Developers
For developers: extensible, open, and lock-in-free
Charla is built to be extended, integrated, and observed, an open agentic platform, with no vendor silos, that you adapt to your stack instead of adapting your stack to it.
If you're going to integrate an agentic platform into your product, three questions matter: what it locks you into, how you extend it, and how much you can see of what happens inside. Charla was designed with clear answers to all three: a swappable AI model, interoperability through an open standard, and a real-time event stream you can follow from start to finish.
Charla is model-agnostic
In Charla, the AI model is a swappable layer, not the foundation of the platform. That means you aren't locked to a single model provider: the layer can be replaced without rewriting the application around it. The current generation runs on Google's Gemini via Vertex AI, but that choice is an engineering decision, not an architectural constraint.
For integrators, the payoff is practical: your business logic, your integrations, and your interface talk to the platform, not to the quirks of one specific model provider.
Open interoperability via MCP
Charla connects assistants to external tools and systems through MCP (Model Context Protocol), an open standard. Charla acts as an MCP client: it consumes the tools you expose, rather than requiring you to program against a proprietary, closed API. Tools are namespaced per connection, which keeps it clear where each capability comes from and avoids collisions between integrations.
Betting on an open standard is deliberate: an agentic platform should be open and observable, with no vendor silos. By speaking MCP, Charla fits into a tooling ecosystem you can also use elsewhere.
Custom skills, with no platform code changes
Where MCP covers the connection to external systems, custom skills cover extending behavior. A skill packages instructions and reference files into a capability the assistant loads on demand: when a task calls for that knowledge, it uses it; when it doesn't, it carries no extra weight.
- You extend an assistant with no changes to the platform's code.
- The capability is loaded on demand, only when the task warrants it.
- Instructions and reference files travel together, as a versionable unit.
Streaming responses, over a structured event stream
Charla's responses arrive in real time, over a structured event stream, rather than as a single block of text at the end. Each event type carries its own meaning, which lets you build interfaces that react as the agent works: show a status indicator, stream tokens as they're generated, signal a tool call in progress, or render an artifact the moment it's ready.
- status: signals about what the agent is doing right now.
- tokens: the response text, streamed as it's generated.
- tool calls: when the agent invokes a tool.
- artifacts: structured outputs, such as tables and charts, ready to render.
- final event: the clean close of the response cycle.
The stream exposes what your interface needs to react (status, tokens, tools, artifacts), but never the model's raw internal reasoning.Charla · notes for integrators
Built-in tools you toggle on and off
Beyond the external tools you connect, Charla ships a set of built-in tools that can be turned on or off per assistant: you expose only what makes sense for each use case.
- Web search, to bring in current information from the world.
- Current time, for time-aware answers.
- Image generation and chart rendering.
- Structured data export.
- Knowledge search and table search.
Portable deployment, with no cloud lock-in
Charla's stack runs in containers behind a load balancer and is portable across clouds or on-prem. The only cloud-specific component is the load balancer itself: everything else travels with you. There is no lock-in to a single infrastructure provider.
Cost is accounted per call, in a transparent credit-based model: you see what each interaction consumes, with no opaque charges hidden behind the platform.
Taken together, these decisions point the same way: an agentic platform you adapt to your stack, integrate through open standards, and observe while it works, rather than a black box your product has to submit to.