screen for four eyes, where the reviewer sees exactly what they approve.
  • A view of the sealed journal: find a refused call, see why, verify the chain.
  • A help and glossary page for the words a beginner meets.

Changed

  • A full audit closed across security, database, operations, accessibility and translations; a broken deployment now rolls back on its own after its smoke test.

Display currency and team invitations

Added

  • Costs shown in one display currency across the dashboard, converted with a daily exchange rate, and payment in EUR or USD at checkout.
  • Team invitations show whether the e-mail was delivered and can be sent again; a test e-mail reports the provider’s exact error.

Analytics with consent, and the home page

Added

  • Product analytics hosted in the EU and loaded only after consent — off by default.
  • Contact capture and a chat widget, also loaded only after consent.
  • The home page answers four questions and says what Azoth is and is not.

Every decision sealed, refusals included

Added

  • Every refusal leaves a sealed, hash-chained trace; a refusal that cannot be sealed shows as a proof gap, never dropped in silence.
  • Every decision is bound to a declared use case and its version, with the rule set and the governance profile in force.
  • The use-case register moved to the dashboard: a different person reviews each declaration, an edit makes a new version, and a prohibited practice needs a documented exception.
  • What ran is compared with what was approved; a blocking gap revokes the approval, and the gap is sealed.
  • One switch for the posture — GOVERNANCE_PROFILE: observe, govern or enforce — observe by default.
  • Human approval for real: a held call is kept encrypted, approved by another person, then run once.
0.13.0

Exportable evidence, PolicyPacks, the MCP server

Added

  • An exportable evidence dossier: the sealed decision journal as a ZIP — the evidence, a manifest of content hashes, a readable summary — verifiable offline by a tool that shares no code with Azoth. Technical evidence that supports obligations, not a certification.
  • Regulatory PolicyPacks map obligations to policy rules. An applied pack lands as draft rules, not enforced until reviewed, and an obligation with no automatable rule is listed as a gap.
  • The Azoth MCP server lets an agent consult the governance layer within its own loop, read-only.
  • A GreenOps footprint API — avoided-impact estimates.
  • Navigation by hub: Decide, Block, Prove, Costs.

Fixed

  • The theme and language scripts run again under a strict content security policy.
  • Accessibility: a main landmark and correct heading levels on the public pages; a horizontal overflow on the enterprise page on phones.
0.12.0

Decision trace, avoided CO₂, outbound hardening

Added

  • A sealed, per-workspace hash-chained journal of every inference and agent decision, exportable as a tamper-evident evidence pack; personal content can be erased without breaking the chain.
  • An auditable record of the CO₂ avoided by each decision that avoids emissions, under a conservative baseline.

Changed

  • Hardened Kubernetes manifests, static analysis on every pull request and stricter typing across the API.

Security

  • Outbound requests to customer-supplied addresses are pinned to a validated public address.
  • Tighter rate limits on security settings.
0.11.0

Reliability, GDPR self-service, accessibility

Added

  • GDPR self-service: export your workspace data, request erasure, read what data we hold, review your consents.
  • Cookie consent per purpose, tied to a version of the policy.
  • An accessibility statement: a WCAG 2.2 self-assessment, with the known gaps listed.
  • Breadcrumbs, and consistent empty, loading and error states across the dashboard.

Changed

  • Retention now bounds every trace table, and heavy reports can read from a replica.
  • Automated WCAG 2.2 AA checks, and secondary text at the highest contrast level.

Fixed

  • The nightly retention purge, and a failure of the production build.
0.10.0

Operator tools and plan cleanup

Added

  • Diagnostics for the signed-in session and for the creation of an API key.
  • A workspace picker when you belong to several workspaces.

Changed

  • The plan lineup simplified to four plans; legacy plans keep working.
  • Requests are accepted from both the www and the bare domain.

Fixed

  • API-key creation failing on a www / bare-domain mismatch, accepted invitations that stayed pending, and billing management for plans set by hand.
0.9.0

Sign-up and interface

Added

  • A plan picker on the sign-up page, translated rate-limit messages, working footer links and nine provider connectors on the home page.
0.8.0

Documentation and company pages

Added

  • The documentation section, and the blog, careers, contact and status pages.
0.7.0

Inventory Object and workers

Added

  • The Inventory Object connector: catalog matching and vision extraction under the data policy.
  • Ingest split into a fast path and a worker that applies PII redaction and the data policy, with a retry safety net.
0.6.0

The GitHub App

Added

  • The GitHub App: each pull request gets a check run from its regression suite, with each case’s result, and a GitHub Actions template.
0.5.0

Platform and analytics

Added

  • Cost per agent, a 14-day forecast and a comparison with reference workloads.
  • Organisations with several workspaces, and four roles: owner, admin, member, analyst.
  • API v1 with Bearer keys, and the GDPR data export.
0.4.0

The first runtime engines

Added

  • Model routing, the semantic cache, prompt optimisation, loop termination and spend analytics.
0.3.0

SDK and integrations

Added

  • @azothos/sdk, with OpenTelemetry-compatible wrappers for OpenAI, Anthropic and AWS Bedrock.
  • Slack and PagerDuty alerts, and signed webhooks retried three times.
0.2.0

Security

Added

  • Two-factor authentication (TOTP), e-mail confirmation, a password strength meter and an append-only audit log.
0.1.0

A new brand

Changed

  • A new home page, navigation and plan names, in five languages.
0.0.1

The first version

Added

  • Sign-in with row-level security, a multi-tenant dashboard, five languages, light and dark themes, billing and security headers.