For teams: coding agents on the self-hosted models you're allowed to run - see early access →
polyglot

Security

Runs in your environment. Here's exactly what it does there.

Polyglot is self-hosted: the agent, the gateway and their logs run on your infrastructure, and none of your code, prompts or model output comes to us. This page lists every network connection it makes, what it stores, how we build and release it, and where we stand on certification. Each item says whether it works today or is planned.

What leaves your network

The complete list of outbound connections. Run the CLI against a local model with web search off, and the only connection that leaves the network is the npm version check.

Polyglot CLI

The model endpoint you configure

The conversation. With a local model it never leaves the machine.

The search backend, only when the agent searches the web

The search query. DuckDuckGo by default, or your own SearXNG.

Remote MCP servers you add

Whatever the model passes to those tools.

npm registry, at startup

A version check. Nothing about your session.

Polyglot gateway (teams)

Your model servers

The requests your agents send through it.

Your own control plane, if you run one

Activity and audit events, to a server you host.

Our licensing endpoint, only if you turn on automatic license renewal

Your license ID, to fetch a renewed license file. Off by default; without it, you install a new license file yourself. Seat limits are checked offline, against the license file.

Details for the CLI, including every file it writes and every setting that controls it, are in Data handling.

What it stores

The CLI keeps transcripts and plans as plain files under ~/.polyglot/ on the developer's machine. Turn persistence off, or set a retention period that deletes them automatically.
Usage telemetry is off by default. If you opt in, it's written to a local file and never sent anywhere.
The gateway keeps its activity and decision log in a database on your own disk.
We don't add our own file encryption: a key on the same disk adds little. Use full-disk encryption, which your endpoint policy likely requires already.

How we build it, and how it protects you

  • Every npm release is built in CI and published with a signed provenance attestation, so you can verify it came from our repositoryWorking today
  • CodeQL security scanning on every change and weekly, and dependency alerts for known vulnerabilitiesWorking today
  • Secret scanning of tool output: API keys, tokens and private keys are flagged or redacted before they reach the modelWorking today
  • Tool calls are only ever read from the model's own replies: text in a file or web page that looks like a tool call is never run as oneWorking today
  • Permission modes: approve every write, command and network call; plan first and approve the plan; or run freely within deny rules you setWorking today
  • A software bill of materials (SBOM) with every releasePlanned
  • Policies as signed, versioned files, signed with a key you holdPlanned
  • Sign-in through your identity provider (OIDC), with SAML and SCIM for EnterprisePlanned
  • A sandbox for the commands agents runPlanned

Certification and assurance

Security questionnaires

We'll complete yours, whether it's CAIQ, SIG Lite or your own format, and walk your team through the architecture.

SOC 2 and ISO 27001

Not certified yet. Polyglot runs in your environment and none of your data comes to us, which keeps the scope small. We'll certify when our first customers' procurement requires it: ISO 27001 or SOC 2, whichever they ask for.

Penetration testing

An independent test of the gateway is planned before the governance features are generally available. We'll share the summary.

Reporting a vulnerability

Please use GitHub's private vulnerability reporting rather than a public issue. Reports are acknowledged as quickly as possible; Polyglot is early-stage, so response times are best effort rather than a contractual SLA. The full policy, including what counts as expected behaviour, is in SECURITY.md. For a security review or questionnaire, use the teams page to get in touch.