MESH

Coding agents · English guide · Updated 15 November 2026

Use MeshCoder with T3 Code

Connect T3 Code to a MESH model running on compatible hardware you own. Start with a separate Codex provider instance, keep your existing setup, and verify one completed task before trusting it with a project.

MESH does not charge for generation on your own connected hardware. Available memory, model downloads, electricity and runtime still limit use. T3, editor and external provider costs are separate.

Before you start

Configure a separate T3 provider

  1. In T3, open Settings → Providers → Add provider → Codex → Configure manually. Name the instance MeshCoder and select your installed Codex binary.
  2. Use a new, dedicated CODEX_HOME directory. Leave Shadow home empty. Do not overwrite your normal Codex configuration or copy its credentials. Separate homes do not continue your normal Codex threads.
  3. Save the Responses-provider config.toml from MESH’s T3 guide inside that dedicated directory. Select model meshcoder. The T3 route ends in /platforms/t3-code/v1.
  4. After creating the instance, add the matching key as a Sensitive provider environment variable named MESH_API_KEY. Keep it out of launch arguments, screenshots, shared configuration and git.
  5. Hide stock OpenAI models only in this instance, add/select the custom meshcoder model and start a new session. Keep MESH running with an available device.

Choose the address for the agent’s machine

Same machine: use MESH’s authenticated loopback endpoint and its local mlk_ key. “Localhost” refers to the computer running T3’s agent process, which may differ from the computer showing T3’s interface.

Remote environment: it cannot reach your computer’s localhost. Public access is a separate, explicit MESH setting. If you choose it, use the reachable public HTTPS endpoint and its mpk_ key. Local and public keys are not interchangeable. Public access involves hosted routing; it is not a fully local connection.

Check the result, then allow a small task

First ask a file-free question. Wait for T3 to mark the turn completed; text appearing on screen alone is not success. Then use a disposable project, review the requested tool permissions, and ask for one small change. Inspect the diff and run the relevant check yourself. The coding client controls file and shell operations.

If a request does not complete, check the selected model, device availability, endpoint reachability, key class and request size. Responses are currently buffered until generation finishes, so silence during inference is not token streaming. Review the error before retrying.

Direct MeshCoder, hybrid and MCP are different

What has been verified?

On 5 November 2026, T3 Beta 0.1.7 with Codex 0.166.0 completed a file-free inference turn and a harmless shell command through authenticated MESH QA loopback on one Mac. This is narrow QA evidence, not a current production compatibility certificate, a coding-quality benchmark, or verification of remote environments and other operating systems.

Direct coding support is experimental. The complete request, including instructions, tool definitions and reply, must fit the admission ceiling shown in MESH. Tool-call accuracy and provider-specific features are not universally supported. Review generated code and permissions.

Workload APIs and privacy

For supported catalog jobs, open a model in the Solutions Hub and choose Show API details. Its /api child page contains authentication, request fields, cURL and status polling. Those REST jobs use MESH’s hosted control plane and your connected devices. The Private setting controls storage; it is not a guarantee of local-only routing.

Sources and help

Use MESH’s in-app T3 guide as the source for your current configuration. T3’s Codex provider guide explains provider instances and isolated homes. Its Claude provider guide describes that separate integration.

Published by Mesh Labs. Questions: support@meshlabs.cloud.

Get MESH and open the T3 Code setup guide →