Coding Harnesses

Most harnesses below work two ways. FireConnect writes the settings for you and restores them with off. Manual setup shows the same settings so you can apply them yourself, with no extra tool installed. Claude Desktop is FireConnect only.

Claude Code

Messages API

Claude Desktop

macOS only

Codex CLI, Codex app, and ChatGPT

Responses API

OpenCode

Chat Completions

Pi

Chat Completions

Cursor IDE

Chat Completions

VS Code

Chat Completions

Copilot App

Chat Completions

Copilot CLI

Chat Completions

DeepSeek Harness

Chat Completions

Before you start

FireConnect
Manual setup

Install FireConnect and sign in once:

fireconnect login

Connect: fireconnect <harness>. Add --model <id> to choose a model or router. Browse IDs with fireconnect model list.
Quit first: fully quit Cursor IDE, VS Code, and the Copilot App before connecting or running off. Restart other harnesses, including Claude Desktop, after connecting.
Check: fireconnect <harness> status is read-only and safe while the app is open.
Undo: fireconnect <harness> off restores the settings FireConnect saved under ~/.fireconnect/.

Create an API key in Fireworks settings and export it:

export FIREWORKS_API_KEY="fw_..."

Every harness uses the same Fireworks endpoints:

Harness API	Base URL
Anthropic Messages	`https://api.fireworks.ai/inference`
OpenAI Chat Completions or Responses	`https://api.fireworks.ai/inference/v1`

Use any model ID from Open Models, such as glm-latest, or a router from FireRouter, such as firerouter/opus. Back up a settings file before you edit it.

Claude Code

Anthropic Messages FireRouter Native web search

FireConnect
Manual setup

fireconnect claude
fireconnect claude status

Start a new session, or exit and run claude --resume <id>. FireConnect selects FireRouter by default. To choose another model, pass --model:

fireconnect claude --model glm-5p3-flash

--model sets the default model for new sessions and adds it to /model. /model also lists auto, the routers, and the Fireworks catalog. Your own picker rows are not changed.To send one Claude tier or your subagents to a specific model, pin that slot:

fireconnect claude --sonnet deepseek-flash-latest --subagent deepseek-flash-latest

Tier flags are --opus, --sonnet, --haiku, --fable, and --subagent. Unpinned slots stay on Claude Code’s defaults. Running fireconnect claude again keeps your pins. Pass native, such as --sonnet native, to release a slot.

.claude

settings.json

.claude.json

.fireconnect

See what FireConnect writes

~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.fireworks.ai/inference",
    "ANTHROPIC_CUSTOM_HEADERS": "X-Fireworks-Api-Key: fw_...",
    "DISABLE_TELEMETRY": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "ENABLE_TOOL_SEARCH": "true"
  },
  "modelPicker": {
    "options": [
      { "model": "firerouter[1m]", "label": "FireRouter" },
      { "model": "auto[1m]", "label": "Auto" }
    ]
  },
  "model": "firerouter[1m]",
  "statusLine": { "type": "command", "command": "... claude-statusline.mjs" }
}

The Fireworks key goes in a custom header, so your Claude login stays available for routes with Claude models.
[1m] marks 1M-context models.
A pinned tier adds ANTHROPIC_DEFAULT_<TIER>_MODEL to env, such as ANTHROPIC_DEFAULT_SONNET_MODEL. A pinned subagent adds CLAUDE_CODE_SUBAGENT_MODEL.
The status line shows the estimated session cost. It is added only if you do not already have one.
FireConnect also turns off Claude Code telemetry and nonessential traffic, and enables MCP tool search.
~/.claude.json records approval for the FireConnect credential so Claude Code does not prompt for it.

Point Claude Code at Fireworks

Add these values to ~/.claude/settings.json:

~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.fireworks.ai/inference",
    "ANTHROPIC_AUTH_TOKEN": "fw_YOUR_FIREWORKS_API_KEY"
  },
  "model": "glm-latest"
}

Or set them for one shell:

export ANTHROPIC_BASE_URL="https://api.fireworks.ai/inference"
export ANTHROPIC_API_KEY="$FIREWORKS_API_KEY"
claude --model glm-latest

Start a session

Run claude. Claude Code may ask once whether to use the API key; approve it.

