AI Assistant

Open AI → Copilot for org-scoped conversations that answer infrastructure questions, leverage live platform context, and run governed Platform MCP tools. HIGH-risk tools pause for human approval before they execute.

ⓘ Powered by the Axio Model Platform.See the Model Platform documentation for supported providers, routing rules, and governance.

What it does

▣

Answer questions

Chat with configured, approved models from your org’s AI providers (via the AI Gateway).

✓

Platform context

Include a secret-masked snapshot of your tenant to ground responses (on by default for every user; only admins can turn it off in the UI).

⚒

Run tools

Execute Platform MCP tools from Copilot or AI → MCP Servers. HIGH-risk tools require approval.

Prerequisites & permissions

The Copilot adapts to what your organization has configured. UI gates and API enforcement use different permission strings—custom roles should grant the set you need.

✓ Open Copilot page: copilot:use

✓ Load providers, conversations, capabilities: ai_gateway:read

✓ Send chat / delete conversations / run tools: ai_gateway:use

✓ Approve HIGH-risk tools: ai_gateway:approve

✓ Configure providers, preview context, toggle context off: ai_gateway:manage

✓ Browse MCP catalog: mcp:use (MCP Servers page)

✓ View Activity / tool run history: copilot:read (AI → Activity)

Running a tool also requires that tool’s own requiredPermission (in addition to ai_gateway:use). Administrators configure providers under Administration → Integrations → AI Providers. Only enabled administration integration providers that are configured (credentials or Ollama endpoint), connection-verified (HEALTHY), and have at least one approved CHAT model (not marked unavailable) appear in the Copilot picker.

Assistant modes

Mode is derived from configured assistant providers and registered Platform MCP assistant tools. Labels match in-product chips and banners.

Mode

LLM Chat

MCP Tools

Behavior

Unavailable

×

×

No eligible AI provider and no MCP tools. Configure a provider or ensure platform MCP tools are registered.

Tool runner

×

✓

Run tools by name (Send label becomes Run); no conversational LLM. Banner offers Browse tools and Configure AI provider.

AI Copilot

✓

×

Chat with the assistant only—no Platform MCP assistant tools registered for the org.

Copilot + tools

✓

✓

Full chat plus tool execution (including natural-language requests that name a tool). Banner includes Browse tools.

Chat experience

✓ Conversation list, create, delete

✓ Provider & model picker — explicit model wins over Model Platform route rules

✓ Connection status chip (Connected, Connection Failed, Disabled, Not Verified)

✓ Include platform context — default on; toggle visible only for ai_gateway:manage

✓ Markdown formatted assistant replies

✓ Starter suggestions (cost, drift, compliance, troubleshooting—or MCP examples in tool-runner mode)

✓ Prompt Library hand-off ( — built-in prompt + optional variables from session)

✓ MCP Servers hand-off ( pre-fills the composer)

✓ Legacy URLs , /agents, and /copilot redirect to

Conversations are private to the creator unless you are an AI org administrator (Owner, Admin, or ai_gateway:manage via custom role), who can list all org conversations. Changing provider or model on an open conversation applies on the next send; start a new chat for a clean thread. Users with ai_gateway:manage see per-message metadata chips (model, tokens, context-grounded, masked-field count) and can use Preview context.

Context Engine

Before each chat turn (when context is enabled), Axio gathers a live, tenant-scoped snapshot for grounding and analysis. All strings pass through the shared redactor (secrets, PII). Tool-runner mode never attaches platform context. Up to the most recent chat messages are included in the model prompt (bounded history).

♙User, org, and counts (projects, stacks, workspaces, runners, repositories, drift monitors, open findings)

↗Recent IaC runs and failed-run count

♢Compliance scores (latest per framework) and open security findings (by severity)

▦Runners (online/total) and Git repositories

▤Sample projects/stacks, cost estimates, drift monitors, recent audit activity

♢Deliberately omitted or masked fields are listed in redactionFlags on the gathered snapshot and on assistant messages (contextRedactionFlags).

Users with ai_gateway:manage can click Preview context to inspect the masked snapshot before sending a message.

Tool calling

Tools are registered Platform MCP capabilities. Each tool declares a risk level (LOW, MEDIUM, HIGH), a required org permission, and input validation. Tool input is scanned for secrets before execution. Only HIGH risk sets approvalRequired and waits for ai_gateway:approve.

1

Request

User asks a question or names a tool (from Copilot or MCP Servers).

→

2

Validate

RBAC (gateway + per-tool permission), input validation, and security scanning run before execution.

→

3

Approve (if HIGH risk)

Execution pauses until a user with ai_gateway:approve approves or rejects.

→

4

Result

Outcome is recorded, audited, and returned in chat or the tool-run history.

Tool runs are recorded under AI → Activity (Tool runs tab). Non-administrators see their own executions; AI org administrators see org-wide history. Pending HIGH-risk executions show Approve / Reject actions for approvers. You can also run and approve tools from AI → MCP Servers.