Agents

Every agent tentacle answers the same calls from Octomate: run a turn for a conversation, stream what happens, raise approvals and questions, and say which models it offers. What differs is the runtime behind it.

Enable an agent

Choose Claude Code, Codex or the experimental DeepSeek Harness to reuse an existing runtime. Use Inkling as the backup when you want a model through Pydantic AI. For a harness agent, install and authenticate the runtime as the account running the server. For Inkling, supply provider credentials and at least one model.

Add the agent to tentacles.yaml, then include its id in a channel's agents list. Set enabled: true if its block is disabled, check the configuration and restart Octomate. Send a request through that channel to verify the agent replies. The Tentacles guide walks through the shared steps. To collect sessions from your own app, editor or terminal, also follow the client setup.

Configuration

tentacles:
  claude:
    type: claude
    permission_mode: default
  codex:
    type: codex
    permission_mode: user_review
  inkling:
    type: inkling
    models:
      - name: anthropic:claude-sonnet-5

Each agent type may be declared once, because each owns a fixed hook route. Every block accepts enabled (default true) and gateway (default true): with gateway: false, that agent's driven turns get none of the routing spells.

Routes and claims

A route is one (agent, model) pair a channel can reach, and every route carries a claim: what it is for, and which thinking efforts it accepts. Agents advertise; the caller picks. The answering agent reads the claims when deciding whether to hand work to somebody better placed.

Claude Code, Codex and DeepSeek Harness report their catalogs at startup, with each model's description and supported efforts where the harness has them. Inkling's routes are its configured models. Where a harness reports nothing, claims fills the gap, keyed by the provider-qualified model name:

tentacles:
  claude:
    type: claude
    claims:
      "anthropic:claude-opus-5":
        ability: Deep review and hard refactors.
        efforts: [medium, high, xhigh]

Reported metadata wins over a configured claim. Each harness keeps its own effort names; choose from the levels offered for the selected model. Inkling uses Pydantic AI's minimal, low, medium, high, and xhigh vocabulary. Configured claims must list the levels they accept; omitting them advertises no effort control.

A claim also carries default_effort, the level a run gets when its caller names none, where that is known: Codex's is its effort setting, then Codex's own model_reasoning_effort, then the model's default; Inkling's is the model's thinking setting. It is set only when the route accepts that level. Claude Code and DeepSeek Harness decide for themselves and report nothing, so theirs stays empty.

Models

A harness agent has no configured default model. Omitting one preserves the harness's own settings, including the model a resumed session had chosen. Inkling's default is the first entry in its models list. A channel exposes every model of every agent it binds; the entry agent's default is what answers a new conversation.

Permission modes helps you choose how much an agent may do without asking. Start a conversation covers using a channel or continuing in your agent's own app, editor or terminal.