Tendwell
A self-hostable agent that monitors production health, retrieves the relevant runbooks and history, and explains what it finds - in plain language. Local-first by default: your data never leaves your infrastructure.
What Tendwell does
Four pluggable layers behind stable interfaces - data sources, LLM backend, knowledge store, and the action surface - so you swap any part without touching the core.
Local-first by default
The default configuration sends nothing off-host - including the LLM and embeddings. Point a backend at a remote address and Tendwell warns you explicitly at startup. No egress unless you opt in.
Any OpenAI-compatible LLM
One client targets a configurable base_url: Ollama, vLLM, llama.cpp, LocalAI, or a LiteLLM proxy. Weak at native tool calls? A prompt-based ReAct fallback keeps it working.
Pluggable data sources
Prometheus, Loki, and a generic HTTP/JSON adapter today, normalized into one result shape. A failing source degrades an SLO to unknown - it never crashes the run.
Knowledge-grounded findings
Runbooks and postmortems are embedded locally and retrieved by relevance. Citations come only from retrieved chunks, so the model cannot invent a source.
Human-gated actions
The model can only propose. Deterministic validation and a human approval gate sit between a proposal and any execution. Partial failure is first-class; dry-run plans without touching anything.
Tamper-evident audit
Every step in an action's life is an append-only, hash-chained event. A removed or altered entry is detectable. Audit cannot be disabled - it is the evidence a regulated buyer needs.
The security model is the product
For teams in security-conscious and regulated environments, the separation between proposing and executing is not a footnote - it is the reason Tendwell is trustworthy near production.
Read-only out of the box, and the open core ships no real executor: mutation is opt-in and bring-your-own. With no action surface configured, the agent is structurally unable to change anything - that property is enforced, not promised.
See it work in minutes
The demo stack runs against a synthetic source with zero real infrastructure, so you get a real health report on the first run.
git clone https://github.com/bmoldo/tendwell
cd tendwell
docker compose up # instant tier: synthetic source, no model downloadA second compose profile adds a small local model for genuine model-driven output. The one-time model pull is a few hundred MB - honest about the cost.
Built for high-stakes environments
Self-hosted, security-conscious, and regulated teams who need agentic operations they can actually trust - and audit.
Tendwell is free and Apache-2.0. If you want it run - or governed AI operations designed - in your environment, that is what ReOps Tech does.