Hivemind AX takes a goal, turns it into a plan, runs the plan with agents, tools and your documents, and records every agent action in a hash chain you can export and verify offline. It ships with 35 agent definitions, more than twenty models behind one gateway, MCP in both directions, an isolated computer for code, and human approval on the actions that need it. Registration is open, with a 14-day trial of Pro. This is the long version of what it does and where it stops.
Hivemind AX is available today. You can register at ai.hivemindax.com, start a 14-day trial of Pro and run your first mission this afternoon. Pro is $29 a month, Team is $149, Enterprise is priced by contract.
The short description has been on hivemindax.com for a while: agent work you can prove. This article is the long one. I want to describe the product the way we describe everything on this site, which means what it does, how it is built, what it costs and where it stops. Everything below comes from a review of the release build of 31 October 2026. When I say something is there, it is in the code that is running, not on a roadmap. Where a capability depends on a setting, a region or a plan, I say so.
What a mission is
Most agent products are a chat with tools attached. Hivemind AX is built around a different unit of work: the mission.
A mission starts from a goal in plain words, plus whatever you choose to give it: documents, a data area, a repository, connections to outside systems, a budget, a time limit, a model preference, a residency choice. From that point the mission is a durable object with a state. It carries the plan, the tasks and the agents assigned to them, every model call with its tokens, every tool call and its result, every human decision taken along the way, the files produced and the verification status of what came out. Nothing you hand to a mission is a permission by itself; the runtime checks what an agent may use at the moment it tries to use it.
The practical consequence is that a finished mission is not a transcript. It is a record you can hand to someone who was not there.
From goal to delivery
- Triage. Before anything is planned, AX decides whether the goal needs a plan at all. A self-contained request with no outside inputs and no outside effects, such as an arithmetic question or a short piece of text, gets a single model call with a tight token limit. Anything that mentions a document, a connected system, a repository or a change to something takes the full path. Small questions do not start a team of agents.
- Preview. Before you commit, you see what the system intends to do: the classification, the proposed plan and roles, the models that can serve it, a cost estimate, an expected duration, and anything that would block it, such as a connection you have not authorised or a computer that is not available in your region. The estimate is an estimate, not a quote.
- Plan. The planner breaks the goal into tasks with roles, dependencies and required outputs, and validates the result as a dependency graph. Plans are bounded on purpose: up to three productive tasks plus a synthesis step for a standard mission, up to eight plus synthesis for a complex one. A plan may not add work you did not ask for, like a security audit or a web search nobody mentioned. Ten templates, from write-report and code-review to security-audit and data-migration, give the planner structure, but every mission still goes through the planner.
- Execution. Tasks run in waves, in parallel where dependencies allow. Each agent gets the tools its role, its task and your workspace policy allow, and nothing else. What an agent says and what a tool returned are stored separately, so “I ran the tests” stays a sentence until a tool result backs it. Failures, interrupted attempts, incomplete answers, approval requests and missing inputs are distinct states, and retries are bounded.
- Questions along the way. An agent can flag an assumption and keep going, and you can answer inside the mission while it runs. When something essential is missing, the mission stops and asks. If your answer changes work already done, AX records the dependency and re-runs the affected tasks instead of presenting the old result as current.
- Delivery. A mission ends as completed, completed with errors, failed or cancelled. Finishing is not the same as succeeding, which is why the result carries three separate labels. More on that below.
Working time and waiting time are counted separately: minutes spent waiting for your approval do not consume the mission’s time budget. Execution runs on durable leases, so a deploy or a worker restart picks up from the persisted state rather than starting over.
Decides whether the goal needs a plan at all. Self-contained requests with no outside inputs or effects take a single model call; documents, connected systems, repositories or state changes take the full mission path.
Agents and models
AX ships with 35 agent definitions. Seven are generalists: software engineer, code reviewer, researcher, data analyst, solution architect, technical writer, writer. Twenty-eight are specialists, grouped around security and privacy (ISO 27001, SOC 2, Cyber Essentials Plus, GDPR data mapping, HIPAA, FCA, DPIA, the EU AI Act, pentest report analysis, third-party and supply-chain risk, business continuity), finance and risk (PSD2 SCA, AML, MiFID II suitability, DORA, operational risk, financial crime), healthcare (NHS DSPT, clinical data security, medical device cyber, patient data breaches, CQC standards), defence (MOD security assessment, classification levels) and legal (contract risk, eDiscovery review, regulatory change).
