After a harsh July, flespi's monthly uptime in August was 100%. All the hard work we did internally, such as upgrading the Postgres database cluster to the latest major version or moving databases between servers, was almost completely invisible from the outside. And this is actually the way it should be. We will continue with silent maintenance operations on the platform and its infrastructure in the subsequent months, as usual, in the background.
This time, the major changes were towards AI-related services. And, on occasion, we completed and released two independent features in the same month – connecting to flespi from external AI agents, such as Claude, Cursor, or ChatGPT, and our own flespi AI agents.
Connecting external AI agents to flespi is done via realms. To some extent, this complicates the account setup by requiring users to create realms and define users within them. But at the same time, it adds an additional security layer, prompting you to consider who can use your flespi account and how. On top of this, realms support corporate identity providers, which generally simplifies user management. Once a realm and its users are configured, providing access to flespi MCP for an AI agent of your choice is a one-button operation.
In this mode, a modern frontier model powering Claude or ChatGPT can gather information from your flespi account or even make changes. All the security controls lie either in the token configuration issued by the realm for a particular user or in the AI agent's harness configuration. Keep in mind that modern AI models with CLI access, when working on your task, can bypass the harness security and use the token directly. That's why, in this mode, we do not recommend providing external AI agents with any tokens other than ACL tokens, with strict security settings.
Our own flespi AI agents are much more secure. All modification operations are gated by explicit approval, and they don't have CLI access to bypass it, so it's generally safe to provide them even with master tokens.
Security and safety of use are exactly what distinguishes flespi AI agents from any other agentic system. Another notable feature is their intelligence specifically in telematics and flespi. We packed all our experience operating and supporting hundreds of telematics projects into our agents – now they probably know more about telematics devices than anyone else in the world. They know which device brand is better to use in which situations, their pain points, and their advantages over one another.
Equipped with actual documentation and capable of analyzing raw device traffic to extract patterns, they are magnificent telematics consultants, outperforming even Codi, AI Assistant, on most topics thanks to access to far more capable tooling.
And of course, they can make real changes – making actual flespi REST API calls and assessing their results. Add to this an extended scripting system, embedded MQTT client capabilities for tracking and reacting to any event in MQTT, and, of course, external communication surfaces such as email, Telegram, or Slack.
We made AI agents extensible with custom connectors, so the same approval or computational systems can be utilized outside of flespi. Just like protocols, connectors can be built privately to provide an agent with access to your company data and made available only to specific flespi accounts. Wialon Hosting is exactly an example of such an external platform connector, giving agents access to both flespi and Wialon Hosting for shared housekeeping.
Codi, our AI assistant, received its own set of updates alongside the newly released AI systems.
First of all, we adjusted model selection depending on the plan you're working with us on, the type of question or task you ask Codi to handle, and how frequently you use Codi. Second, and perhaps even more importantly, we now set a per-plan limit on the number of specific delegated tasks you can assign to Codi each month. He decides whether your question is acceptable and can be answered by a normal human support engineer. If it isn't, Codi may refuse to answer after you reach a certain number of such tasks per day or month. We suggest offloading all such tasks to AI agents.
And the third long-awaited feature was deployed in August – in video telematics, we supported the WebRTC protocol for live streaming. We are currently working on end-to-end voice communication between the driver and dispatcher for certain devices that support this functionality, which we plan to release later this month.
In the tachograph direction, we completed the tacho-file-parse plugin with parsing of all available fields in DDD files. Lisle Design Rack support was further matured, and now our Tacho Bridge Application can be set up to run with a few hundred company cards inserted. And Evgeny Shatilo – our most experienced engineer in telematics and tachography – has published the very first article in his series revealing the naked truth about tachographs that nobody will tell you.
We supported three new protocols, and all of them are non-standard: wirepas (mesh network over MQTT), mettax (video), and thermo-king (API).
In September, we plan to continue working in the same direction – maturing recently released features based on initial feedback and focusing increasingly on internal work.