Self-hosted AI agents, in production · installs in fifteen minutes
Rikoo
CommunityEnterprise

The Studio, module by module.

Everything a team needs to operate an agent is in the Community edition. What a company needs to operate a hundred of them, with a security officer and an auditor looking on, is in Enterprise. That is the whole rule.

34
Model providers
3 174
Models in the price registry
9
Services to operate
Air-gap
Supported deployment
Agents · Community

Versions you can point at, months later

System instructions with their history, a model chosen among your connected providers, ephemeral sub-agents that delegate with a subset of their parent's rights, never more, and reusable skills loaded on demand and frozen with the version.

  • Validated JSON output schemas, with bounded retry on drift
  • Input and output guardrails: scope, content, format
  • Image, PDF, Office and CSV attachments that announce truncation instead of staying silent
  • A playground on any version, with a live trace, without polluting production metrics
  • Test suites with evaluators, and a blocking publish gate below a pass threshold (Enterprise)

Agents and versions

Tools and MCP · Community

Allow-listed per server, then tool by tool

Remote MCP servers over HTTP, SSE and Streamable, with OAuth 2.1 and PKCE towards Notion, Linear, GitHub and others, and tokens sealed in the vault. Project HTTP tools are described in JSON Schema, authenticated by a vault secret, and unit-testable.

  • Three embedded MCP servers, for files, shell and browser, off by default
  • Each run gets its own confined space under a dedicated system identity; the shell has no network
  • A circuit breaker and a health probe per server
  • Large outputs offloaded to object storage and re-read on demand
  • Tool media carried as artifacts, never as bytes in a log

Browse the MCP library

The MCP Servers screen of the Studio: ten connected servers, each streamable and reporting its connection state.
Guardrails · Community

A guardrail is not a paragraph in the prompt.

Templates run outside the model, are pinned to the agent version, and cannot be talked out of their job by the conversation. They compose: one agent can carry a scope gate, a redactor and a semantic judge at once, in a defined order.

  • Personal data

    Detected on the way in and on the way out, then redacted, blocked or escalated depending on the template. What was redacted is recorded; the value itself never enters a log or a trace.

  • Semantic guard

    Judges the answer against a policy you write in plain language: no commitments the company cannot honour, no advice outside the licensed perimeter, no invented figures. A failed check retries once, then falls back to a safe response.

  • Format and citations

    A validated output schema with bounded retry on drift, and an optional rule that every factual claim carries a citation from the knowledge collection. An answer without a source is not delivered.

  • Every decision is visible

    Blocks, redactions, retries and escalations appear as steps in the run trace with the reason attached. A support lead can see why a customer got a refusal, and an auditor can count them.

  • Untrusted content is framed

    Attachments, retrieved extracts, tool output and server descriptions are framed as data before they reach the model. One implementation for the whole platform, deliberately not two.

  • Asserted before publishing

    Test suites assert on guardrail decisions, and a publication gate can block a version that fails them. Enterprise.

Models and knowledge · Community

Thirty-four providers, and your own documents

European providers are there too: Mistral, OVHcloud, Scaleway, Nebius and Nscale, alongside self-hosted vLLM and Ollama, and any OpenAI-compatible endpoint as a custom provider. Credentials go into the vault, the connection is tested, models are discovered and activated project by project.

A price and capability registry for 3,174 known models ships with the product. No network call at run time, and no cost silently reported as zero.

Knowledge is organised in collections: the unit of sharing and of permission. Hybrid vector and lexical search runs on PostgreSQL and pgvector with an HNSW index, citations are visible, and tightness between collections and between organizations is enforced inside the SQL query rather than filtered afterwards.

Knowledge collections

The LLM APIs screen of the Studio: five connected providers, each with the number of models it exposes and a sealed key.
Surfaces · Community

One brain, many surfaces

A versioned public API with read or full scoped keys, restrictable per agent, described by a published OpenAPI contract. A web widget in one script tag with enforced origins. Slack through a workspace token or OAuth, answering in thread.

Slack is in Community on purpose: an agent unreachable from the tool people actually work in is not a platform.