An agent definition is a role: a method, a set of capabilities, the tools it may ask for and the shape of its output. It is not a separate model trained for that name, and it does not make your organisation compliant with the framework in its title. It makes the analysis structured and repeatable. You can define your own agents per workspace; their instructions sit below the platform’s constraints and cannot grant themselves tools or permissions.
Underneath, one gateway routes every model call. Automatic routing in this release uses Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, plus Claude Fable 5.1 outside the EU pin. Beyond those, 22 models from 12 providers are selectable by explicit pin under the Global setting and 11 under the EU pin, among them Qwen, Mistral, DeepSeek, GLM, Nemotron, MiniMax, Nova and GPT-6 families; the full list is on the pricing page. Three routing preferences decide how the gateway chooses: max for the strongest models available, balanced for the default trade-off, efficient when cost matters more than headroom. Every call, whichever route it took, lands in a ledger with its model, its token usage and, where the rate is known, its cost. A cost the system cannot measure is recorded as unknown, not as zero.
Residency is a workspace setting. Platform data is hosted in the EU, in Amsterdam. Model inference can be pinned to EU regions (AWS Bedrock in Frankfurt) per workspace, and that pin is the default for workspaces in the EU, the EEA, the UK and Switzerland. Under the EU pin, inference calls go to European regions only; the edge network stays global, and the code sandbox, which runs elsewhere, does not run. “Global” means no geographic restriction, not a region called Global.
max · balanced · efficient unknown, never $0 Tools, connections and the mission computer
The built-in tools are deliberately few: read a workspace document, make an HTTP request, compute dates, compute arithmetic. The HTTP tool is not open internet. Destinations go through a network guard that rejects internal and special addresses and re-checks the resolved address at connection time, and each call is classified for approval by where it is going and what it does.
MCP works in both directions. Any MCP-compatible client can drive AX through its MCP server: nine tools, from creating a mission and uploading a document to reading results and pulling the audit trail, plus mission resources and three prompts. In the other direction, agents can use tools from external MCP servers. The catalogue starts with GitHub, Stripe, Neon, Notion, Atlassian, monday.com, Hugging Face, Cloudflare Docs and DeepWiki, and you can add your own servers; OAuth connections use PKCE and a checked destination. Every remote tool gets a permission of Auto, Ask or Never. Never wins over everything, and Auto does not let a server lower its own risk rating. Grants are re-read at call time, so a revoked tool stops at the next call, not at the end of the mission.
For work that needs a terminal, a mission can get a computer: an isolated environment, built on E2B, where agents run commands, write files, start processes and build projects. Two sizes, standard (2 vCPU, 2 GiB) and build (4 vCPU, 8 GiB). The image is pinned: Ubuntu 24.04, Node.js 24, Python 3.12, Git, build tools and a set of common packages. The computer has no internet access; npm and PyPI installs come from read-only local mirrors built from locked package lists, so a mission cannot pull arbitrary code from the network. Each task works as its own Unix user in its own clone, and work moves between tasks through explicit handoffs.
When an agent reports that the tests pass, AX checks. A separate clean environment imports the project bundle and runs the declared verification commands; the verdict comes from exit codes, not from the agent’s summary. Files are delivered with their provenance, up to 10 MiB per file and 50 MiB per mission, with a Git bundle for the whole project. Export to GitHub goes to a dedicated branch and opens a pull request, after a human approval. Previews of a running app go through a dedicated origin with its own session, not through an open port.
The computer is available under the Global residency setting, within each plan’s VM allowance. It does not run under the EU pin yet.
Node.js 24, Python 3.12, Git and build tools. npm and PyPI installs come exclusively from read-only local mirrors built from locked package lists. Each task runs as its own Unix user in its own clone.
When an agent claims tests pass, a separate clean environment imports the project bundle and runs the declared verification commands. Deliverables: ≤10 MiB/file, ≤50 MiB/mission + Git bundle. GitHub PR export requires human approval.
Any MCP-compatible client can create a mission, upload a workspace document, inspect mission state, read deliverables and pull the tamper-evident audit trail using workspace-scoped tokens.