The first reply comes from the Fireworks model you set.

Optional: use FireRouter

Set "model": "firerouter/opus". For Claude models in the route, connect an Anthropic Provider Key, or add x-anthropic-api-key to ANTHROPIC_CUSTOM_HEADERS.

Claude Code’s /model picker does not list Fireworks models you set by hand. Change models in settings.json or with claude --model <id>. Claude Code may log unrecognized_model for Fireworks IDs; requests still succeed.

Routers and credentials

Bare firerouter and routes with Claude models can use your Claude login when you connect with FireConnect.
Optional: save an OpenAI key once with fireconnect configure --openai-api-key sk-... so FireRouter can also reach GPT models. FireConnect sends it on bare firerouter and on routes that name a GPT model ID, such as firerouter/gpt-5.6-sol.
Family routes such as firerouter/sol and firerouter/astra don’t use the saved key. Connect an OpenAI Provider Key, or send x-openai-api-key with manual setup.
See Routing Preferences for --routing-preference and Harness Compatibility for cross-harness support.

Usage and cost

For the status line and fireconnect claude usage, see Usage and Cost. Claude Code’s own in-app cost uses Anthropic list prices; use the FireConnect status line for the Fireworks estimate. For a side-by-side comparison, see the Side-by-Side Demo.

Troubleshooting

Text-only models fail on pasted images. Use /rewind, or switch to a vision model such as glm-5p3-flash.
Resume or restart the session after changing models.
Auto mode says classifier requests are billed: this is expected, once per session, and auto mode keeps working. Its safety checks run on your Sonnet slot and bill to Anthropic. To bill them at Fireworks rates, pin Sonnet with fireconnect claude --sonnet <id>.

Claude Agent SDK

Anthropic Messages The Claude Agent SDK can reuse the Claude Code settings above. Connect Claude Code first, with FireConnect or manually:

fireconnect claude --model firerouter/opus

In the SDK query options, include settingSources: ["user"]. The SDK then reads the Fireworks endpoint, model, and credentials from ~/.claude/settings.json. fireconnect claude does not configure Claude Desktop. See Claude Desktop.

Claude Desktop

Anthropic Messages macOS only FireConnect runs Claude Desktop’s Chat, Cowork, and Code on Fireworks models. It requires macOS and a stored credential from fireconnect login. Quit and reopen Claude Desktop after on or off.

FireConnect

fireconnect claude-desktop on
fireconnect claude-desktop status

Reopen Claude Desktop. The picker starts on auto, followed by the Fireworks models, in place of your claude.ai subscription models.on brings your connectors along automatically. Each one asks you to sign in once the first time you use it. off puts your previous setup back, including your claude.ai models.

Claude Desktop model picker in Cowork showing Auto selected, then Kimi K3 Fast, Kimi K3, MiniMax M3, GLM 5.3 Fast, GLM 5.3, and DeepSeek V4.1 Flash — Claude Desktop after FireConnect, with Auto selected

See what FireConnect writes

on writes a third-party inference provider profile, Claude Desktop’s own supported mechanism. The profile points at a small local gateway that FireConnect runs as a launchd service. It starts at login and restarts after a crash.
on does not copy custom skills or plugins into third-party mode. Conversations stay in the mode they were made in.
Connector OAuth clients you register are saved in ~/.fireconnect/claude-desktop/oauth-clients.json, readable only by you. They survive off and on.
off restores your previous provider profile and the device settings FireConnect changed, then removes the local gateway service. Nothing of yours is deleted.

What moves with you

Connectors move automatically with on. Sign in to each one once the first time you use it.
Connector OAuth clients you register are saved and survive off and on.
Conversations, custom skills, and plugins stay where they are. Third-party mode does not show your claude.ai sessions, and your subscription does not show sessions you start while connected. Nothing is deleted.
MCPs that need an OAuth client ID and secret do not move. Create your own client, such as a Desktop app client in Google Cloud for the Google-hosted connectors. If you need one of these connectors, reach out to the team.

Run fireconnect claude-desktop mcp sync when you add a connector later and want it in third-party mode too.

Where did my previous conversations go?

Your conversations are still there. Conversations stay in the mode they were made in. Claude Desktop’s third-party mode does not show conversations from your claude.ai subscription, and your subscription does not show conversations you start while connected. Nothing is deleted. Run fireconnect claude-desktop off and reopen Claude Desktop to see your previous conversations again.

