# Integrations

> The model providers, identity providers, AI coding tools, source control, vendor audit feeds, engineering tools and deployment targets Neural Inverse works with, and how far each one goes.

Source: https://neuralinverse.com/integrations/

---

![](/photos/bench.jpg)

Integrations

# Works with the tools your engineers already use.

Model providers, identity providers, third-party AI tools, source control, vendor audit feeds, engineering vendor tools and where it runs. Each entry says how far the integration goes.

Overview

This page lists what works today. Where an integration has limits, the limit is stated next to it, and work still in progress is kept apart in a list at the end. Product names belong to their owners.

[Models ↓](https://neuralinverse.com/#models)[Identity ↓](https://neuralinverse.com/#identity)[AI coding tools ↓](https://neuralinverse.com/#ai-tools)[Source control and CI ↓](https://neuralinverse.com/#source-control)[Vendor audit ↓](https://neuralinverse.com/#vendor-audit)[Engineering tools ↓](https://neuralinverse.com/#engineering-tools)[Where it runs ↓](https://neuralinverse.com/#deployment)[What's next ↓](https://neuralinverse.com/#whats-next)

Model providers

## Hosted providers and models you run yourself.

Govern's model gateway reaches providers through connections your administrators add. Every model is denied until it is allowed for the organisation, a team or a project.

[Model access](https://neuralinverse.com/docs/govern/model-access/)[Models in the Build apps](https://neuralinverse.com/docs/build/models/)

- **Through the Govern gateway:** Connections to Anthropic, OpenAI, Google, Azure OpenAI, Amazon Bedrock and Google Vertex AI. Provider keys are held in the vault and never returned. Bedrock requests are signed with AWS Signature Version 4.
- **Self-hosted models:** Any OpenAI-compatible server and Ollama, on private networks the operator lists. Their models count as self-hosted, so restricted work can be limited to them.
- **Gateway compatibility:** The gateway is compatible with the OpenAI and Anthropic APIs, so clients built for either can call it with a Govern credential.
- **In the Editor and the Agents app:** Your organisation's models through Govern, or a provider connected with your own key where policy allows. Local models need no key.
- **In the CLI:** Providers work as soon as their credentials are present. Ollama is detected automatically, and any server that speaks the OpenAI API can be added.

Limits: through the gateway, Amazon Bedrock serves Anthropic models only, and Google and Vertex AI Gemini models are reached through their OpenAI-compatible endpoints.

Identity providers and directory sync

## Sign in with your identity provider, provision from your directory.

Govern is the identity provider for the Neural Inverse apps. It signs your people in through your own identity provider and keeps them in step with your directory.

[Identity and access](https://neuralinverse.com/docs/govern/identity/)[Single sign-on with SAML](https://neuralinverse.com/docs/govern/sso-saml/)[Directory sync with SCIM](https://neuralinverse.com/docs/govern/scim/)

- **SAML 2.0:** Single sign-on with presets that fill in the NameID format and attribute names. A connection is tested by a real sign-in before it is activated.
- **OpenID Connect:** Single sign-on with the authorisation code flow and PKCE, with the same test-before-activate rule.
- **Social sign-in:** Per organisation.
- **SCIM 2.0:** Your identity provider creates, updates and deactivates people and groups. Deprovisioning ends sessions, deactivates keys and revokes role bindings before the response returns.
- **LDAP and Active Directory:** Partial: the directory's groups sync into Govern by hand or on a schedule. LDAP is not a sign-in method; use SAML or OpenID Connect for sign-in.
- **Workload identity:** Workloads that already have an identity exchange it for a short-lived gateway token instead of holding a key.

Third-party AI coding tools

## Governed as far as each tool allows, and no further.

Enforced: a blocking hook in a managed location the developer cannot remove. Managed configuration: the tool's settings are locked; its hook is advisory. Network: control holds while the tool's traffic goes through Govern. Monitored: recorded, nothing blocked.

