Agent Tools

The canonical reference for everything the Oxagen agent can do. For the why and how — one registry, three call sites — see Agent Overview.

Each table below lists the tool name, the surfaces it's exposed on, the credit cost per successful call, and a one-line description. Costs and policies are pulled from the live registry; if a number here disagrees with the dashboard, the dashboard wins.

Capability platform

Every tool below is registered through one shared capability layer. Whether you call a tool from the in-app chat, an MCP client, or the REST surface, the same registry runs validation, permission checks, credit reservation, and result caching.

These four guarantees are why every tool in the tables below can be called identically from chat, MCP, and REST.

Ontology — read

The agent's most-used tools. Search, fetch, and list typed entities and relationships in your workspace knowledge graph.

Ontology — write

Typed writes against the graph. The agent uses these to record what it learned in a conversation so the next session has it too. Destructive writes (merge_nodes, delete_*) require explicit user confirmation in the chat UI.

Graph algorithms

Requires a Neo4j-routed workspace.

Code graph

Available once a GitHub connection has ingested code into the workspace. These tools traverse the typed code graph (File, Function, Class, Module, plus CALLS, IMPORTS, DEFINES, READS, WRITES edges) and are how the agent answers questions like "what calls processPayment and what tests cover it?".

CI / test layer

Available when the workspace has a CI integration connected.

Code — write

The single primitive that lets an agent propose a code change rather than describe one. The agent opens a draft pull request against the target repo; nothing merges automatically.

Memory

Durable agent memory — patterns the agent has learned, procedures it has been taught, and recent work it should remember. See Agent Memory for the deeper model.

Agent sessions

Multi-agent coordination primitives. When two agents work the same workspace concurrently, these tools prevent them from stepping on each other's writes.

Workspace + tenant

Light-touch tools the agent uses to read or update workspace-level configuration.

Web

A narrow outbound surface so the agent can pull in external context when answering a question.

Notify

Visualization (app surface only)

Used by the chat agent to render an inline visualization inside an assistant message. Not available on MCP — your IDE renders its own UI.

Security

Available when GitHub code scanning, Dependabot, or Semgrep is connected to the workspace.

Tool policies

Every tool carries a policy that determines whether it can run silently or must prompt the user first. The chat UI surfaces a confirmation card for anything stricter than auto.

Confirmation cards arrive in the chat as a structured oxagen_confirm part — the user clicks Approve or Decline in-thread and the server executes the pending tool call.

Credits

Credits are reserved before a tool runs and refunded automatically if the handler errors. A failed call costs nothing; a successful call deducts the cost shown in the table.

Buy credits from the Oxagen web app at Settings → Billing. The Pro plan includes a monthly credit allowance; overage is billed month-to-month. Enterprise plans include configurable credit budgets, per-user caps, and a real-time spend dashboard.

Calling tools yourself

If you want to invoke a tool outside the agent loop — say from a Celery job that needs the same dedupe semantics as the agent — every capability is also a REST endpoint.

curl -X POST https://api.oxagen.ai/v1/cap/ontology.search \
  -H "Authorization: Bearer $OXAGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "validateCard",
    "top_k": 20
  }'

The full schema for every capability is published at https://api.oxagen.ai/openapi.json and rendered in the API Explorer.

Documents, slides, sheets

Agent-authored artifacts in Google Workspace. The connection_kind="google_workspace" binding resolves the workspace's active OAuth grant; calls in workspaces without a connected Google account return a 409 with the missing-connection details.

PDF export

Generic LibreOffice-backed PDF conversion. Accepts any office document (Google Docs / Slides / Sheets export, Word, PowerPoint, Excel, ODT) and returns PDF bytes plus page count. See PDF generation.

Branding

Per-workspace brand kits — palette, typography, spacing, logo, favicon, watermark, voice doc, footer. Every artifact tool above picks up the active kit automatically when present. See Brand kits.

Install MCP server · Read the API docs · Agent memory