- Remediate: Agents and Pix, 31 built-in agents, Pix Call
- Build: Coding agents, MCP, skills, hooks, Belai
- Govern: AI Firewall, 13 providers, one policy
- Prove: Inventory and compliance, AI-BOM, EU AI Act, ISO 42001
31 built-in agents, and Pix to coordinate them
Each agent has one job: scan a repository, patch a finding, re-check a verdict, plan a feature, reproduce a bug, fix a failing test, review a pull request or resolve a merge conflict. Belai, the open source coding harness, runs them, and Pix coordinates them by phone.
Security fixes
A scout runs the Vulnesis scanners, patchers fix each finding on its own branch, and a verifier re-checks every verdict with the scanner before it is recorded.
Features, bugs and interface
A planner splits a request into small build items, a designer specifies interface changes as user journeys, a bug investigator names the root cause, and builders and reviewers do the rest.
Dependencies by ecosystem
12 agents, each triaging the dependency findings of one group of package ecosystems: Go modules, npm and Deno, Python, Maven and Gradle, Cargo, Composer, Bundler, NuGet, SwiftPM and CocoaPods, CI pipelines, containers and infrastructure files, and the rest.
Upkeep
Agents that fix failing tests and lint drift, review a named pull request read only, and resolve merge conflicts.
We pay the data cost once, upstream
Every advisory is reconciled against the package registry, the distribution advisory and the upstream commit before it reaches you. A wrong affected-version range costs one developer the same hour whether the finding was real or not, and that is the bill bad data sends every team running a scanner on raw CVE feeds.
Call Pix
Pix takes calls from the Resolve Platform header and the installed app. It answers questions about your runs, sessions and hosts, and it reads a plan back before it starts or stops a session, acting only on a plain yes.
Run the factory by phone · Pix Personal Assistant · Software Factory and its agent fleet · Announcing Pix (historical)
The Software Factory, on a phone call
Call Pix from the Resolve Platform or the installed app. It tells you what your agents did and what it cost, starts and steers sessions after a read-back, and the same phone carries every stop down to the sandbox container.
Intel on the go
What is running, what finished, what waits on you and which host is offline, read from your records.
Launch and steer
Start up to five sessions on one yes, pass a running agent a message, or answer its question by voice.
Cost and savings
Spend against budget, and the same work priced by hand from the rates in your settings.
Stops you can reach
Stop a session by voice. Pause automations, cut A2A relay, throw the kill switch or destroy the sandbox container in one press.
Hand the queue to agents that never clock off.
A crew of built-in agents works a finding to a draft pull request, with System One decisions, plain-rule quality gates and a person approving. The Software Factory is what you run; the Pix Sandbox is the machine it runs on, with controlled egress.
- System One: withheld: injection
- AI Firewall: passed
- Package Firewall: refused: malware
- Egress: passed
Your agent installs the dependency. Something should check it first.
A guide per agent, because the path each one reads is different. An agent only knows what its tools tell it. Vulnesis answers at the moment it adds a package, in the agent's own context, and stays silent when your policy is already satisfied. Reach it through the MCP server with no install, or wire hooks and skills locally with one command.
31 MCP tools and 18 workflow prompts
KEV deadlines, exploit intelligence, EPSS and SSVC, ranked remediation and pre-install risk checks, answered from live data. One URL, one header, in twelve MCP clients.
18 Skills, 5 subagents and 3 guard hooks
Typosquat checks, safe fixes, VEX publishing, PR security review and a dependency guard at the tool call. Apache-2.0.
$ vulnetix agent install
Detects the agents you have, wires each one. Free on every plan. AI analysis on the MCP server unlocks on Pro.
A safer LLM coding harness
Belai ships the guardrails the big labs leave out. Sealed delimiters with a nonce and SHA-256 hash, a security classifier on every untrusted tool result, and fail-closed posture gates, built into the harness rather than bolted on.
- Research malware through an agent that classifies tool output before it trusts it, so the bytes stay evidence rather than a running process.
- A forged context block, an unknown nonce, or a failed hash is stripped at egress. An agent cannot widen its own tool surface.
- The role manager segments untrusted content and Belai logs nothing. Your context window stays yours.
- A fixed-argv native tool catalogue, plan and goal modes, and a prompt library as files.
Govern every LLM call without touching a client library
One gateway in front of OpenAI, Anthropic, Mistral, Groq and nine more providers, 13 in all. Your engineers change base_url and nothing else. Model allowlists, regex guardrails and PII redaction apply identically whether the call came from a curl, a Codex turn, a Kiro session or a Claude Code tool result. The gateway and your key vault are free at every tier; policy is included on Pro and guardrails on Business.
chat
OpenAI SDKs, LangChain, curl, most everything
responses
Codex CLI, whose config accepts no other wire API
messages
Anthropic SDKs, Claude Code, Claude Desktop
kiro
Kiro IDE and CLI, signed in with a stored Kiro login instead of an API key
- BYOK: Bring your own provider keys. They are held per organisation as KMS-encrypted ciphertext and decrypted just in time, never riding the request.
- LOGS: No prompts logged. Streaming responses are relayed untouched, as Server-Sent Events (an AWS event stream for Kiro).
- ROLLOUT: Roll a rule out as flag, then redact, then block, so nobody discovers the policy by having a build fail.
Agent governance · Business
Rules for the tools, MCP servers, skills and clients your agents carry. The strip action removes one tool from a request, so the model is never offered it and the session keeps working.
What the repository declares, and what actually ran
AI-BOM from the code
Four evidence passes find every coding agent, SDK and model a project uses, including agents that leave only a commit trailer. CycloneDX 1.7, env var values never read.
Inventory from live traffic
Every tool, MCP server, skill and client your agents send through the gateway, including the MCP server someone wired in locally. Free on every plan.
Your AI inventory is already the evidence
The same inventory, firewall telemetry and triage decisions feed three AI reports, period-scoped and exportable as PDF, OSCAL and an evidence bundle.
Transparent and auditable: metadata only, every decision on the record.