sekia.ai

I spend my days leading cloud infrastructure security across AWS, GCP, and Azure. That means a constant stream of alerts, access requests, vulnerability findings, and cross-platform noise that demands attention. Most of it is repetitive. Most of it follows patterns. And most of it was eating hours I should have been spending on actual security engineering.

So I built sekia, an open-source multi-agent event bus that connects the tools I already use and lets me automate workflows with Lua scripts and built-in AI. It ties together GitHub, Slack, Linear, Gmail, and Google Calendar over an embedded NATS message bus. Seven small Go binaries, zero external dependencies, one brew install to get started.

This post covers what sekia does, how it handles security, and the three workflows that have had the biggest impact on my day-to-day work.

What sekia Actually Is

sekia is a daemon (sekiad) with an embedded NATS server and JetStream for durable messaging. Lightweight agents (sekia-github, sekia-slack, sekia-linear, sekia-google) connect to the bus, publish events from external services, and execute commands dispatched by Lua workflows. A CLI (sekiactl) and an optional MCP server (sekia-mcp) round things out.

The architecture is simple: agents emit events, Lua scripts react to them, and scripts can send commands back to any connected agent. Everything flows through NATS subjects like sekia.events.github or sekia.commands.slack-agent. Agents heartbeat every 30 seconds, auto-register on connect, and support named instances for multi-tenancy.

The Lua API is intentionally small. sekia.on() subscribes to event patterns. sekia.command() talks to agents. sekia.ai() and sekia.ai_json() call an LLM (Claude by default) directly from your workflow. sekia.schedule() runs handlers on a timer for autonomous agent behavior. sekia.conversation() gives you multi-turn state for threaded interactions. That’s the whole surface area, and it covers a surprising amount of ground.

Workflows live in ~/.config/sekia/workflows/ as plain .lua files. Hot-reload is on by default, so editing a file picks up changes immediately. No deploys, no containers, no CI pipeline for a config change.

Security by Design

I built sekia with the same mindset I bring to production infrastructure. Secrets never sit in plaintext configs. sekia supports three encryption backends that can be mixed in the same TOML file: ENC[...] for local age keypair encryption, KMS[...] for AWS KMS, and ASM[...] for AWS Secrets Manager. You pick the backend that fits your threat model. Age keys can live off-machine via SEKIA_AGE_KEY or SEKIA_AGE_KEY_FILE. AWS backends use the standard SDK credential chain.

Lua workflows run in a sandboxed VM. Only base, table, string, and math are available. No os, no io, no debug, no dofile, no load. A workflow cannot touch the filesystem or execute arbitrary commands on the host.

For supply-chain integrity, sekia supports workflow verification. Enable workflows.verify_integrity and each .lua file is checked against a SHA256 manifest before loading. sekiactl workflows sign generates the manifest. Tampered or unsigned workflows do not execute. When hot-reload is active, updating the manifest triggers a full reload.

The web dashboard (htmx + SSE, embedded in the binary) ships with Content-Security-Policy, X-Frame-Options, X-Content-Type-Options, and Strict-Transport-Security headers on every response. SSE connections are capped at 50 to prevent resource exhaustion. All state-changing requests require CSRF protection via double-submit cookie.

The NATS bus listens on 127.0.0.1:7600 by default and the HTTP API runs over a Unix socket. Nothing is exposed to the network unless you explicitly configure it.

Workflow 1: Auto-Remediating Security Operations Issues

The workflow that saves me the most time handles the triage and remediation loop for security findings. Vulnerability scanners like Wiz generate a constant stream of issues that land in our secops repo. Most follow well-known patterns: a misconfigured S3 bucket policy, an overly permissive IAM role, an unencrypted resource.

