August brought Connic MCP and a native Slack connector. In September, agents started taking phone calls. Connic Voice handles the conversation with realtime voice models, and new connectors link agents to phone numbers, SMS, WhatsApp, and RCS. The new Playground runs deployed agents from the dashboard. Agents can now work through websites with built-in browser tools and pause to ask a person for input. Environments gained retention, redaction, access, and copy controls. The base price of an LLM agent run dropped by 28%, and paid plans got higher limits.
Spoken Conversations with Connic Voice
Connic Voice, now in beta, lets an agent hold a live spoken conversation. A native realtime model processes the caller's audio directly and answers with speech. Prompts, Python tools, Retrieval, MCP servers, discoverable tools, and delegation all work during a call. Voice runs on OpenAI realtime models through OpenAI or Azure OpenAI, and on Gemini native audio models through Google Gemini or Vertex AI, with the project's own provider credentials.
version: "1.0"
name: voice-assistant
type: llm
model: openai/gpt-realtime
description: "A helpful voice assistant"
system_prompt: |
Keep replies short and natural. Reply in the user's language.
voice_config:
transcription_model: gpt-realtime-whisper
greeting: "Hi! How can I help?"
thinking_sound: trueThe voice_config block also sets the voice, language, turn detection, interruptions, and an idle timeout. A quiet tone plays while tools run and pauses whenever someone speaks. The agent can end the call itself, and its farewell always plays in full first. For models that support it, such as Gemini 3.8 Live Extended Thinking, the agent's reasoning_effort sets the reasoning level.
For phone numbers, we also introduced new voice connectors, such as Twilio Voice and Telnyx Voice, which route incoming calls on an existing number to a deployed voice agent. The Voice Customer Support template provides a starting point with a Gemini Live agent that answers from Retrieval and demonstrates ticket creation with a mock tool.
The Connic Voice announcement covers the launch, and the Voice documentation lists supported models and current beta limitations.
SMS, WhatsApp, and RCS Conversations
Customers can now message an agent over SMS and MMS, WhatsApp, or RCS. To support this, we added new messaging connectors, including Twilio Messaging and Telnyx Messaging. An inbound connector starts the linked agents for each incoming message with its text, sender details, and attachments as files.
Replies go out through an outbound connector in any of the delivery modes introduced in August: automatic, agent tool, or middleware. A message can carry text, media URLs, or an approved template. When no recipient is set, the reply goes back to the customer who wrote in. Each incoming message also carries a stable conversation ID, which can serve as the session key so the agent remembers earlier messages:
session:
key: input.conversation_id
ttl: 86400A saved Twilio or Telnyx connection is shared between that provider's voice and messaging connectors, so one account setup covers both calls and messages.
The Twilio Messaging and Telnyx Messaging documentation covers sender setup and channels.
Run Agents in the Playground
The Playground, now in beta, runs an agent by hand and shows the result next to the input. It opens as a resizable panel below the current project page and stays open while moving between pages. Switch it on in the project sidebar or click Open Playground on an agent's page, choose an active agent, enter text or JSON, and click Run. The output, traces, logs, and recent manual runs appear in the panel. Runs use the selected environment's tools and connections.
The Playground replaces the manual trigger drawer on the agent page. The quickstart uses it for the first run.
Browser Automation for Agents
Agents can now operate a browser. The new web_browser_* tools open and read pages, click, type, take screenshots the model can see, move the mouse, manage tabs and dialogs, and download and upload files. One entry in tools enables the full set.
version: "1.0"
name: browser-agent
model: connic/gpt-5.6-sol
session: true
description: "Agent that interacts with websites"
system_prompt: |
Open the requested website and inspect its available controls.
Use the browser tools to complete the requested task.
Close the browser when finished.
tools:
- web_browser_*With session: true, cookies and local storage carry over between the agent's runs, so a site the agent signed in to stays signed in. Persistent sessions now control conversation history and browser state separately through session.history and session.browser. Browser time costs €0.03 per minute while a browser is open, with a one-minute minimum per session, and every browser closes when its run ends.
The Browser Automation Demo template combines the browser tools with the human input requests described below. It asks for an email, a masked password, and a project title, then signs in to the Connic dashboard and creates the project.
Our browser automation guide walks through a complete example, and the web tools reference documents every browser tool.
