Give a model a role, point it at your own data through the same tools and permission checks the rest of Overseer uses, and let it finish by proposing — on your hardware or on a frontier model, bounded on the card, watched step by step.
Ordered the way the questions arrive: what you allow, what you add, what starts it, what it knows before it starts, what it does, and what you decide.
Sixteen agents on this estate, each a role with its own brief, hue and bounds. An incident can page several specialists at once — the hosts team rules the machine in or out while the network team blames the path — and every one of them lands a draft on the same timeline. The eight below are illustrative fixtures.
INC-20260826-M7DG
Invidious (VaporCloud, port 3001) container exited
Establish whether the container stopped deliberately or crashed; state what could not be established. Start with get_briefing.
Overseerget_briefing 410 msget_entity_confidence 220 msrun_probe 1300 mspropose_hypothesis 60 msrecord_investigation 80 msINC-20260826-7K3Y
Invidious down — coverage and blast radius unconfirmed
Resolved without a post-incident review. Draft the review from the timeline; name what the handling missed.
get_briefing 380 msget_incident_context 240 msrecord_investigation 90 msINC-20260822-QA2M
PKVault Drop unreachable to external checks
Paged specialist for hosts and rack. Rule the host in or out before anyone blames the service.
get_briefing 400 msget_update_posture 90 msrun_probe 1100 msadd_incident_note 70 msINC-20260822-QA2M
PKVault Drop unreachable to external checks
Paged specialist for network. A check failing from one carrier only is a path fault.
get_briefing 390 msget_incident_impact 150 mspropose_hypothesis 60 msrecord_investigation 80 msINC-20260821-M136
Kiwix unreachable
Impact has not been stated. Map the blast radius from the graph and say who outside can tell.
get_briefing 420 msget_incident_impact 180 msadd_incident_note 70 msINC-20260822-EXM0
Home Assistant internal TCP check down
The service this touched has no page. Draft one from what the incident established.
get_briefing 400 mslist_undocumented 120 msdraft_documentation 95 msINC-20260821-CRHA
Kiwix unreachable (4th in 3 days)
This keeps happening. Decide whether it is one fault or several, and draft the problem record.
get_briefing 410 msget_known_errors 130 msget_entity_confidence 200 mspropose_hypothesis 60 msINC-20260821-TNR3
Flaresolverr state unverified on VaporGet
Paged specialist for incidents and actions. If exactly one allowlisted action fits, propose it — never run it.
get_briefing 400 msrun_probe 1200 msrecord_investigation 85 msA capability is a set of tools, and a tool is the only way an agent touches anything. There is no tool that resolves, publishes or pages — so there is no capability that could.
Where the model runs is a setting on the agent, not a different product and not a decision you make once for the whole estate. Run a documentarian locally and an investigator on a frontier model, if that is the trade you want.
Point Overseer at an Ollama server you already run and choose a model it has pulled. Nothing leaves your network, nothing is metered per token, and the estate data an agent reads never touches anybody else’s machine. This is the default, and the reason the feature exists in a self-hosted tool at all.
For the work a small local model genuinely struggles with — long chains of evidence, ambiguous causes, anything needing real judgement. The key lives in an environment variable and Overseer reads it at run time without being able to show it back to you. What the agent reads leaves your network, and the card says so every time.
If you already have Claude Code, an agent can run on it instead of buying API credits separately. A small runner stays on your host holding the login, because the Overseer container may be exposed to the internet and your subscription credentials should never be. It reaches your tools back through Overseer under a token you mint and can revoke — with the host’s own shell, files and web tools disallowed.
Every tool in this space now claims AI, which tells you nothing. The question worth asking is what it is able to do when it is wrong — and that is a question about boundaries, so these are Overseer’s.
There is no tool that resolves an incident, publishes a status update, writes to your documentation or pages anybody. The most an agent can do is produce a draft and mark it for your attention. That is a property of the tools it is handed, not an instruction it has been asked to obey.
Your grants are the ceiling and each agent narrows them again. A hosted agent reaches its tools through your own endpoint under a token with its own role, so the boundary is checked on the server on every call — rather than living in a prompt the model might talk itself out of.
A run stops after a set number of steps or a set amount of time, whichever comes first; a summariser gets fewer steps than an investigator, and every agent has a daily token budget it cannot exceed. The numbers are on the card, beside the run you are watching.
Agents separate what the evidence shows from what it does not, and state their unknowns rather than round them away. An investigation that quietly asserts a cause is worse than no investigation, because somebody will act on it.
For any one service it can ask exactly what the pages tell you: whether its death would be detected, whether its green checks are credible, whether the catalogue and the runtime disagree, whether this keeps happening — and, for a host, what patches are outstanding.
Inventory, monitoring, incidents, backups, patching and the graph involve no model and never have. The agent surface is off until you switch it on, and switching it back off leaves you a tool that was never waiting on one.
This is the state a fresh install is in: the endpoint switch unset, no capability granted, and an agent that will refuse rather than half-work. An empty queue and a switched-off subsystem look identical unless the interface is explicit, and only one of them is worth acting on.