Every tenant gets an isolated runtime with its own agent team and memory. What makes the platform compound is a shared, privacy-scrubbed learning layer that all those isolated agents read from — so every server’s lessons make every other server’s agents smarter.
Every paid plan deploys the full team into your own runtime. You’re not buying capability unlocks, only scale. The Orchestrator reads intent and hands each job to the right specialist, and every action is metered and capped.
Your runtime’s main loop. Reads every Discord/Telegram message and event, then answers directly, delegates to a specialist, or provisions a durable cron job. Owns your per-tenant memory and skills.
Operates the server end to end: crash-diff analysis, power, file/config edits, and modpack install/update. Power and file mutations are gated by the trust-gate and your trust tier.
Answers your players’ and staff questions from your own docs, rules, and chat history. Read-only and cheap — retrieval plus a small model — escalating to the Orchestrator only rarely.
Watches chat and enforces your rules through a cost-safe pipeline: a deterministic pre-filter (zero tokens), then a small triage model, then an Orchestrator escalation only when needed.
Builds mods, plugins, and integrations for you — including crash-fix patches — compiled and boot-tested in an isolated scratch sandbox, then delivered via a signed link or pushed to your server on your approval.
The Development Team’s read-only second pair of eyes: adversarially reviews the Developer’s code and crash-fix patches for correctness, concurrency, loader/version compatibility, supply-chain provenance and security — then escalates findings. It reviews and reports; it never ships or mutates.
Plus the platform agents Tecton runs for everyone
Onboarding
The signup → connect bot → add server → pick agents wizard. Tecton runs it once for everyone; you never deploy or pay per-action for it.
Billing & metering
Meters every action, enforces your caps, and holds the pay-as-you-go hard-stop. Control-plane infrastructure, not an agent with a face.
The Researcher
The platform curator behind the unified learning layer — it scrubs and promotes generalized lessons into the shared knowledge base and skills. See the learning-layer section below.
Credits track real usage. Each request is billed by the tokens the agent actually consumes to do the work — a quick chat costs a little, a deep crash-diff or build costs more. You pay for work done, not per seat.
A credit is pinned to a fixed slice of baseline-model usage, so its value stays predictable. Heavier models cost proportionally more credits because their tokens genuinely cost more — the rate is derived from real token cost, never a number someone made up.
Run each agent on the model that fits its job — a light model for Support, a heavyweight for crash-diffs. You pick per agent, and credits reflect what each model’s tokens actually cost.
Every paid tier carries a credit cap and a hard-stop, so a runaway agent can never overspend or surprise you. Credits track real model cost; temporary NVMe scratch is included from Starter up.
This is how the platform compounds. Isolated tenant agents contribute lessons up through a privacy scrub, and read a shared knowledge base and skills library back down. Every server’s lessons make every other server’s agents smarter. The shared knowledge base and the Researcher that curates it are the learning system’s design; the ingest and curator pipeline are being built out — we don’t claim a pre-indexed catalog of mods as live.
A curated, tenant-agnostic store of facts, crash-signature → cause → fix entries, light how-tos, and public mod/modpack metadata. Any agent, on any tenant, reads it directly — deduped so each lesson is stored once and referenced by all.
Versioned, industry-grade how-to procedures any agent can load and reuse. The agent that solved a Forge mixin conflict once turns that fix into a skill the whole fleet can run instantly.
When an agent learns something worth sharing, it lands in a tenant-private, row-level-secured queue first. Identifiers, paths, IPs, and world names are stripped and the payload is generalized to a signature before anything is promoted. Tenant-private data never enters the shared layer.
A platform curator (the Researcher) is the only bridge from the private queue to the shared brain. It scrubs, generalizes, dedupes, and promotes — running with cross-tenant read only for that transform, never exposing raw tenant data.
Sharing knowledge and sharing data are not the same thing. Your raw data — servers, secrets, chat, memory — stays sealed inside two rings of isolation. Only generalized, scrubbed lessons ever leave, and only through the Researcher.
Each tenant runs its own Hermes process under an unprivileged system user with its own home directory. One tenant, one runtime, one blast radius — zero access to anyone else’s servers or secrets.
Inside that runtime, every action is checked against an allowlist you control — observe, propose, or act. And at the database, Row-Level Security pins the active tenant so a query can only ever see that tenant’s rows.
Run on managed keys with token-priced credits. Whatever you choose, keys are envelope-encrypted — we can’t read them and neither can another tenant.
Agents run on platform-managed keys with token-priced credits. Pick a model per agent — heavier models spend more, lighter ones less.
Point an agent at any OpenAI- or Anthropic-compatible endpoint — NVIDIA NIM, OpenRouter, your own gateway. Keys are envelope-encrypted; the URL is SSRF-validated against internal and metadata addresses.
Managed plans from $9/mo. Full agent team on every paid tier. Cancel anytime.