Connectors

Claude Desktop’s third-party mode hides the in-app connector browser. Manage connectors with FireConnect instead:

on imports the connectors you already use, from Desktop session history, your previous Desktop config, and your claude.ai organization when the Claude Code CLI is installed and signed in.
fireconnect claude-desktop mcp add <name> <url> and mcp remove <name> manage entries by hand. Removals stick across syncs.
Each connector asks you to sign in once the first time you use it after on.

Google-hosted connectors (Gmail, Google Drive, Google Calendar, BigQuery) need a Google OAuth client you own, created as a Desktop app client in Google Cloud. One client covers all four:

fireconnect claude-desktop mcp add gmail https://gmailmcp.googleapis.com/mcp/v1 \
  --client-id "<ID>.apps.googleusercontent.com" --client-secret "<SECRET>"

The hosted Figma connector does not work in third-party mode. Turn on the Figma desktop app’s local server in Dev Mode, then run fireconnect claude-desktop mcp add figma-desktop http://127.0.0.1:3845/mcp.

Troubleshooting

status shows the local gateway is not responding: run fireconnect claude-desktop on to restart it.
Models are missing from the picker: run on once while online, then restart Claude Desktop.
The Code tab says the API key was rejected: run on again to refresh the key from your stored credential.
A connector sign-in error won’t clear: mcp remove it, then mcp add it again with your own OAuth client (Google) or the local server (Figma).

Codex

OpenAI Responses FireRouter Native web search The Codex CLI, the Codex app, and the ChatGPT desktop app share one config in ~/.codex/config.toml, so one setup covers all three. Quit the Codex app and the ChatGPT app before you change it so their model lists refresh.

MCP servers and plugins in the ChatGPT desktop app don’t work reliably while FireConnect is on. See Harness Compatibility.

FireConnect
Manual setup

fireconnect codex
fireconnect codex status

The default model is auto. To add FireRouter to the Codex and ChatGPT model pickers, run fireconnect codex --model firerouter. Exit Codex and run codex resume <id>, or start a new session.

~/.codex

config.toml

fireworks-model-catalog.json

See what FireConnect writes

~/.codex/config.toml

model_provider = "fireworks-ai"
model_catalog_json = "~/.codex/fireworks-model-catalog.json"
model = "auto"

[model_providers.fireworks-ai]
name = "Fireworks"
base_url = "https://api.fireworks.ai/inference/v1"
wire_api = "responses"
experimental_bearer_token = "fw_..."
requires_openai_auth = false

FireConnect keeps your other TOML settings, saves the key with file mode 0600, and writes a model catalog so Codex knows each model’s context window and capabilities. The catalog also makes Codex auto-review use the model you picked, so approvals are reviewed on Fireworks. Connected before FireConnect 0.9.8? Run fireconnect codex once, or auto-review keeps denying every approval.

Add a Fireworks provider

~/.codex/config.toml

model_provider = "fireworks-ai"
model = "glm-latest"

[model_providers.fireworks-ai]
name = "Fireworks"
base_url = "https://api.fireworks.ai/inference/v1"
wire_api = "responses"
env_key = "FIREWORKS_API_KEY"

Run Codex

codex

Codex answers from glm-latest.

Without a model catalog, Codex warns that it has no metadata for Fireworks models and uses fallback limits. FireConnect writes that catalog for you.

ChatGPT desktop model picker showing FireRouter selected alongside Auto and Fireworks models — ChatGPT desktop with FireRouter selected

When you resume an old session, select the provider explicitly:

codex resume <id> -c model_provider="fireworks-ai"

MiniMax models don’t work on Codex. Codex may insert assistant messages between tool_calls and tool_results, which MiniMax templates reject. Use a Chat Completions harness such as OpenCode for MiniMax.

OpenCode

OpenAI Chat Completions FireRouter

FireConnect
Manual setup

fireconnect opencode
fireconnect opencode status

Restart OpenCode. The default model is auto, stored as fireworks-ai/auto. Short IDs such as glm-5p2 expand to full Fireworks paths.

