Skip to content
part of the groupApertiaDraivix

01AI and integrations · module

MCP for external AI agents

Connect Claude Code, Codex and other agents to your ERP under specific user permissions

AvailableBook a consultation
  • API key
  • Personal access token
  • MCP token
  • AI conversation
13
capabilities
4
record types
3
automations
12
integrations and AI tools

02What the module does

Less routine. More value.

MCP (Model Context Protocol) lets external AI agents such as Claude Code or Codex work with data in Draivix. The agent receives the same tool catalog as the built-in assistant, and every call runs under a specific user’s permissions and API key. Your business can connect its own agents without bypassing access controls or audit logs.

  • The agent knows what it may callAn external AI agent downloads a catalog of tools complete with input and output descriptions. Your agents work with data against a documented interface rather than by guesswork.
  • 85 tools at a single endpointThrough a single endpoint, an agent can access all 85 allowed tools of the built-in assistant, keeping custom agent integrations straightforward.
  • Built-in safeguard for change proposalsWrites through the capability bridge return a proposal by default. The agent or a user confirms it with mutation_confirm or cancels it with mutation_cancel.
  • Keys scoped to selected toolsThe API key an agent uses to sign in can be restricted to a named list of tools. The agent gets only what your company permits and nothing more.
  • Key permanently pinned to clean-room modeAn API key can be permanently pinned to clean-room mode for sensitive data. Such a key only sees clean-room tools in the catalog.
  • Agents never exceed user permissionsAn agent’s effective access is the intersection of the user’s permissions and the API key’s ceiling. The agent can only do what both allow, and the key itself never unlocks any record on its own.
  • Dedicated server for your agentsThe standalone MCP (Model Context Protocol) server newerp-agents-mcp offers 53 tools built on top of the REST API. Its writes execute immediately within the user’s permissions, without approval proposals.
  • Authentication tailored to the agentAn agent signs in with a passed token, a personal access token, or by signing a challenge with an SSH key—the latter being ideal for command-line agents.
  • Agent work traceable on the task12 tools link the agent’s work to tasks and code commits in GitLab, so you can see exactly what the agent did right on the task record.
  • Chat and document tools on demandThe server can be extended with optional tools for Mattermost and Gemini-powered document analysis (which requires a Gemini API key). Enable only what you actually use.
  • Every call recorded in the audit logEvery agent call is written to the audit log with the mcp source, so you can always tell what an external agent did versus what a person did in the app.
  • Capped at 600 calls per minuteThe capability bridge—the agent interface built directly into Draivix—is limited to 600 calls per minute, keeping agent load within clear boundaries.
  • Conversation handoff to HermesA conversation from the assistant can be handed over to the Hermes agent service with a short-lived 300-second credential. It runs under the permissions of the user who started the conversation and is logged in the audit trail.

03How it works

How it works step by step.

  1. Two paths for agents

    The first path is the capability bridge built directly into Draivix. An agent downloads the tool catalog and calls tools through a single endpoint. The catalog includes the allowed tools of the built-in assistant (85 out of roughly 100 registered tools) plus nine actions: confirming or cancelling a change proposal, editing items in a bulk proposal, sending an email and closing a displayed canvas. The catalog is versioned; if nothing has changed, the server replies that it is up to date so the agent does not need to download it again.

    Read in detail

    The second path is the standalone MCP server newerp-agents-mcp, which wraps the general REST API for coding and operations agents. It provides 53 tools: working with records and metadata, relationships, attachments, email, tasks, activities and reports, global search, calendar, 12 tools for linking agent work to GitLab tasks and commits, and optionally 7 tools for Mattermost. The document analysis tool requires a Gemini API key. The server runs over stdio or HTTP and truncates responses so they fit within the model’s context window.

  2. Permissions and API keys

    An agent always acts as a specific user. Capability bridge permission only lets the key access the interface—by itself, it does not grant access to any record. Effective access is the intersection of the user’s permissions and the API key’s ceiling. A key can be restricted to a specific list of tools and permanently pinned to clean-room mode; such a key only sees clean-room tools in the catalog.

    Read in detail

    The newerp-agents-mcp server has no permissions of its own: every call carries a user token and goes through the standard REST API. Agents can authenticate with a passed bearer token, a personal access token, or by signing a challenge with an SSH key—which is ideal for command-line agents.

    The capability bridge is rate-limited to 600 calls per minute. Every call runs within an audit context tagged with the mcp source and a trace identifier. Whether a tool modifies data is passed to the agent as a standard MCP hint (read-only or mutating).