With sekia, I have a scheduled workflow that periodically reviews open issues, classifies them with sekia.ai_json(), and takes action based on severity and type. For findings that match known remediation patterns, the workflow opens a PR with the Terraform fix directly. For everything else, it triages into the right priority bucket in Linear and pings the on-call channel in Slack.

sekia.schedule(300, function()
    local result, err = sekia.ai_json(
        "Review these security findings and classify each by severity "
        .. "and whether auto-remediation is possible:\n\n" .. findings,
        { system = sekia.skill("secops-triage") }
    )
    if err then return end

    for _, finding in ipairs(result.findings or {}) do
        if finding.auto_remediate then
            sekia.command("github-agent", "create_comment", {
                owner = "myorg", repo = "secops",
                number = finding.issue_number,
                body = "Auto-remediation PR opened. Review and merge.",
            })
        else
            sekia.command("linear-agent", "create_issue", {
                team_id = SECURITY_TEAM_ID,
                title = "[" .. finding.severity .. "] " .. finding.title,
                description = finding.summary .. "\n\nGitHub: " .. finding.url,
            })
        end
    end
end)

The key here is the skill system. The secops-triage skill is a markdown file that encodes my team’s remediation playbook: what constitutes auto-remediatable, what the Terraform patterns look like, and how to prioritize. The AI does not freelance. It follows documented procedures.

Workflow 2: Cleaning Up My Gmail Inbox

Email is a tax on security engineers. Vendor alerts, compliance notifications, access request confirmations, certificate expiry warnings. Important signal buried under noise.

The Google agent polls Gmail and publishes gmail.message.received events. My cleanup workflow uses AI to classify each message into one of a few buckets: actionable (keep in inbox), informational (archive with a label), or noise (archive immediately). Actionable messages that relate to security operations get forwarded to the appropriate Slack channel as well.

sekia.on("sekia.events.google", function(event)
    if event.type ~= "gmail.message.received" then return end

    local result, err = sekia.ai_json(
        "Classify this email. Respond with action and reason.\n\n"
        .. "From: " .. event.payload.from .. "\n"
        .. "Subject: " .. event.payload.subject .. "\n"
        .. "Body: " .. event.payload.body,
        { system = "Classify as: actionable, informational, or noise. "
          .. "If security-related and actionable, set security=true." }
    )
    if err then return end

    if result.action == "noise" then
        sekia.command("google-agent", "archive", {
            message_id = event.payload.id,
        })
    elseif result.action == "informational" then
        sekia.command("google-agent", "add_label", {
            message_id = event.payload.id, label = "auto/informational",
        })
        sekia.command("google-agent", "archive", {
            message_id = event.payload.id,
        })
    end

    if result.security then
        sekia.command("slack-agent", "send_message", {
            channel = C_SECURITY,
            text = "Flagged email from " .. event.payload.from
                .. ": " .. event.payload.subject,
        })
    end
end)

In a typical week this archives 60-70% of my incoming mail automatically and surfaces the security-relevant messages where I actually see them.

Workflow 3: Daily Email Summaries to Slack

The third workflow ties things together. Every morning at 8am, a scheduled handler scans the last 24 hours of Gmail activity, builds a digest with AI, and posts it to my Slack DM. No more opening Gmail to figure out what happened overnight.

sekia.schedule(86400, function()
    local conv = sekia.conversation("internal", "daily-digest", "singleton")

    local summary, err = conv:reply(
        "Summarize the key emails from the last 24 hours. "
        .. "Group by: security alerts, access requests, vendor updates, "
        .. "and everything else. Keep it concise."
    )
    if err then return end

    sekia.command("slack-agent", "send_message", {
        channel = MY_DM_CHANNEL,
        text = summary,
    })
end)

The conversation API here is a nice touch. Because it maintains multi-turn state, I can follow up in Slack (via the slack.mention event) and ask clarifying questions about specific emails. The context carries over.

Getting Started

sekia installs in one line:

brew install sekia-ai/tap/sekia

Then install whichever agents you need, set your tokens as environment variables, drop a .lua file in the workflows directory, and run sekiad. The documentation walks through setup for each agent.

The project is Apache 2.0 licensed and on GitHub. Contributions, feedback, and workflow ideas are welcome. If you work in security and you are tired of context-switching between six different tools to do the same repetitive tasks, give sekia a look.

What sekia Actually Is#

Security by Design#

Workflow 1: Auto-Remediating Security Operations Issues#

Workflow 2: Cleaning Up My Gmail Inbox#

Workflow 3: Daily Email Summaries to Slack#

Getting Started#