Ask your infrastructure anything.

The LLM proposes an operation. Cynative resolves it to the IAM actions it requires, authorizes those actions against a read-only policy, and only then attaches a credential. A denial returns an error to the LLM and the loop continues.

The trust boundaries in full →

8 LLM calls · 5 tool calls · 2 verifiers · 184.3s

The LLM writes JavaScript and Cynative runs it safely. This script discovers every enabled region, lists the EKS clusters in each one concurrently, and follows pagination. Only the summary returns to the LLM.

// Discover regions, then list EKS clusters in every region concurrently,
// following pagination - only the summary returns to the LLM.
const r = await http_request({
  method: "GET",
  url: "https://ec2.us-east-1.amazonaws.com/?Action=DescribeRegions&Version=2016-11-15",
  auth_provider: "aws", aws_auth: { service: "ec2", region: "us-east-1" },
});
const regions = [...r.body.matchAll(/<regionName>([^<]+)<\/regionName>/g)].map((m) => m[1]);

const all = await mapConcurrent(regions, async (region) => {
  const clusters = [];
  let token = null;
  do {
    const url = `https://eks.${region}.amazonaws.com/clusters` +
      (token ? `?nextToken=${encodeURIComponent(token)}` : "");
    const resp = await http_request({
      method: "GET", url,
      auth_provider: "aws", aws_auth: { service: "eks", region },
    });
    const body = JSON.parse(resp.body);
    clusters.push(...body.clusters);
    token = body.nextToken;
  } while (token);
  return { region, clusters };
});

console.log(JSON.stringify(all.filter((x) => x.clusters.length > 0), null, 2));

• Runs in an ephemeral process with access to the tools Cynative exposes, and to those only.
• Network and host access stay outside the sandbox.
• Secrets are redacted and never returned to the LLM.
• Every script is shown in full before it runs.

45 built-in agents are working examples. An agent is a Markdown file: frontmatter for the name and description, a body for the methodology. Drop it in ~/.cynative/agents/ and run it by name. It inherits the same action gate, sandbox, approvals and audit log as freeform deep-research sessions. Agents compose: pipe one agent's findings into another with -p, and the piped input is fenced as data. What comes back is a finding with its evidence and its trace, in text you can route into any process your team already runs. The framework finds and verifies; remediation stays your team's decision, made with the evidence in hand.

• Agent file format ↗Frontmatter, body, shadowing, piped stdin
• Running agents ↗--agent with -p, --auto-approve, --config and stdin
• Agent catalog ↗All 45 built-ins with their descriptions

Connectors are discovered at runtime from credentials already in your shell. Cynative works within the permissions you already hold, and each credential stays where it is. Each ceiling is a policy you can change in configuration before the run. The LLM never sees it and cannot move it.

• awspolicy = arn:aws:iam::aws:policy/SecurityAudit · STS session re-vended read-only
• gcprole = roles/viewer
• azurerole definition = Reader
• ekscluster role = view · fetched at runtime, enforced per request
• gkecluster role = view · fetched at runtime, enforced per request
• akscluster role = view · fetched at runtime, enforced per request
• kubernetescluster role = view · self-managed clusters
• githubpermissions = default=read, secret-scanning=none
• gitlabpermissions = default=read, ci-variables=none · GraphQL denied

Even though it's read-only, GitHub's secret-scanning endpoints and GitLab's GraphQL API stay blocked.

Connector guides ↗

Including the one your cloud already runs. Point Cynative at Bedrock, Vertex AI or Azure OpenAI in the account it audits and the model, the agent and the data share one perimeter.

Cynative talks to LLMs through the embedded Bifrost SDK: support for 28 providers, covering OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, Google Vertex, Gemini, Cohere, Mistral, Groq, Ollama and vLLM among others. Set it through environment variables.

$ export CYNATIVE_LLM_PROVIDER=bedrock
$ export CYNATIVE_LLM_MODEL=us.anthropic.claude-opus-5
$ export AWS_REGION=us-east-1

All 28 provider guides ↗

• ThroughputSandboxed code fans out calls concurrently so tasks resolve in fewer tokens and less time.
• FindingsA verifier cross-checks every finding against live evidence.
• Read-onlyThe LLM can't cross its trust boundary.
• SecretsSecrets are redacted from tool output before the LLM sees them.

The documentation lives in the repository.