The HTTP tool is not open internet: destinations pass a network guard that rejects internal/special IP ranges, re-verifies resolved addresses at connection time, and classifies every request for approval.
Your documents
Documents live in data areas inside your workspace. Supported formats are PDF, DOCX, XLSX, PPTX, CSV, TSV, TXT, Markdown and JSON, up to 20 MiB per file on Trial, Pro and Team and 1 GiB on Enterprise through a streaming upload path. Office files are read through their container format; macros, formulas and embedded objects are never executed. Scanned PDF pages go through OCR in English and Italian, up to about 16 scanned pages per upload. Accepted, read and indexed are three distinct states, and a mission cannot cite a document that has not finished indexing.
Retrieval works on passages of roughly 512 characters, embedded locally with a small open model rather than an external embedding API, and searched by cosine similarity inside PostgreSQL, filtered by workspace, data area and, when you choose them, specific documents. Every passage handed to an agent carries its source label, so a finding can be traced to the document and position it came from. A similarity score is not a truth score, and an empty result does not prove absence; the verification layer is built with that in mind.
Each document is encrypted with its own key, wrapped by a key derived from your workspace secret (AES-256-GCM, PBKDF2 with 310,000 iterations), and every read by a mission is written to an access log. File names, sizes and status stay readable, because the system needs them. Retention periods are defined per plan: 7 days on Trial, 90 on Pro, 365 on Team, by contract on Enterprise, with a 30-day grace period whenever a policy gets shorter. Workspace erasure walks every table and removes stored objects before it reports done. Account deletion needs re-authentication and cannot leave a workspace without an owner.
Who decides
Every action an agent wants to take is classified into one of four tiers. Tier 1 and 2 actions run on their own: reading a document, calling an approved endpoint, running a command inside an attested computer. Tier 3 needs one human decision. Tier 4 needs two different people, and a workspace with a single owner cannot approve it alone. An approval is bound to what was shown; if the request changes materially, the old approval does not carry over. Standing approvals let you pre-authorise defined actions for a mission or a workspace, and they are revocable. The halt control stops new actions across a workspace and reaches into queues, the model gateway and running computers. It does not undo an email already sent or a transaction already accepted by an external system. Nothing can.
An API token cannot approve anything. Tokens carry read or create capability, are scoped to a workspace, and are rejected on the oversight endpoints, so an integration can start a mission but cannot sign off on its own actions.
Accounts use server-side sessions with opaque tokens, Argon2id password hashing, a check against known breached passwords, TOTP, passkeys and recovery codes, with re-authentication for sensitive operations. Roles are owner, admin, operator, viewer and auditor; viewers and auditors are unlimited on every plan. SSO is not in this release.
Tenant isolation runs in layers: workspace-scoped queries, a tenant context on every transaction, and PostgreSQL row-level security on tenant tables, with an application role that cannot bypass it. Secrets for integrations live in a vault that injects them into outbound requests to allowlisted destinations; an agent never sees the credential itself.
External content, whether a document, a web page or a tool result, is treated as data and sanitised before it reaches a model. We do not claim immunity to prompt injection. The approval tiers, and the fact that agents cannot grant themselves tools, are what limit the damage when a sanitiser misses something.
What “verified” means here
A finished mission carries three separate answers, each of which can be yes, no or not evaluated:
- Delivered: did it produce a usable output?
- Grounded: is that output supported by sources that were actually read, fetched or produced? Agent prose, an operator’s assumption and a file the agent wrote itself do not count as evidence.
- Verified: did the checks run, over what scope, and did they pass?
The checks are deterministic where they can be. Six run against the evidence: the requested parts are present, quotes match the source text, cited identifiers exist in the sources, figures appear in the sources, arithmetic recomputes, and a claim that something is absent says what was searched. A model-based content review can run on top. A coordinator can approve a task, send it back (twice at most) or escalate, and a coordinator that cannot be reached escalates rather than approving by default. A quality score across accuracy, completeness, relevance, actionability, attribution and coherence, with a threshold of 70 out of 100, is anchored to the outcome, so a failed mission cannot score itself into looking good. Read that score as a signal, not as a probability of being right.
Output is usable, backed by documents or tool results actually read, and all deterministic checks pass over the declared scope.
