◈Enterprise & sovereign
Run it where it has to run
The default offering is hosted; the differentiator is that Mephit can run anywhere — your cloud, your datacentre, your closed network — on a model you control.
◈LOCAL LLM
Bring your own model
Run inference on a model you control. Plug in any OpenAI-compatible endpoint — Ollama, vLLM, llama.cpp server, TGI, or a private Anthropic / OpenAI / Bedrock / Azure deployment. Mythos uses your endpoint; nothing leaves your perimeter except what you ask it to.
◈SELF-HOSTED
Single-tenant on your infra
Docker Compose for staging, Kubernetes / Helm for production. Bring your own Postgres (14+), object storage, and reverse proxy. We ship the platform images, runbooks, and upgrade paths.
◈AIR-GAPPED
Disconnected mode
Everything off the public internet: local LLM, internal intel feeds, no telemetry, no auto-update. The platform refuses to dial out. Suitable for SCIF-style deployments and classified networks.
◈CUSTOM SKILLS
Wrap your internal tools
Expose your in-house APIs (sandbox, malware DB, asset inventory, EDR, SIEM) as skills. The router activates them alongside the built-ins. Source stays in your repo; we provide the SDK.
◈INTEL ADAPTERS
Swap providers per skill
Replace any public OSINT provider with your own: MISP, OpenCTI, internal threat-intel feeds, regional CERT data, classified sources. Drop-in adapters — same skill, your data.
◈IDENTITY · RBAC
SSO, SCIM, role-based access
SAML 2.0, OIDC, SCIM provisioning. Role-based access control over projects, servers, skills, and audit logs. Group-mapped permissions; auto-deprovision on offboard.
◈RESIDENCY
Choose your region
Cloud deployments come in EU-only, US-only, or country-specific configurations. Self-host gives you full control — pick the rack.
◈SLA · SUPPORT
Production-grade ops
99.9% uptime SLA, 4-hour critical-incident response, dedicated Slack / Teams channel, named technical account manager. Quarterly upgrade cadence with backport patches for two minor versions.
◈Local LLM
Your data never meets a model you don’t control.
Mephit treats the LLM as a swappable adapter. Point it at an OpenAI-compatible endpoint of your choosing — running on a machine you own, in a network you operate — and the entire inference path stays inside your perimeter. The platform still drives skills, tools, audit, and projects exactly the same way.
Combine with disabled OSINT adapters and you have a system that does nothing outbound — usable inside an air-gapped network.
config · llm.endpoint
yamlllm:
# any OpenAI-compatible endpoint
base_url: https://llm.internal.gov/v1
model: llama3.1-70b-instruct
api_key: ${LOCAL_LLM_KEY}
timeout: 120s
intel:
whitelist: [misp, opencti, internal-sandbox]
block_external: true
network:
egress_policy: deny-by-default
allow: [llm.internal.gov, misp.local]