  3. Change control

    The capability bridge follows the same rule as the built-in assistant: create, update and delete tools return a proposal by default. Either the agent or a person confirms it using mutation_confirm or cancels it using mutation_cancel. Confirming re-runs the actual operation under the user’s permissions.

    Read in detail

    The newerp-agents-mcp server does not use proposals or clean-room mode. Its writes are executed immediately, always within the permissions of the user whose token the agent is using. For agents where you want to approve every change, use the capability bridge instead.

  4. Delegation to the Hermes agent service

    A conversation in the built-in assistant can be handed over to the Hermes agent service. Hermes receives a short-lived credential valid for 300 seconds, bound to the company, and calls back through the capability bridge. Calls run with the permissions of the user who started the conversation; clean-room mode cannot be widened. Every delegated call is recorded in the audit log as security-relevant.

04What runs automatically

Automations that save time.

  1. No redundant catalog downloads

    The tool catalog is cached for 15 minutes per company and version. When an agent checks in and the catalog hasn’t changed, the server doesn’t resend it and the agent doesn’t have to download it again.

  2. An audit trail for every call

    Every agent call automatically creates an audit log entry tagged with the mcp source and a trace ID, making it easy to look up any specific action later.

  3. Credentials expire automatically

    When a conversation is handed over, the Hermes service receives a company-bound credential that expires after 300 seconds, keeping access strictly short-lived.

External integrations

  • Claude Code as an agentConnect Claude Code to Draivix via MCP (Model Context Protocol). It works with data under a specific user’s permissions and API key, without bypassing permissions or audit logging.
  • Codex under the same rulesCodex connects to Draivix via MCP and receives the same tool catalog as the built-in assistant. Every call runs under a specific user’s permissions.
  • Hermes takes over the conversationThe Hermes agent service receives a handed-off conversation from the assistant and calls back through the capability bridge. Clean-room mode cannot be widened, and every call is recorded in the audit log.
  • Document analysis via GeminiThe newerp-agents-mcp server includes a document analysis tool powered by Gemini (requires a Gemini API key), allowing the agent to work with document contents as well.
  • Agent work alongside commitsTools in newerp-agents-mcp link the agent’s work to tasks and commits in GitLab, so you can trace what the agent did directly on the relevant task.
  • Mattermost when you need itThe newerp-agents-mcp server can be extended with 7 optional Mattermost tools, letting the agent work with team chat whenever your company wants it to.

What the AI agent handles

  • Data access for your agents

    An external agent can read, search and count records of any entity (record type), seeing only what the user and their API key are permitted to access.

  • Changes start as proposals

    Through the capability bridge, an agent only proposes a record change by default. Once the agent or a person confirms it, the operation runs again under the user’s permissions.

  • Fine-tune bulk proposals beforehand

    Individual items in a bulk proposal can be edited before confirmation, so you don’t have to reject the entire batch over a single incorrect item.

  • Agents can send email too

    An agent can send an email as one of the actions in the tool catalog. The action runs under a specific user and is recorded in the audit log, so you always know who is behind it.

  • Attachments and email via the server

    Through the newerp-agents-mcp server, an agent can work with attachments and email. Writes execute immediately within the user’s permissions; use the capability bridge if you want to approve every change.

  • Agent work linked to the right task

    The agent links its work to tasks and code commits in GitLab, so during review you can see which task its changes belong to.

A person approves whenever money, a customer, or an exception is involved.

05Frequently asked questions

Can I connect Claude Code or another AI agent to Draivix?

Yes. The agent connects via MCP—either through the capability bridge in Draivix or the standalone newerp-agents-mcp server—and operates with the permissions of the user whose key or token it uses.

Can an agent bypass permissions over MCP?

No. Every call runs as a specific user, and effective access is the intersection of that user’s permissions and the API key’s ceiling. The key can also be restricted to selected tools.

Does the agent write changes immediately?

Through the capability bridge, create, update and delete operations return a proposal by default that must be confirmed separately. The newerp-agents-mcp server writes immediately within the user’s permissions.

Can I trace what an agent did in the system?

Yes. Every call through the capability bridge is logged in the audit trail with the mcp source and a trace ID, and delegated calls from the Hermes service are flagged as security-relevant.

06Related modules

Modules that work great together.

All modules in AI and integrations

Let’s start together

Tell us how your business works.

We will show you what a system that takes over part of that work could look like.

Book a consultation