Wabi

AI agents

Build a support agent from your knowledge, test and evaluate it, ship versions from draft to production, and read every session's trace.

An agent answers customers on a number from its knowledge sources, cites what it used, calls the tools you allow within their permissions, and hands a conversation to your team with a summary when evidence is thin, the customer asks for a person or a rule says so. Every turn is traced. Try it on the sandbox number on any plan; live numbers need Starter or higher.

Versions

Configuration is versioned. You edit a draft (update_agent), promote it to test (the sandbox number and test chats) and then to production (live numbers) with promote_agent_version. A promoted version is frozen; rollback_agent restores the previous production version, and running sessions keep the version they started with. get_agent_config_schema describes the configuration: model, instructions, knowledge, tools, escalation and approval policies, language, budgets and limits.

Build one in the sandbox

# 1. An agent with version 1 as a draftnpx @meterapp/wabi create_agent --json '{"name":"Support"}'# 2. Knowledge it answers fromnpx @meterapp/wabi add_knowledge_source --json '{"kind":"sitemap","name":"Help center","url":"https://example.com/sitemap.xml","path_prefix":"/help/"}'# 3. Try a turn without messaging anyonenpx @meterapp/wabi test_agent --json '{"agent_id":"<id>","text":"Can I change my booking to Friday?"}'# 4. Ship to test, put it on the sandbox number, talk to itnpx @meterapp/wabi promote_agent_version --json '{"agent_id":"<id>","version_id":"<draft id>","to":"test"}'npx @meterapp/wabi deploy_agent --json '{"agent_id":"<id>","phone_number_id":"<sandbox phone_number.id>","environment":"test"}'npx @meterapp/wabi simulate_inbound_message --text "Hi, where is my order?"

Test and evaluate

  • test_agent runs one turn and returns the decision (respond, approve or escalate), the reply, cited sources, confidence, tool calls (side effects simulated) and cost.
  • Golden answers: create_eval_case, run_eval_suite and get_eval_run (Growth and up). replay_conversation re-runs a real conversation against a draft before you promote it.
  • search_knowledge shows exactly what the agent would retrieve for a question.

Observe

list_agent_sessions and get_agent_session are the session explorer: every turn with the sources retrieved and cited, checks, tool calls, tokens, latency and cost, and why it stopped or escalated. get_agent_metrics adds resolution, containment, escalation, cost per resolved issue and per-tool success rates; audit_conversations finds recurring unresolved issues and the documentation to write. The same numbers are in Analytics.

What it costs

Each turn that writes a customer message or a suggested reply on Wabi's model key is one AI reply against your plan. From Growth up you can bring your own model key (set_agent_secret); those turns are not counted.