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

Early access

Govern every AI agent your teams run, on the models you're allowed to run

Your teams are building and running AI agents, and some of your data can't go to a cloud LLM. We're building a self-hosted layer between your agents and your models: only approved agents run, every tool call is checked before it executes, and you keep the record. It starts with coding agents, where the open-weight models you can self-host most often skip or garble tool calls, and makes those calls reliable.

No account created, nothing charged - just lets us reach out as this takes shape. Your email is only used for that; it's never shared or used for anything else. See the privacy page for details.

Same model, same 30 tasks

qwen2.5-coder 14B, self-hosted

  • Polyglot23/30
  • goose + toolshim21/30
  • pi0/30
  • goose0/30
  • Hermes0/30
  • opencode0/30
Most agents make no tool calls at all on this model. The Gateway repairs tool calls with the same engine. Method and every run.

What it includes

One Gateway between your team's agents and your models: it decides what agents may do, keeps the record, and keeps tool calls working on self-hosted models.

Govern

See every agent, decide what it may do, keep the record.

  • Inventory, discovered from traffic

    Every agent, tool and model in use, and who ran it. Nobody fills in a form.

  • Approval per environment

    Approve agents and tools for development, staging and production separately; suspend one in seconds.

  • Policy as code

    Rules in your repo, signed with your key, tried in monitor mode first. A refused call is explained to the model.

  • Activity and audit trail

    Every tool call and decision: which agent, which model, which rule.

Run

Agents that keep working on the models you're allowed to run.

  • One shared endpoint

    OpenAI-compatible, for your team's agents; self-hosted or cloud models behind it.

  • Tool-call recovery

    Repairs the malformed or missing tool calls that make agents stall on self-hosted models.

  • Model check

    Does it make proper tool calls, does its chat template fit, is its context window cutting prompts short.

  • Drift alerts and scorecard

    Failed or refused calls, cut-off prompts, latency and cost, per model, with alerts on the dashboard and to Slack.

Qualify

Building next

Know which models can be trusted with agent work before you approve them.

Qualification report

Test a model or quantization against your own tool schemas, in your environment; get a signed report.

Upgrade comparison

A newer model side by side with the approved one, on every measure, so upgrades don't stall in review.

Re-qualification

A model is flagged when its weights, quantization or behaviour change.

Built to pass a security review

It works on its own - no other AI platform required.

Self-hosted, air-gap ready

Nothing about your code or models leaves your infrastructure, and there's no phone-home.

Your policies, your key

Versioned files in your own repo, signed with your own key. We hold no key that can change them.

Sign-in through your identity provider

Building next

OIDC SSO on every plan; SAML and SCIM on Enterprise. No separate passwords to manage.

Exports to the tools you already run

Building next

JSON Lines, syslog and OpenTelemetry, for your SIEM and governance suite.

Indicative pricing

We haven't finalized this - showing a real number here so early signups reflect actual interest, not just curiosity. Tell us if it's off.

Starter

Up to 15 developers

$29/developer/mo

  • Shared endpoint with tool-call recovery
  • Reliability scorecard and drift alerts
  • Qualification report for one model at a time
  • OIDC single sign-on

Team

16-75 developers

Talk to us

  • Everything in Starter
  • Activity feed and audit trail
  • Model approval workflow and re-qualification
  • Policy in monitor mode, per-user budgets, SIEM exports

Enterprise

75+ developers, or any regulated team

Talk to us

  • Everything in Team
  • Policy enforcement and approvals
  • Signed evidence exports
  • SAML, SCIM and an air-gapped install

Get early access

We'll reach out as this takes shape - to ask what you'd need, not to sell you anything yet.

Your email is only used to reach out about this - never shared or used for anything else. See the privacy page for details.