PyxGrant / Solutions

The same verdict, shaped to the way your team uses AI.

Each solution below is a policy pack and a set of connectors on the same control plane. Nothing is a separate product with a separate console.

Book a demoBrowse solutions
one control plane, every teamlive
Coding agents

Let developers run agents in YOLO mode without handing them the keys.

Cursor, Claude Code, Copilot, and CI get the same checkpoint. The checkpoint trusts the working tree it hashed, not the commit pin the agent reported.

Secret paths. ~/.ssh, .env, and cloud credentials are denied, even with YOLO on.
Trust handoffs. Writes to .claude hooks, .mcp.json, fsmonitor, or the docker socket are refused.
Commit-shaped branches. A branch named like a 40-character commit is not accepted as a pin.
cursor-sam · eng-mbp-114YOLO on
ALLOW  npm test                        R-PERMIT
DENY   cat ~/.ssh/id_rsa               R-SECRET-PATH
DENY   rm -rf /var/backups             R-BACKUP
DENY   cp payload .claude/hooks.json   R-EXEC-CONFIG
DENY   tree digest ≠ advertised pin    R-TREE
AI apps and agents

LLM, MCP, and A2A traffic under one policy.

Your own agents call the decide API or run behind the MCP guard. Peer agents must present a signed card and an attestation bound to a person.

Tool pins. Schema, description, and the full tools/list digest are pinned on the plane.
Results checked. A tool result carrying injection or a secret is refused before it re-enters the model.
Shared run budget. Every child of a run draws on one ledger.
a2a.hop · claims-peerattested
card signature      valid
human               u-dev
attestation         signed by control plane
depth               1 of 3
verdict             ALLOW

same hop, no attestation
verdict             DENY  R-ATTEST
Healthcare

Minimum necessary, enforced on the tool call.

PHI is read only for treatment, payment, or operations. It routes only to models with a BAA. Once a session has read PHI, it cannot write it to Slack, email, or a public issue.

Hidden streams. Speaker notes and embedded instructions in documents are quarantined before context assembly.
OCR evidence pack. Export the decisions an investigator will ask for.
claims-bot · ses-claimstaint PHI
ALLOW       ehr.query MRN, purpose payment    R-PERMIT
DENY        slack.post "summary ready"         R-SLACK
DENY        llm.complete → gpt-public          R-PHI-BAA
QUARANTINE  q3-claims.pptx speaker notes       R-HIDDEN
Payments and irreversible actions

Nobody moves money on the model's say-so.

A refund above $500 escalates to an approver who does not own the agent. A payment capture needs an independent observer to sign the same arguments, and then a grant for that exact argument hash.

payment.captureamount 100
no observer                  DENY      R-READBACK
observer signed args hash    ESCALATE  R-PAYMENT
+ approver grant, same hash  ALLOW     R-PERMIT
observer key = appliance key DENY      R-READBACK
Industrial

Parse the frame, then decide the write.

Modbus TCP frames are parsed for function code and register. Reads are allowed. Writes are irreversible and need an observer and a grant. This is a Modbus TCP parser, not an IEC 62443 certification.

ot-line · plcmodbus tcp
MBAP  tx 0001  proto 0000  len 0006  unit 01
FC    06 write single register
REG   40001
verdict DENY  R-READBACK (no observer yet)
MSSP

Parent tenants see counts. Child bodies stay with the child.

A managed provider sees event summaries across child tenants. Requests for child event bodies return 403, so a clinic's patient data never shows up in the parent's console.

GET /api/mssp/clinic-westparent view
GET /api/mssp/clinic-west
200  event summaries only, no bodies

GET /api/mssp/clinic-west/events
403  child bodies are not visible to the parent

Bring your hardest agent workflow.

We'll write the policy with you and show the verdicts on your traffic.

Book a demo →