  • Zone-based widget design, icon catalogue, automatic contrast, restrained motion
  • Appearance changes live; what a visitor may ask is a decision of the version
  • In-page capabilities only when the operator enables them and the visitor consents
  • Verified visitor identity: your backend signs an assertion, the widget picks up the right thread

Widget and SDK

Monitor and administration · Community

What happened, what it cost, and who changed what.

  • Per-run traces

    Steps, model calls with tokens and cost, tool calls, cited extracts, parent-child delegations and authorization decisions. Usage and cost by organization, project, agent, surface and model, with CSV export.

  • Native OpenTelemetry

    OTLP export to your collector, be it Grafana, Datadog, Langfuse or Honeycomb, following the generative-AI conventions, with Prometheus metrics and logs correlated to the trace. Without a collector, nothing fails.

  • Forty-eight rules, watching

    The platform watches itself: restart loops, queue backlog, failure rates, exhausted provider billing, cost deviation, failed logins, blocked egress. A rule that changes severity announces itself in the Studio, by email, or to a webhook address you state. The twelve security rules cannot be switched off.

  • A dashboard, ready to import

    One JSON file, built from what the platform actually exposes rather than adapted from a template. Import it into your Grafana against Prometheus, VictoriaMetrics, Thanos or Mimir, and it draws. Nothing extra has to run for it to work.

  • Accounts and roles

    Local accounts with Argon2id, progressive lockout, two-factor with recovery codes, sessions listed and revocable. Invitations by email or at sign-up, usable without a mail server. A role per project and ownership transfer.

  • Secret vault and budgets

    Written once and never readable again, with atomic rotation, a rollback window and a "referenced by" graph. Budgets and quotas per organization and project, and a per-run cost ceiling carried by the agent version, one dollar unless you raise it.

Deployment · Community

Runs where you already answer to

One Compose file starts the whole platform on a single machine. A Helm chart deploys the same services on a cluster, with the stores either bundled or pointed at managed ones, and migrations running as a job before the application starts.

Three Terraform modules go further and provision the cluster too, on AWS, Google Cloud or Azure, with the managed PostgreSQL, Redis, object storage, load balancer, certificates, logs and encryption a production deployment needs on its first day. They deploy the same chart rather than a second definition of the platform.

  • Every deployment path is a repository you can read, fork and pin to a version
  • Managed PostgreSQL, Redis and object storage wherever you prefer them to the bundled ones
  • Air-gapped installs supported: the model registry ships with the product and is never fetched at run time
  • The same per-run sandbox isolation on a cluster as under Compose

Self-hosting

Positioning

Where Rikoo sits, and what the others do well.

CriterionRikooPersonal agentsClosed platformsWorkflow builders
Designed forAn organization exposing agents to customers and employeesOne developer or operator, on their own machineCompanies already living in the vendor ecosystemTeams assembling step chains
Where it runsYour site, a sovereign cloud, or air-gappedA workstation or a personal serverAt the vendor, under their jurisdictionSelf-hosted or hosted
Models34 providers, European and self-hosted included, with automatic fallbackSeveral providersTheir model, or a curated catalogueSeveral providers
Roles and tenancyOrganizations, projects and roles at both levelsSingle userTheir own model, often per seatWorkspaces
AuditabilityImmutable versions, frozen grants, append-only audit logDepends what you wroteWhat the vendor exposesRun history
Companies have agent prototypes, not agents in production. The gap is not the model. It is everything around it.
RikooWhy we built this
What is not shipped

Declared, not yet delivered.

Discord, WhatsApp and Telegram

Slack is shipped and in Community. The others are not, and a connection to them cannot be declared today. The ingress and egress path Slack uses was built to take a second protocol, so adding one is work rather than a redesign. Until it is done we would rather say so than list them as available.

A hosted offering

A hosted beta is open at beta.rikoo.io, run on our infrastructure, and Rikoo stays self-hostable in full. What is not there yet is the commercial offering around it: published plans, billing and a service commitment.

Named customer references

They are on the way. Until they are approved for publication, the quote on this page is a placeholder and is labelled as one.

Next

Install it and judge for yourself.

Nine services, one command, fifteen minutes. No account, no key, no call with us.