[Third-party AI coding tools](https://neuralinverse.com/docs/govern/third-party-tools/)

| Tool | Policy control | Model traffic through the gateway |
| --- | --- | --- |
| Claude Code | Enforced, with managed settings | Yes (Network) |
| Cursor | Enforced, with the enterprise hooks file | Partial |
| Windsurf | Enforced, with the system hooks file | No |
| GitHub Copilot CLI | Enforced, with the policy hook directory | No |
| OpenAI Codex CLI | Advisory until verified | Yes (Network) |
| Gemini CLI | Managed configuration, advisory hook | No |
| Kiro CLI | Monitored, advisory hook | No |
| Cline | Monitored, advisory hook | Yes (Network) |
| Amazon Q Developer (IDE plugins) | Monitored, through the vendor audit connector | Not documented |
| Continue | Not controlled | Yes (Network) |

### Routed through the model gateway only

| Tool | Level | How |
| --- | --- | --- |
| Aider | Network | OpenAI-compatible base URL |
| Roo Code | Network | OpenAI Compatible provider |
| GitHub Copilot | Monitored | Seats, usage and metrics through the vendor audit connector |

Enforced levels assume the endpoint agent has installed the compiled policy at the tool's managed location. Cursor does not cover file edits; Network control can be bypassed without an egress lock on developer machines. The documentation lists every gap per tool.

Source control, trackers and CI

## Your repositories, your trackers, your CI as the final gate.

Agents work on your repositories with your existing review flow, and merges and releases are decided in your CI on the exact commit.

[Context sources](https://neuralinverse.com/docs/context/spaces-and-sources/)[Pull requests](https://neuralinverse.com/docs/agents-app/pull-requests/)[CLI and GitHub](https://neuralinverse.com/docs/cli/github/)[Trackers and release gates](https://neuralinverse.com/docs/editor/intent-and-projects/)

- **GitHub repositories and issues:** Context indexes repositories, issues, pull requests and recent failed GitHub Actions runs through a GitHub App with read-only permissions.
- **GitHub pull requests:** The Agents app opens a pull request from a session and can start a session from one. The CLI works with pull requests and issues from a session. The Agents app's pull request features work with GitHub only.
- **GitHub Actions:** A ready-made action runs the CLI on one prompt and reports the answer, session and cost as step outputs.
- **Any CI system:** The CLI runs non-interactively with text or JSON output and exit codes. Merge and release gates run in your CI on the exact commit, and bind when they run from protected configuration.
- **Repository addresses:** Projects are matched to repositories by address, including GitLab subgroups and Azure DevOps paths.
- **Trackers in the Editor:** Work items and requirements linked to commits; only identifiers and links are sent, never source code. Each adapter is tested against recorded responses from the vendor's documented API, not live instances.

Vendor AI audit connectors

## Bring other AI services' own records into one place.

Connectors read the audit and usage logs of AI services Govern does not sit in front of. They observe; they do not give Govern control over the vendor's service.

[Vendor AI audit connectors](https://neuralinverse.com/docs/evidence/vendor-connectors/)[Push receivers](https://neuralinverse.com/docs/evidence/push-receivers/)

- **Model and cloud platforms:** Pulled from the vendor's API with credentials held in your vault.
- **Coding assistants:** Pulled from the vendor's API.
- **Workplace AI:** Pulled from the vendor's API.
- **Push receivers:** The tool sends OpenTelemetry events to Govern with a revocable ingest token.
- **Audit webhook:** Any service that can post JSON audit records, authenticated with an ingest token.

Pull connectors call the vendor's service, so an air-gapped installation can use only those whose service is reachable inside your network.

Engineering vendor tools

## Domain modules drive the vendor tools you qualified.

The vendor's own compiler, simulator or host system stays yours; the module drives it. Modules are pre-release, and no capability is verified by a customer yet. The state of each is given as the documentation states it.

