Memory that supersedes
A customer moves from Growth to Enterprise. Both facts were true once; one is true now.
Three engines benchmarked.
All three quoted the cancelled plan.
The difference
Everyone else watches. Nixet runs it.
Tracing tools attach to an agent you host and hand back a waterfall. The memory is still yours to build, so is the retrieval, the tenant boundary, the version history and the machine it runs on.
Traces that name the culprit
Tool arguments, results, tokens per turn, and where each tool came from.
33 event types · 14 categories
Cost per customer
Priced per call, summed onto the run, the conversation and the user.
Six decimals · avg $0.00215
Tools, with policy on your side
A tool's tags are the server's claim about itself. The MCP specification is explicit that a server can declare readOnlyHint: true and delete your files anyway — so tags decide nothing here. Nixet resolves your policy before the model is ever offered the tool.
Deny is deny, whatever the server says.
Knowledge bases
Twelve formats, converted locally in Rust, embedded on a worker you can watch.
Memory
Most memory only ever adds.
Nothing marks the old fact as no longer true, so the agent picks whichever the retriever ranked highest — and quotes the cancelled plan back at the customer paying for the new one.
Append-only store2 conflict
plan = "Growth"active
plan = "Enterprise"active
seats = 12active
The agent saysYou're on the Growth plan, so that feature isn't included.
Nixet memoryresolved
plan = "Growth"superseded
plan = "Enterprise"current
seats = 12current
The agent saysYou're on Enterprise, so that's included on all 12 seats.
Control
An agent you can tell no.
A model that can call a tool will eventually call the wrong one. Every mechanism here refuses something before it happens, on the Nixet side, where a prompt cannot argue with it.
Tool policy
enforcedA server's tags are its own claim. The MCP specification says plainly that it can declare readOnlyHint: true and delete your files anyway, so the policy you wrote is what runs.
tool.denied zendesk.delete_ticket · policy: deny
resolved before the model was offered the tool
resolved before the model was offered the tool
Redaction
enforcedTwenty-two detection labels across four modes. The provider receives a token where a name was, and the answer is restored on the way back out.
input [GIVEN_NAME_1] at [BUILDING_NUMBER_1]
output restored before it reaches your customer
output restored before it reaches your customer
Egress
enforcedA customer's URL is resolved to an address and checked against private ranges — then checked again on every redirect, because a public link that 302s to the metadata endpoint is the whole technique.
refused 169.254.169.254 · private or reserved
re-checked at hop 2 of 5
re-checked at hop 2 of 5
Spend
enforcedA workspace that has used its allowance is refused at admission, before the model is paid for. Runs already executing finish; nothing is killed mid-flight.
402 used all 1,000 runs in the Free plan
resets 2026-11-12 · upgrading lifts it now
resets 2026-11-12 · upgrading lifts it now
Approvals
buildingA tool tagged destructive pauses the run and waits for a person. Today that policy is refused at publish rather than stored, because an agent that cannot stop should not claim it will.
tool.policy ask → refused at publish
until a run can wait for a human
until a run can wait for a human
Evaluations
buildingRun a published version against a dataset in a sandbox, score it, and compare it to the version it replaces — so a prompt change is a measurement rather than a hope.
v3 → v4 · 240 cases · sandboxed
blocked on a regression, not merged
blocked on a regression, not merged
Trust
Agents that can handle regulated data.
An agent reads whatever your customer typed, and most of what they type is personal. Nixet masks it before the model provider sees it, keeps the original encrypted, and records who looked.
PII redactionlive
22 detection labels across four modes. The provider receives a token where a name was.
Encryption at restlive
Masked values go to an AES vault keyed per deployment. Postgres never holds the plaintext.
Tenant isolationlive
Memory and knowledge are keyed to your customer. A partial key addresses nothing, not more.
Audit traillive
Key issuance, revocation and every reveal of a masked value, with the member who did it.
GDPRdpa ready
EU region, a signed DPA, and deletion that reaches the vault and the vector store.
SOC 2 Type IIin progress
Audit underway. We will say so here when the report exists, and not before.
What the provider actually receives
CustomerHi, I'm Dana Reyes at 14 Hanover Street. My card ending 4412 was charged twice.
ProviderHi, I'm [GIVEN_NAME_1] [SURNAME_1] at [BUILDING_NUMBER_1] [STREET_NAME_1]. My card ending [CREDIT_CARD_1] was charged twice.
CustomerYes, Dana — we have refunded the duplicate charge on the card ending 4412.
The agent reasons over tokens and the answer is restored on the way out, so the customer reads their own name and the provider never saw it.
Pricing
Billed per run. Never per seat.
A run is one question answered, one ticket triaged, one document processed — a number already visible in your own product. No per-seat fee, no per-connection fee, no charge for the second person to look at a failure.
* SOC 2 Type II audit is in progress. Redaction, the encrypted vault, tenant isolation and the audit trail ship today on every plan.
Unlimited seats, including Free
The console is where you go when something went wrong. Charging for the colleague you show the trace to is charging for incident response, so we don't.
Your own key, no markup, ever
Point Nixet at your provider key and pay the platform fee, nothing else. If you'd rather we ran the gateway, managed inference is +15% on Pro and +8% on Scale.
Free
$0
A month of real development at the measured rate.
Start free- Runs1,000
- Retention7 days
- Agents3
- Seatsunlimited
- Redactionall 4 modes
Pro
$29 /mo
The plan a team of two is already past the point of.
Start building- Runs25,000
- Overage$2 / 1k
- Retention90 days
- Seatsunlimited
- Redactionall 4 modes
Scale
$249 /mo
Half a million runs, a year of traces, an SLA to point at.
Start building- Runs500,000
- Overage$1.20 / 1k
- Retention1 year
- Seatsunlimited
- ComplianceDPA · SOC 2*
Enterprise
Custom
A committed rate, SSO, and an engineer who knows you.
Talk to us- Runsnegotiated
- Retentionconfigurable
- Seatsunlimited + SSO
- ComplianceDPA · HIPAA
Start
Your agent already works.
Give it somewhere to run.
Create an agent, publish a version, and watch the first run land in the console.
1,000 runs free · no card