Results come out as PDF, Markdown, DOCX or JSON reports in technical, executive and regulatory presentations, and findings export to CSV and to a JSON format that Jira can import. The Jira export is a file; it does not create tickets in your project.
The record
Every mission writes an audit chain: each event carries the hash of the event before it, authenticated per workspace. Alter an entry, remove one or break a link, and verification names the sequence number. A chain that verifies proves that what was written is intact and in order. It does not prove that nothing is missing, and the product page says so.
You can export the chain as a signed file: a Merkle tree over the canonical events, inclusion proofs, a digest that binds the root to the mission and its sequence range, and an RSA-SHA256 signature with the certificate attached. The verifier runs offline, in your browser at hivemindax.com/verify or from the command line, and answers verified, partial or failed. Partial means the export does not cover everything it claims to, and it is never promoted to verified. The verify page carries the fingerprint of our production signing key, so a reader can tell whether we signed an export or somebody else did. A valid signature with an embedded certificate proves consistency, not identity, which is why you can also pin a certificate obtained from us directly.
Content that must stay deletable is stored beside the chain, not inside it. The chain keeps a salted hash, so you can erase a document without rewriting history.
9f4a… prev: genesis 3c81… doc: salted_hash b72e… prev: 3c81… 6d10… prev: b72e… Day to day
The console at ai.hivemindax.com holds missions, computers, approvals, data areas, evidence, reports, model routing, usage and settings. Missions stream live over SSE and WebSocket, with the REST API as the source of truth when a connection drops. A mission can be shared by link as a read-only report, with an optional password and expiry. Findings carry comment threads and a status workflow from new to remediated, verified or accepted risk. Paid plans can schedule recurring missions by cron expression and time zone; each run goes through the same quotas, budget and approvals as a manual one. Notifications reach the console, email, Slack, signed webhooks and browser push.
If you would rather not use the console, there is a REST API with an OpenAPI specification, a TypeScript SDK, a CLI with mission and verify commands, and the MCP server described above. All of them use the same workspace-scoped tokens.
Plans
The trial gives you 14 days of Pro with $15 of usage. A card is required at signup, and the trial converts to Pro on day 14. The paid plans are priced in US dollars, monthly or annually, and each includes a monthly usage allowance equal to its price.
| Pro | Team | Enterprise | |
|---|---|---|---|
| Price | $29/month or $290/year | $149/month or $1,490/year | By contract |
| Usage included | $29 a month | $149 a month | By contract |
| Mission-running seats | 3 | 5 | By contract |
| Viewer and auditor seats | Unlimited | Unlimited | Unlimited |
| Hash chain, independent verification, signed export | Yes | Yes | Yes |
Usage is metered at the published rate card, per million tokens by model and route, plus computer time by the hour, and drawn from a wallet in US dollars. The plan’s allowance is spent first; beyond it you prepay credit, with optional auto-recharge under a monthly ceiling you set. Each mission has a spending cap. The amount is reserved when the mission starts and settled when it closes, so two missions cannot spend the same dollar. In this release a mission that fails, runs out of budget or ends without its verification being evaluated is not settled as a billable mission. Prices exclude VAT, which is added at checkout where we collect it; businesses can enter a VAT number for reverse charge.
Where it stops
- An agent named after a framework does not make you compliant with it. The specialists structure an analysis; the judgment stays with your people.
- Verification checks what can be checked: structure, quotes, figures, identifiers, arithmetic, the commands you declared. It does not establish that a conclusion is true.
- A verified chain proves integrity and order, not completeness.
- The mission computer has no internet access, installs only from our mirrors, and does not run under the EU pin yet.
- Replaying a mission reconstructs the recorded transitions. It does not reproduce the same model reasoning.
- The halt control cannot undo effects an external system has already accepted.
- Sanitisers reduce prompt injection; they do not eliminate it.
- SSO and seat expansion beyond the plan sizes are not in this release.
- We have not yet published a representative benchmark of mission quality. When we do, it will come with the losses, the way the HivemindEval release did.
If you run work that someone else will have to trust, this is the proposal: sign up, run one mission, export the evidence and hand it to that person. The trial takes a card, gives you $15 of usage and converts to Pro after 14 days. If you need a contract, a residency commitment or a procurement conversation, write to business@hypereum.tech.