[Domain modules and their verification states](https://neuralinverse.com/docs/build/domain-modules/)

- **Embedded firmware:** Running a firmware image in QEMU is verified. Everything else is built and unit-tested on recorded tool output; not yet run against IAR, Keil, TRACE32, J-Link, ST-LINK or STM32CubeProgrammer, on Windows or on a board.
- **HDL (FPGA and ASIC):** The open flow is verified on Linux. The vendor tool paths are kit shipped or modelled, not yet run against the vendor tools. Nothing is verified on Windows.
- **PLC:** Built and tested on Linux. No vendor tool has been run and no controller connected; verification kits are ready.
- **DCS:** Modelled. No DeltaV system or real plant export has been used; the export has a verification kit.
- **Energy (IEC 61850):** The SCL reader and checks are verified on 292 public sample files; online MMS against open-source servers on Linux. No relay vendor tool or relay has been run.
- **Mainframe:** The workstation path is verified on Linux. Nothing has run against a z/OS system yet; a verification kit lets a customer run the module's requests on their own.
- **Automotive:** DBC and ARXML reading, the codec and ODX are verified on sample files; XCP against XCPlite; the bus monitor on virtual interfaces. Vector, PEAK and Kvaser interfaces are kit shipped. No CAN hardware or real ECU yet.

Deployment targets

## Your servers, your cloud, a closed network, or ours.

Govern is a single server binary with PostgreSQL as its only required dependency. The Build apps run on engineers' machines and sign in to your Govern deployment.

[Deploy and operate](https://neuralinverse.com/docs/deploy/)[Air-gapped installation](https://neuralinverse.com/docs/deploy/air-gapped/)[Install the Build apps](https://neuralinverse.com/docs/get-started/install/)

- **Docker Compose:** A single host, for evaluation, development or a small installation. Docker 24.0 and Compose v2.20 or later.
- **Kubernetes with Helm:** Production on your own clusters, on premises or in your cloud. Kubernetes 1.29 or later and Helm 3; each plane runs as its own workload.
- **Your cloud account:** On your Kubernetes cluster with a managed PostgreSQL service, your object storage and your secret store.
- **Air-gapped:** Installed from an offline bundle with checksums, with your registry, identity provider, mail relay and model servers inside the boundary.
- **Database and storage:** PostgreSQL 15 or later. Object storage on a shared directory or an S3-compatible bucket, including on-premises stores with Object Lock for audit archives.
- **Hosted by us:** We run Govern in the region your organisation chooses. Sandbox and build workers can run in your own AWS or Azure account.
- **Engineers' machines:** The Editor and the Agents app on macOS, Windows and Linux. The CLI on macOS (x64, arm64), Linux (x64, arm64; musl on x64) and Windows (x64).

What's next

## Not yet working, and said so.

None of these is available today. Tell us which tools your team runs, and we will say plainly what works.

[Request access](https://neuralinverse.com/contact/?topic=access)

- **Embedded Linux:** Being built: a module for Yocto Project and Wind River Linux builds, in early development behind a setting. Not ready for use.
- **PLC: CODESYS and TIA Portal:** Planned: CODESYS V3.5 and Siemens TIA Portal are in scope after TwinCAT 3. Not started.
- **Other control systems:** Planned: Honeywell Experion, Yokogawa CENTUM VP, Siemens PCS 7 and PCS neo, and ABB 800xA are in scope for the DCS module. Not built.
- **CNC and telecommunications:** Being built: dedicated domain modules. Today these teams use governed AI coding, named approvals and the audit record from Build and Govern.
- **OpenAI Codex CLI enforcement:** In verification: enforcement through Codex's managed hooks is not yet verified against a released Codex CLI, so it is treated as advisory.

![](/photos/team.jpg)

## Talk to our team.

Neural Inverse is available by request. We work with your engineering, security and procurement teams from the first conversation.

[Request access](https://neuralinverse.com/contact/?topic=access)[Contact us](https://neuralinverse.com/contact/)