Ask a Person for Input
Approvals already paused a tool call until someone approved it. Now an agent can also pause to ask a person for information it cannot get anywhere else, such as an MFA code. Each named input under approval.inputs becomes a tool that Connic generates without any Python code, and the person's answer becomes the tool result.
approval:
inputs:
- get_mfa:
prompt: Use this tool to request a 2FA code when needed.
params:
- reason: str
- account_email: str
label: MFA code
sensitive: true
timeout: 300The request appears on the Approvals page with the agent's parameters above an input field, and the project team is notified as with any approval. Answers can also come through the REST API or Connic MCP. With sensitive: true, the input is masked, the stored response is encrypted, and the literal value is redacted from logs and traces, while the model still receives it.
Deployment tests can script the answer: an approval_decisions entry now accepts a response alongside the decision.
The approvals documentation covers inputs, parameters, and sensitive responses. Our human-in-the-loop guide shows where approvals and input requests fit in a workflow.
Retention, Redaction, and Run Deletion
Each environment now controls how long its runs are kept and which recorded fields are hidden. Both settings are in the environment drawer under Settings → Git & Environments.
{
"token": "root-token",
"user": {"token": "[REDACTED]", "name": "Ada"},
"items": [{"token": "[REDACTED]", "id": 1}],
"message": "token=user-token"
}Redaction only changes what is recorded. Agents, tools, and callers still receive the original values. Deleting a run removes its inputs, outputs, context, and traces but keeps its status and duration, so usage metrics stay complete.
The environment documentation explains retention and the redaction path syntax. The runs documentation covers deletion.
Environment Permissions and Copying
Environments can now restrict who acts on them. In each environment drawer, On-demand deployments selects the permission groups that may deploy manually, including through connic deploy, and Copy from this env selects the groups that may copy its data into other environments. Automatic deployments and PR workflows from a connected Git branch run as before.
Existing items in the target are skipped unless Replace is selected. Sensitive variable values and API credentials are copied on the server and stay hidden in the browser.
The environment documentation covers permissions and copying, and the REST API reference explains API key scopes.
Lower Run Prices and Higher Plan Limits
The base price of an LLM agent run dropped from €0.047 to €0.0337, a 28% cut. Larger run packages and prepaid terms reduce costs further.
Volume savings apply to LLM agent starts, judge evaluations, runtime, storage, and Retrieval ingest. Tool and sequential agents no longer pay a per-run fee. Their runtime still applies, and each step of a sequence is its own run, billed by its agent type. Web searches now cost €0.015 each and page reads €0.015 per extracted page, instead of one additional run per call.
Paid plans also got higher limits. Developer now allows 10 parallel runs per agent, 15-minute run timeouts, 25 active connectors, and 90 days of data retention. Pro allows 100 parallel runs per agent, 60-minute run timeouts, 5 concurrent deployments, 50 active connectors, and 365 days of data retention.
The pricing page has the calculator and the full rate table.
More Improvements
- •New managed models:
connic/gemini-3.8-flashadds a fast multimodal model with a 1M-token context window, andconnic/qwen3.8-27ba compact open-weight model with image understanding. Managed model catalog - •Environment selection in the CLI:
connic devasks whether to start a quick test or use a reusable named environment and remembers the named environment.connic deploylists target environments and asks for confirmation.connic dev --quick,connic deploy --list, andconnic deploy --env staging --yeskeep scripts and CI non-interactive. CLI reference - •Deploy & skip tests: Git-connected projects can start a manual deployment without the deploy gate from the Deployments page, for example to ship a hotfix while a flaky test is fixed. Deployments triggered by a push still run the gate. Deploy gate documentation
- •Tool result assertions:
expected_tool_callscan now check the value a tool returned through theresultbinding, including the error flag and content parts of MCP tools. Assertion reference - •Sequential steps as child runs: Each step of a sequential agent now appears as its own linked child run with a separate trace. Runs and traces
- •Faster run details: Log and trace data render only when their section is opened, so large runs open faster. Responses that consist of a JSON code block are formatted and highlighted as JSON.
- •Run timeouts: Retrieval queries and managed model requests now end at the run's deadline, and model retries stop early when the remaining time is too short.
- •Connic in German: The website, documentation, and blog are now available in German. Visitors whose browser prefers German see a suggestion to switch on English pages.