mesh

MIT licensed · public release in preparation

Your runtime.
Your infrastructure.

Own the harness your agents depend on. Inspect the MIT-licensed core, deploy it on infrastructure you control, and choose your inference providers. Texture hosting is an option, not a runtime dependency.

The source isn’t public yet.

The repository and container image are currently access-restricted while we prepare the public release. There isn’t a public download or quickstart available yet.

Ask about source access ↗

Model portability

The model will change.
Keep your harness.

Choose supported models through OpenRouter or Ramp Router, connect directly to OpenAI, or use an operator-hosted OpenAI-compatible endpoint. Bring your own keys and keep model selection in your hands.

For private inference, connect to servers such as Ollama, vLLM, or llama.cpp with a compatible, tool-capable model. You operate the server and choose the open-weight model, hardware, and capacity. A failed private endpoint does not silently fall back to a cloud provider.

The compatible endpoint is configured per deployment. Model capabilities and tool support depend on the server and model you select.

Private networks & airgapped environments

Bring inference inside.
Plan the whole boundary.

Self-hosting and local inference give you a foundation for keeping agent work inside your environment. A fully airgapped deployment also needs internal messaging, execution, authentication, staged dependencies, and an offline update process.

Local inference is supported today. The complete offline deployment profile, local execution worker, and qualified installation bundle are in development. The hosted service and cloud connectors are not an airgapped configuration.

Discuss a private or airgapped deployment ↗
MESH / RUNTIME MAPARCHITECTURE, NOT A TERMINAL
mesh/
  • identity & actors
  • conversation graph
  • perspective & memory
  • turn runtime
  • models & tools
  • connectors
  • execution backends
Conceptual modules
SlackTelegramMessaging connectors
Mesh runtimeIdentity → conversation → context → judgment
PostgreSQLDurable stateModel providerInferenceWorkspace backendOptional execution

Know what you’re taking on

Small core.
Real operational responsibilities.

Mesh is a Go service with an embedded web UI and PostgreSQL-backed state. Self-hosting means you operate that service—and choose the providers your agents use.

Prefer managed infrastructure? →
Compute & database
A host for the runtime and PostgreSQL with the pgcrypto extension. You manage updates, availability, backups, and recovery.
Connector connectivity
Cloud messaging connectors need network access. Slack requires a reachable event URL. Private-network deployments need a connectivity plan; local inference alone does not remove connector dependencies.
Secrets & model access
A master encryption key held outside the database, plus your model and connector credentials. Provider usage is billed to you.
Optional execution
Configure a supported remote workspace provider if agents need code or repository tools. Execution is separate from inference; a local model endpoint does not provide a local shell backend.
FIRST BOOTILLUSTRATED STATE
0agents
and that’s valid.
Your first coworkerPersona · model · connector
Create agents through the UI without restarting the process. No default agent is baked into the deployment.

No blessed bot

Start empty.
Build your team.

Agent setup is runtime state, not a pile of environment variables. Add another agent with its own persona, model configuration, and connectors through the dashboard.

Agents can share an installation while keeping separately owned identity, conversation, and memory state. The process hosts the team; it doesn’t merge their perspectives.

The separation that matters

Hosted is an option.
Not a dependency.

The core runtime does not need Texture accounts, SaaS billing, or hosted provisioning to do its job. Those concerns live in a separate service that operates Mesh instances.

Connectors, model providers, and execution backends sit behind explicit interfaces. Extending them is code-level integration work today—not a promise of an already-shipped plug-in marketplace.

Want to review the source before the public release? Contact Texture and mention Mesh. Access and timing are not guaranteed.

READ THE DESIGNWhy the room comes firstIdentity, perspective, and differentiation →Follow the runtimeTurns, memory, records, and roadmap →Understand the boundariesData ownership and current limitations →