Early access
It knows how sure it is.
Exhibit 01 · one receipt
orders-db warn: volume /var/lib/postgres at 86 percent capacity
Bandwise runs your yes-or-no, pick-one and score decisions on TypeSafe's System One models. Every answer comes back with a band, a reason and a cost. You set the lines. Each band gets an action: act, send to a person, fall back to a default, or ask a larger LLM. Every run records what it cost and what it saved.
Install the CLI
npm install -g @bandwise/cliThe hosted cloud is in early access. The kit is free and open source today.
How a decision flows
Step [ 3 / 3 ]
Write it once in a versioned spec, with the answers it may give: yes or no, one option from a list, or a level on a rubric. A System One model such as Jev answers with a confidence. For a yes-or-no question, it is the probability of yes.
Your thresholds sort that number into High, Medium or Low, and each band maps to an action. In this example the yes line is 0.8 with a review margin of 0.1, so 0.74 lands between 0.7 and 0.8 and goes to a person. The lines live in the spec, so a change is a reviewed version.
Every run returns the same envelope: the answer, its band, the action and what the call cost against a larger LLM. The CLI, the HTTP API and the console all read this one shape.
$ npx @bandwise/cli run --local log-line-pager.spec.json line.json --json
{
"answers": {
"page": { "type": "noul", "noul": 0.74 }why
},
"decisions": {
"page": {
"value": true,
"band": "medium",band
"action": "review"
}
},
"cost": {
"systemOneCostUsd": 0.000017,cost
"counterfactualLlmCostUsd": 0.0007,
"savingsUsd": 0.000683
}
}
An excerpt with illustrative values: the confidence is not recorded model output. The costs come from our price book for a 400-token call, against Claude Haiku 4.5 at list prices. Local mode answers from fixtures, so use it to check your spec and wiring, not accuracy.
$0.042 per million input tokens
Jev's list price. Output is not billed. Read from the price book in our model registry, which holds the live value.
67.8% at about $0.0004 and 0.4 s per case
Jev on TypeSafe's 4-workflow benchmark. A vendor benchmark whose reference answers are the average of two frontier models, not human labels. Source: DataCamp.
Up to 194x faster, 445x cheaper
TypeSafe reports this against LLMs on its own workflow evaluations. Your numbers will differ. Source: Vercel blog.
From the Log-line pager template
Does this log line show a problem that a person on call should look at right now, because users, data, or money are affected or soon will be?
trueAt 0.8 · falseAt 0.2 · reviewMargin 0.1
At 0.8 and above, the answer is a confident yes and the set pages someone. At 0.2 and below, it is a confident no and the line is only logged. Anything between goes to a queue a person reads during working hours. An unsure model never wakes anyone up.
checkout-api error charge failed: provider returned 503 for 42 of the last 50 attempts
High0.94yesactsPage the person on call
orders-db warn volume /var/lib/postgres at 86 percent capacity
Medium0.74yesa personQueue for a person during working hours
session-cache warn redis connection reset, retrying (attempt 1 of 3); reconnected
High0.06noactsLog only, nobody is paged
The confidence values are illustrative, not recorded model output. The bands and actions are what the template's policy does with them.
Why a small model
Jev costs $0.042 per million input tokens, and output is not billed. It scores below the largest models. That is why the bands exist.
| Model | Score | Cost per case | Time per case |
|---|---|---|---|
| Jev | 67.8% | about $0.0004 | 0.4 s |
| GPT-5.6 Terra | 67.9% | $0.0304 | 10.1 s |
| Claude Opus 5 | 73.1% | $0.1761 | 37.8 s |
| GPT-5.6 Sol | 74.1% | $0.0836 | 23.3 s |
A vendor benchmark. The reference answers are the average of two frontier models, not human labels. Source: DataCamp.
Most answers in a workflow are easy. Let the small model take the ones it is sure of, and send the rest to a person or to one of those larger models. TypeSafe reports its models as up to 194x faster and 445x cheaper than LLMs on its own workflow evaluations (Vercel blog). That is TypeSafe's framing, and your numbers will differ.
A confidence of 0.8 does not mean the answer is right 80% of the time. It describes how spread out the answer is. So Bandwise measures accuracy on your own labeled decisions, per band, before a band is allowed to act.
New sets start in shadow: they log what they would have done and act on nothing. Controlled lets only the High band act. Full turns everything on. A person can pause a set at any stage.
Savings calculator
System One calls cost little enough that the number worth watching is how often you escalate to an LLM. Move the rate and watch which lane grows.
Estimated difference: $934.20 a month. Most of what you still pay is the escalation lane.
Templates
Each is a working spec, checked to run and not measured for accuracy.
Is this pull request safe to merge without a reviewer?
yes or no
Is this scanner finding reachable, and how bad is it?
yes or noscore
Who owns this error, and did the last release cause it?
pick onescoreyes or no
Does this log line need a person on call now?
yes or no
Which queue gets this email, and does it need a reply today?
pick oneyes or no
Does this email need the recipient soon?
yes or noscorepick one
How ready is this account to buy, and what next?
scorepick one
Should this event wake the agent now, later or never?
pick one
Does the agent still need this item for its task?
yes or no
Headless first
Every capability in the console is an operation you can also call from the command line, the HTTP API or an MCP client. No setting hides behind a screen.
# Run a template on your machine. No account, no key, no network.
$ npx @bandwise/cli run --local log-line-pager.spec.json line.json
# Same run as a shadow rollout, as JSON for a script or an agent.
$ npx @bandwise/cli run --local log-line-pager.spec.json line.json \
--rollout shadow --jsonIt prints each decision with its band and action, the System One cost and the estimated savings. Local mode answers from fixtures, so use it to check your spec and wiring, not accuracy.
Runs specs on your machine now. Commands to publish, roll out and run evals arrive with the cloud.
Call a published set by ID or slug from any backend. Every run returns the same envelope with bands, actions, cost and savings.
Let Claude Code or another agent read set health, draft changes and propose rollouts.
Agents propose, people approve. Publishing or widening a rollout waits for a human. Pausing and rolling back never wait.
Providers
Pick the route per organization and per set. Model keys stay on our servers and never reach a browser or a customer tool.
| Route | Status | Notes |
|---|---|---|
| TypeSafe, direct | Supported | The default route, with pinned model versions for controlled and full rollout. |
| Vercel AI Gateway | Supported | Sets on this route stay in shadow until the gateway offers a pinned model version. |
| OpenRouter | Planned | Proposed as a further route. Not available yet. |
The free kit
bandwise-kit is the open part of Bandwise: the spec format, the local runner, the CLI, the templates, and a Claude Code skill that reads your repository and suggests decisions worth handing to a System One model. It is licensed under Apache-2.0.
The bandwise command. Runs a spec on your machine with no key.
Spec format, compiler, confidence router, cost and savings math.
Nine ready question sets with example and borderline states.
Runs in Claude Code inside your repository and drafts specs for the decisions it finds. It sends nothing to any service.
band=medium score=0.74 cost=$0.000017
Early access
The hosted Bandwise adds versioned sets, rollout stages, review queues and a ledger of what every run cost and saved. We are starting with a small group of design partners. Pricing is not public yet.