How to run Claude Code, Codex or OpenCode on a local model with strata connect

Our earlier Strata guides covered running Qwen3.8-Flash-Next and DeepSeek V4.1-Flash on a gaming PC. This one is about the next step: getting a terminal coding agent to talk to that local model. Everything below was checked against Strata's docs and source for version 0.2.8, released October 29, 2026.

What does strata connect do?

It writes the config file and environment variables each coding app needs to use your Strata server, with the address, key, model name and context size filled in. It knows eight apps: OpenCode, Codex CLI, Claude Code, oh-my-pi, pi, Hermes Agent, OpenClaw and Strata's built-in DeepSeek Harness. Run strata connect on its own to list them. Each app gets three modes:

strata connect opencode           # print the file and environment
strata connect opencode --write   # write it after asking; anything replaced is backed up
strata connect opencode --check   # send one prompt and one tool call the way that app does

Swap opencode for codex or claude-code. Without flags it reads the settings from the model you launched; otherwise it assumes http://127.0.0.1:8080 and a 32K context. You can override any of it with --url, --key, --model and --context. If you prefer the monitor screen, press 7 for Connect, then c to copy a setup, w to write it or h to check it.

Step 1: install Strata and start a coding model

Strata runs on Windows, Linux and Apple Silicon Macs. NVIDIA cards need driver 580 or newer for its CUDA 13 packages. The install needs no admin rights:

# Linux or macOS
curl -fsSL https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh && sh install.sh

# Windows (PowerShell), then open a new terminal
irm https://github.com/Niko1221/Strata/releases/latest/download/install.ps1 | iex

For coding agents, Strata's model guide offers ISTA-DASLab's Coder build of Qwen3.8-Flash-Next, which keeps 256 of each layer's 512 experts. Its authors report 91% of the full model's SWE-bench Verified score and 99% of its LiveCodeBench score. Its first file is 29.6 GB, so a 32 GB RAM PC can run it if the GPU has enough memory (Strata's launch plan tells you), and it is weaker than the full model outside of code. Start it with room for an agent:

strata launch --family coder --model IQ1_M --context 64k

Strata measured this model at about 55 tokens per second in short chat on an RTX 5070 12 GB, and 43 at 128K context. Those are Strata's own numbers on an older engine build, not ours. Skip DeepSeek V4.1-Flash for agent work: its context is fixed at 8K, which an agent's prompt and tools nearly fill on their own.

Step 2: connect your coding agent

Print first, write second, check third. Run strata connect claude-code (or codex, opencode) and read the output. Then add --write, confirm, and finish with --check. The check exits with code 0 only when both the prompt and the tool call work, and --output json gives you the details.

Before it writes, Strata copies any existing file to <name>.strata-bak-<date-time> and makes the new file readable only by you, since it can hold the API key. Every config was tested against a Strata server at 32K context. App by app:

• Claude Code talks to Strata's Anthropic-style /v1/messages endpoint. The base URL is the server address without /v1; with it, requests go to /v1/v1/messages. No Anthropic login is used. Strata also sets CLAUDE_CODE_MAX_CONTEXT_TOKENS and a 4,096-token output cap, because without them Claude Code assumes a 100K window and compacts at the wrong time. At 40K context or less it adds CLAUDE_CODE_SIMPLE=1, which cuts Claude Code to 3 tools and about 1.25K tokens of prompt. It also turns off Claude Code's telemetry, auto-updater and other non-essential traffic.
• Codex CLI uses the Responses API. It will not send anything without STRATA_API_KEY set; any text works when your server has no key. The message “Model metadata ... not found” on every start is harmless.
• OpenCode uses the OpenAI-compatible chat endpoint. It needs limit.input, which Strata sets just under your context (28,672 at 32K); without it OpenCode never compacts. Its first start downloads a plug-in from npm, so it needs internet once.

What it changes on your machine

These are your apps' real config files, so the change is not limited to one project or one terminal. Three things worth knowing before you press yes:

1. Claude Code: every project switches. ~/.claude/settings.json is Claude Code's user-wide settings file, so after a write every Claude Code session on that account goes to Strata, not just the one you are testing. If you also use Claude Code with your Anthropic plan, use the printed output instead and set the variables in one terminal only, or keep the backup handy.
2. Codex: approvals are set to never. The written config.toml sets approval_policy = "never" with sandbox_mode = "workspace-write", so Codex runs commands without asking, inside its sandbox. Because the file is replaced rather than merged, your other Codex providers and settings are gone until you restore the backup. Where the sandbox cannot start (no bubblewrap on Linux), Strata's own advice is to bypass approvals and the sandbox only in a throwaway folder. We would add: not with a model you have not vetted. Our post on Cohere's Qwen finding is why.
3. There is no undo command. To go back, rename the .strata-bak- file over the current one. For Claude Code you can also delete the env keys Strata added.

Strata is clear that connected apps are separate programs with their own network behavior, and it does not control what they send. Its engine has no networking code and its server listens on 127.0.0.1, but if you open it to your network with --access lan or a tunnel, keep the API key on: without one, anyone who can reach the port can use the model and its tools.

How much context do you need?

32K is the floor Strata tested; 64K is more comfortable for longer sessions. At 32K, Claude Code's full prompt would take about half the window, which is why Strata switches it to the slimmer 3-tool mode. Codex and OpenCode leave more room. To change the size, restart the model with --context (for example --context 64k --restart) and run strata connect again so the apps pick up the new limit. A bigger window costs memory and speed, so raise it only as far as your sessions need.

Using Strata with Bodega One Code

strata connect does not list our app, and it does not need to. Bodega One Code takes any OpenAI-compatible server through its Custom OpenAI-Compatible provider slot, on top of its 10+ BYOLLM presets. Set the base URL to http://127.0.0.1:8080/v1 (note the /v1 this time), enter any value for the API key unless you set one in Strata, and type the model name Strata reports. Use the Custom slot rather than the llama.cpp preset: that preset also defaults to port 8080 and can try to start its own server.

With the model on your PC and air-gap mode on, Bodega blocks the agent's web tools, network commands, updates and cloud features, and in Ask mode changes wait for your approval. We have not benchmarked Bodega One Code on Strata, and we are not claiming numbers. If you want to keep Claude Code as well, Bodega can run it as an agent inside a chat over ACP (ACP agents are switched off while air-gap mode is on); our Claude Code alternatives page covers how that works and its limits.

Sources

• Strata strata connect and launch documentation: github.com/Niko1221/Strata/blob/main/docs/TERMINAL_LAUNCH.md
• Strata connect config generators (files written, backups, per-app limits): github.com/Niko1221/Strata/blob/main/strata_cli/launch/harness.py
• Strata model guide (Coder IQ1_M sizes and speeds): github.com/Niko1221/Strata/blob/main/docs/MODELS.md
• Strata safety and privacy page: github.com/Niko1221/Strata/blob/main/docs/SAFETY.md
• Strata 0.2.8 release notes, October 29, 2026: github.com/Niko1221/Strata/releases/tag/v0.2.8
• Claude Code LLM gateway docs: code.claude.com/docs/en/llm-gateway