Jev (optional)
An optional, off-by-default third-party service that decides which tools, skills and context the agent sees — and what it sends.
Jev (TypeSafe System One) is an optional decision service added in v0.3.16. When it is on, the agent asks Jev small questions during a turn — which tools to offer, whether to load a skill, how much of a large tool result the model needs to see — and uses the answers to keep the model's context smaller and cheaper.
Jev is off by default, and off means no request is ever sent to it — the agent behaves exactly as it does without Jev. It is a separate service run by TypeSafe: you bring your own Jev API key and TypeSafe bills you for it, on top of your model provider.
Turning Jev on sends conversation data to a third party. Except on sensitive sites (below), page content and tool results go to the Jev endpoint you configure. NevoFlux asks you to confirm this the first time you switch Jev on. Read the privacy section before enabling it.
Turn it on
Open Settings → Jev (nevoflux://settings, Jev in the sidebar).
- Under Connection, check the Endpoint (default
https://api.typesafe.ai/v1/systemone) and paste your API key. - Click Test connection. It reports the median latency of a few requests and suggests a timeout; Use … ms applies it.
- Switch on Use Jev and confirm the notice.
Every change saves immediately to the [jev] section of
config.toml.
Decision points
Each decision point can be switched separately under Decision points:
| Switch | Default | What Jev decides |
|---|---|---|
| Choose the tools for each task | Off | Which tools (beyond the core set) the model is offered for this task. Off by default because a per-task tool list defeats prompt caching and cost more than it saved in testing. |
| Load a skill when the request clearly needs one | On | Whether to load one skill at the start of the turn (only above a confidence threshold). |
| Grade large tool results | On | How much of a large tool result the model sees now; the full result is kept and can be recalled later in the turn. |
| Rebuild earlier turns when it pays off | On | Whether to rebuild the conversation history sent to the model, when the token saving outweighs losing the prompt cache. |
| Ask before risky actions | Off | Whether a tool call that your execution tier would allow should still ask you first. It can only add a confirmation — it never approves anything the tier would not. |
If Jev is slow or unreachable, the agent falls back to its own built-in rules for that decision and carries on; a timeout never fails the turn. Hover a reply's token line in the sidebar to see how many Jev requests it made — it also says when most of them fell back.
Where it applies. Jev's context decisions only apply to providers the agent drives itself (cloud APIs and custom providers). With a provider that runs its own agent loop — Claude Code, Gemini CLI, Antigravity or OpenClaw — only Ask before risky actions applies; with an on-device model, Jev does not apply at all. The settings page says so under the switch.
What Jev sees
- Normal pages: page content and tool results relevant to the decision.
- Sensitive sites: only the domain, the page title and an element count — never
page text or tool results. A built-in list covers common banking, payment, mail,
medical and government sites (expand Built-in list to see it); add your own
under Sensitive sites, one domain per line (subdomains are included). Private
and loopback IP addresses, single-label hosts and intranet suffixes (
.local,.internal,.lan,.corp,.home.arpa,.intranet) are always treated as sensitive. - Content with no known page: MCP tool output and content from a tab that can no longer be identified get the sensitive-site treatment (metadata only).
- Content that is not from a page — your own messages, local file contents, shell output — is not filtered by the sensitive-site list.
Other safeguards: a remote endpoint must use HTTPS (plain http:// is only allowed
for a loopback address), redirects are not followed, and while an
on-device model
is active only a loopback endpoint can be reached.
[jev]
[jev]
enabled = false
endpoint = "https://api.typesafe.ai/v1/systemone"
api_key = "" # empty → NEVOFLUX_API_KEY_TYPESAFE env var
timeout_ms = 800 # 100–10000
model = "jev-latest"
sensitive_domains = [] # added to the built-in list
tools_k = 10 # tools Jev may pick beyond the core set
skill_threshold = 0.8 # load a skill only at this confidence or above
[jev.points]
tools = false
skills = true
visibility = true
rebuild = true
permissions = falseThe key can come from config.toml or from the NEVOFLUX_API_KEY_TYPESAFE
environment variable; like other keys, it is stored in plain text in config.toml if
you set it there.