MeshAgent 0.22.0 adds Anthropic support, automatic context compaction for OpenAI models, expanded secrets and service templating support, agent containerization improvements, and agent identity simplification. Check out the release highlights below!
MeshAgent now supports Anthropic through the meshagent-anthropic package and ships an AnthropicMessagesAdapter which allows you to use Anthropic models in a MeshAgent Room. MeshAgent handles routing, pricing validation, usage tracking, and MCP tooling for both OpenAI and Anthropic models. This gives you flexibility to choose the model you prefer for different tasks and switch between providers as needed. Both OpenAI and Anthropic models can be used in the same Room.
If you are using agent frameworks like LangChain, PydanticAI, CrewAI, or others, MeshAgent's room router allows you to plug them in and take advantage of both OpenAI and Anthropic models through our proxy infrastructure.
Long conversations eventually hit model context window limits. MeshAgent now handles this automatically for OpenAI models by using OpenAI's native compaction API. The OpenAIResponsesAdapter tracks token usage from each response and checks whether your next request would exceed the model's context window (reserving space for max_output_tokens). If it would overflow, MeshAgent automatically calls OpenAI's responses.compact endpoint, which shrinks the conversation history while preserving important context. The compacted context replaces the original messages, and your request proceeds without hitting the context window limit.
This all happens automatically when using the OpenAIResponsesAdapter so you don't need to manually call compaction or implement custom summarization logic.
For other LLM providers, MeshAgent exposes needs_compaction() and compact() hooks at the LLMAdapter level if you want to implement custom compaction strategies.
MeshAgent now has APIs to create, update, and validate services directly from ServiceTemplates.
A ServiceSpec is a concrete, runnable configuration where all values are fixed. A ServiceTemplateSpec is parameterized, it can define variables that get filled in at deploy time, when MeshAgent converts the ServiceTemplateSpec to the runnable ServiceSpec. For example, when sharing agents in Powerboards, you send someone an install link that references a ServiceTemplate this allows others to install the agent and provide their own values for any requested variables (e.g. an email address for the agent to use).
Before this release, you could only use the CLI and API to create services from ServiceSpecs. Now ServiceTemplateSpec can be directly deployed as services through both the CLI and API.
Services also receive their configuration as a mounted file accessible via the MESHAGENT_SPEC_PATH environment variable, giving your service access to its own spec at runtime.
MeshAgent now supports request and provide workflows for secret sharing between participants. Agents can now request specific credentials only when needed and your app can approve and provide them when required. This ensures access is scoped tightly and credentials are only provided when needed. The MeshAgent CLI now exposes secret commands alongside OAuth flows.
meshagent join now works inside room containers by setting MESHAGENT_TOKEN as an environment variable. This gives you a lightweight path for running always-on agents in a Room, the agent will connect to the Room as a participant and run. We recommend using meshagent join for agents that are running in the Room. For agents hosted outside the room you will need to use meshagent service which runs your agent as an HTTP service that MeshAgent can call on demand and manage its lifecycle and health.
Join our Discord Server to stay up to date with MeshAgent releases, ask questions, and share feedback with our team.
Check out the MeshAgent documentation to start building today!
MeshAgent Studio, SDK, and Server give you everything to build, test, and deploy agentic applications, from development to production.
