Agents

Agent steps run on Mate, on your own AI key or on your own runner: add a provider key, see the models, the limits per run and the run history.

An agent is an AI that carries out the agent step of an automation: it reads the record that started the automation, uses the same CRM tools as Mate and reports what it did. Where an agent runs is the runtime of the step: Mate (Customermates' hosted agent, paid with the AI credits of your plan, Cloud only), Your own key (an agent in the Customermates cloud or on your own self-hosted instance that uses an API key from your AI provider, billed by the provider) or Your own runner (Claude Code or Codex on your own machine, under your own login).

Open Agents in Settings to see which runtimes are available, add provider keys, see the runners of the workspace and read the recent runs. The page is for people who may change workspace settings.

How do I add a provider key?

Open Agents, click Add key and choose the Provider (Anthropic, OpenAI or Google), give the key a Label, paste the API key from the provider's console and choose a Default model. Customermates asks the provider whether the key works before it keeps it. A key the provider rejects is not saved, and a provider that cannot be reached right now asks you to try again. A key the provider accepts but does not let list its models, for example a restricted OpenAI key, is saved with a note that it could not be checked.

The key is stored encrypted and is never shown again: the list shows the label, the provider and the last four characters only. A workspace has one key per provider, so adding a key for a provider that already has one replaces it. Test asks the provider again and updates Last checked, Remove deletes the key for good. If the server's encryption secret changed after the key was saved, Test and agent runs say that the key cannot be read, and you save it again. Keys work on every paid plan and on self-hosted instances, where the key is yours anyway; the Free plan cannot add one.

What does it cost?

An agent on your own key costs you what the provider charges for the calls, on the provider's bill. It does not use the AI credits of your Customermates plan. An agent on Mate uses the credits of the person who owns the automation, the same allowance that Mate chats and routines use, and each run shows the credits it used. A run on Mate holds credits round by round and stops before it uses more than 10 credits or more than is left of that allowance.

Which models can I use?

The list per provider is maintained in Customermates and shows the models of the providers you added a key for:

  • Anthropic: Claude Sonnet 5.5 (the default), Claude Opus 5.5, Claude Haiku 5.5, Claude Sonnet 4.5 and Claude Haiku 4.5.
  • OpenAI: GPT-5 mini (the default), GPT-5, GPT-5 nano, GPT-4.1 and GPT-4.1 mini.
  • Google: Gemini 2.5 Flash (the default), Gemini 2.5 Pro and Gemini 2.5 Flash-Lite.

A step saved without a model keeps the default model of the workspace's only provider key from then on. With several keys the step needs a model, so removing a key never moves a step to another provider. When a key's default model is no longer offered, the run uses the provider's default model and says so in its transcript. Customermates never asks for a Claude Pro or Max login: the own-key runtime works only with an API key from the provider's console.

What may an agent step do?

An agent step runs as the owner of the automation, with the owner's permissions, without anyone watching. It can read and change records, read the Knowledge Base and use the other workspace tools of Mate. It cannot use the interface and has no web search. Anything that needs an approval, such as deleting records or inviting people, is declined and the agent reports what remains.

A run has two caps: at most 25 tool calls and 10 minutes. When a cap is reached, the run stops and says so. When the event that started the automation came from outside the workspace, for example a form, a webhook or an incoming message, the agent treats the record and everything in it as data, never as instructions. The same holds for a record name that the instruction uses through {{record.name}}: it reaches the agent marked as data and cut to 200 characters. A run that outside text can reach in one of these ways may not send emails or chat messages, send connection invitations, or create, change or start webhooks and routines: the agent says in its report what should be sent or changed. The same holds when the record that started the automation was written by a public form, an incoming webhook or an agent step that outside text could reach, whatever the trigger of the automation is. No agent step creates, changes or deletes Knowledge Base pages. Nor does an agent step start a routine or create or change one that reacts to events, because such a routine could start the automation again without end. A step that an incoming message started is told which conversation to read. A run that is cut off after it started, for example by a server restart, is never started a second time, because it may already have sent or changed something: it ends as timed out a few minutes after its time limit has passed. A workspace runs at most five agent steps on Mate or on its own key at the same time; further ones wait in the queue until one is done. A run that could not be started is tried again within a few minutes, and ends as timed out if it has still not started after 15 minutes. A run that fails because of the provider shows the provider's error class, never the key.

Where do I see what an agent did?

Recent agent runs on the Agents page lists the runs newest first, with the status (Queued, Starting, Running, Completed, Failed, Timed out or Cancelled), the runtime, the model, the tokens and, for Mate, the credits used. Open a run to read its transcript: the instruction, every tool the agent called with a short digest of the arguments, a short summary of each result and the final report. Runs are kept for 90 days. People who see all automations see all runs, everyone else sees the runs of their own automations. The instruction, the transcript and the final report of a run are shown only to the owner of the automation and to administrators; everyone else who sees the run sees its status, runtime, model, tokens and times.

For agents

The agent step is the effect agent of a rule in manage_rules over MCP: runtime is mate, ownKey or runner, prompt is the task (it may use {{record.name}} and {{rule.name}}), model picks a model of the key's provider for ownKey, and runnerId pins one runner for runner. A runtime that is not available makes the effect fail with the reason plan, noKey, noRunner, selfHosted or notConfigured.

Frequently asked questions

Can I use my Claude Pro or Max subscription?

No. Customermates runs agents on an API key from the provider's console, which the provider bills by use. A Claude Pro or Max login is for Claude itself and is not used by Customermates.

Is my key safe?

The key is checked with the provider, stored encrypted and shown again nowhere, not even to administrators. It is opened only for the moment of a run, for that run's calls to the provider, and is never written to the run history.

Why is Your own key not available?

The Agents page names the reason next to each runtime: your plan does not include it, or no provider key has been added yet. A runner is available when one was seen in the last two minutes.

Does a run keep going when I close the page?

Yes. A run is started on the server by the automation and finishes there. The page only shows what was recorded.