July introduced managed models, Project credit, and the Marketplace. In August, Connic MCP brought live Project work into ChatGPT, Codex, Claude, Cursor, and other MCP clients through scoped permissions. Slack can now trigger agents and receive their results. Outbound connectors gained automatic, agent-tool, and middleware delivery modes. We also added a conversation view for persistent sessions, file returns from custom tools, more sign-in methods, and five managed models.
Connic MCP in Your AI Client
Connic MCP connects an AI client to one Project through OAuth. During authorization, you choose which environments it can see, which read actions it can use, and whether to grant write actions. The client can never exceed your own Project permissions, and you can edit or revoke every authorization under Project Settings → API Keys & MCP Auth.
A connected client can read runs, traces, logs, deployment tests, connector status, Retrieval, Database, Sessions, budgets, approvals, and governance records. With write access, it can manage the supported resources directly from the conversation. Stored secret values, payments, credential management, and Project deletion stay outside the MCP surface.
codex mcp add connic --url https://mcp.connic.co/mcp
codex mcp login connicThe full Connic plugin bundles the MCP connection with the Connic skill for supported clients, and the server is now listed in the official MCP Registry. API keys use the same action-level permission catalog, so an automation can receive only the operations it needs. Read the Connic MCP announcement or follow the AI agent setup guide.
Slack as a Native Connector
Slack is a native connector in the Marketplace. Create the Slack app from a generated manifest, save its signing secret and bot token in a reusable Connic connection, and link that connection to the agents that should answer in Slack. The app remains in your Slack account, so you control its name, channels, installation, and removal.
An app_mention starts every linked agent with the Slack channel, user, and thread context. Link an outbound Slack connector through the same connection to return the answer to the originating thread. A run can also target another channel explicitly, or use the connector’s default channel.
Follow the Slack connector setup for the complete inbound and outbound flow.
Choose When Outbound Connectors Deliver
Each outbound connector link has a delivery mode. Choose it when you link a connector to an agent, then change it from the connector flow if the routing decision moves elsewhere.
send_connector from Python when application code should make the routing decision.Run traces now show inbound and outbound connector activity alongside the rest of the execution. Deployment tests record agent-tool and middleware connector calls without delivering them, so routing can be asserted without sending test messages to a real destination.
Read the outbound routing guide for examples across the supported connector types.
Inspect the Conversation Behind a Session
You can now inspect the memory that persistent sessions carry across requests. Open Storage → Sessions, select a row, and the new details drawer renders the sanitized recent conversation alongside the runs that used the same session identity.
The conversation view includes user, assistant, system, and tool messages. Tool arguments sit beside their recorded results or errors; attachments show file metadata; retained reasoning appears when it is available, and explicitly redacted reasoning is marked as such. Select a related run to move from the conversation into the exact execution trace.
Run details also have copy controls for inputs, outputs, context, and trace values. Structured values copy as JSON, and reasoning-token usage appears in run traces and cost breakdowns.
See how persistent sessions retain conversation history or configure persistent sessions.
Files from Custom Tools
A custom tool can return a ToolFile when the next model turn needs a PDF, image, audio file, or another binary result. Give it the correct MIME type and either inline bytes or a URI. Lists and tuples can combine files with text and JSON results.
from connic import ToolFile
def export_invoice(invoice_id: str) -> ToolFile:
"""Generate an invoice PDF for the agent to inspect."""
pdf = build_invoice_pdf(invoice_id)
return ToolFile(
mime_type="application/pdf",
name=f"invoice-{invoice_id}.pdf",
data=pdf,
)Read the custom tool file reference for inline and URI-backed results.
Account Security and Sign-In Options
Account Settings puts sign-in security and connected identity providers in one place. A recent sign-in is required before sensitive changes, and Connic keeps at least one working sign-in method attached to the account.
Open account settings to review the methods connected to your account.
Five New Managed Models
The managed catalog added five durable model IDs in August. The two DeepSeek entries expand the long-context reasoning choices for coding and agentic work. Three Gemini Flash entries add multimodal options for high-volume and long-context workloads.
connic/deepseek-v4-flash-0731
connic/deepseek-v4-pro-fast
connic/gemini-3.5-flash-lite
connic/gemini-3.6-flash
connic/gemini-3.7-flashPublished prices also fell for several existing models. The catalog carries the current input, cached-input, and output rates for every managed ID.
Browse the managed model catalog for exact capabilities, limits, and prices.
More Improvements
- •CLI updates:
connic updatecan check or install updates for the Composer SDK, Project skill copies, and Connic plugins already installed in Codex and Claude Code. Review the CLI update commands - •More complete migrations: The LangChain and Google ADK migrators follow additional cross-file imports and re-exports, retain prompts and tools from more project structures, and convert simple sub-agent wrappers into Connic delegation. Follow the LangChain migration guide
- •CI-ready test results:
connic test --jsonreturns the test-run, deployment, and environment IDs. Exit codes distinguish failed or cancelled suites and request errors from infrastructure or server errors. Review the test and CI behavior - •Project data sharing: A default-off toggle under Project Settings → General lets a Project opt into anonymized product analytics. Public aggregate statistics use data only from opted-in Projects