OpenCode model picker showing Auto selected and Fireworks DeepSeek, GLM, Kimi, and other models — OpenCode model picker after FireConnect, with Auto selected

See what FireConnect writes

~/.config/opencode/opencode.json

{
  "provider": {
    "fireworks-ai": {
      "options": { "apiKey": "fw_..." },
      "models": {
        "auto": { "name": "Auto", "limit": { "context": 1048575, "output": 131072 } },
        "glm-latest": { "name": "GLM (Latest)" }
      }
    }
  },
  "model": "fireworks-ai/auto"
}

The file is saved with mode 0600. FireConnect never changes auth.json. Use --config-path for a config stored elsewhere.

Fastest: in OpenCode, run /connect, search fireworks.ai, paste your key, and pick a model with /models.Or edit the config to add routers and pin a default:

~/.config/opencode/opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "fireworks-ai": {
      "options": { "apiKey": "{env:FIREWORKS_API_KEY}" },
      "models": {
        "glm-latest": { "name": "GLM (Latest)" },
        "firerouter/opus": { "name": "FireRouter Opus" }
      }
    }
  },
  "model": "fireworks-ai/glm-latest"
}

opencode run -m fireworks-ai/firerouter/opus "hello" answers through FireRouter.

Pi

OpenAI Chat Completions FireRouter

FireConnect
Manual setup

fireconnect pi
fireconnect pi status

Restart Pi if it was running. The default model is auto.

~/.pi/agent

settings.json

auth.json

models.json

See what FireConnect writes

~/.pi/agent/settings.json

{
  "defaultProvider": "fireworks",
  "defaultModel": "auto",
  "enabledModels": ["fireworks/accounts/fireworks/routers/*", "fireworks/auto"]
}

~/.pi/agent/auth.json

{ "fireworks": { "type": "api_key", "key": "fw_..." } }

models.json adds the Fireworks catalog to Pi’s built-in fireworks provider. FireConnect tracks the IDs it adds so off removes only those. Use --settings-path for a settings file stored elsewhere.

Pi has a built-in fireworks provider that reads FIREWORKS_API_KEY. Set it as the default:

~/.pi/agent/settings.json

{
  "defaultProvider": "fireworks",
  "defaultModel": "accounts/fireworks/routers/glm-latest"
}

pi -p "hello" answers from GLM.

Cursor IDE

OpenAI Chat Completions FireRouter with Provider Keys fireconnect cursor configures Cursor IDE. Cursor CLI (agent or cursor-agent) is not supported. Fully quit Cursor IDE before connecting or running off.

FireConnect
Manual setup

fireconnect cursor
fireconnect cursor status

FireConnect points every existing mode in modelConfig at the Fireworks default and registers the catalog in the picker. Reopen Cursor IDE and choose a Fireworks model.

Cursor IDE model picker showing Auto selected with Fireworks DeepSeek, GLM, Kimi, and MiniMax models — Cursor IDE after FireConnect, with Auto selected

See what FireConnect writes

Cursor IDE stores AI settings in SQLite (state.vscdb under ~/.config/Cursor, ~/Library/Application Support/Cursor, or %APPDATA%\Cursor):

Setting	Value
`cursorAuth/openAIKey`	Your Fireworks key
`openAIBaseUrl`	`https://api.fireworks.ai/inference/v1`
`aiSettings.userAddedModels`	The Fireworks catalog, tracked for a clean `off`
`aiSettings.modelOverrideDisabled`	Cursor’s built-in models, hidden while connected
`aiSettings.modelConfig[<mode>]`	The Fireworks model for each mode

Your previous auth state is saved under ~/.fireconnect/cursor/. Use --db-path for a non-default state.vscdb.

Open model settings

In Cursor IDE, open Settings, then Models.

Add your Fireworks key

Under API Keys, paste your Fireworks key into OpenAI API Key. Turn on Override OpenAI Base URL and enter https://api.fireworks.ai/inference/v1.

Add a model

Add a custom model with a Fireworks ID, such as glm-latest, then select it in chat.

While connected, only Fireworks models work. Cursor IDE’s built-in models are hidden until you run off. Cursor IDE also enforces a server-side allowlist, so not every Fireworks model is selectable. Some features may still use Cursor’s own backend, depending on plan and version.

VS Code

OpenAI Chat Completions FireRouter FireConnect adds a Fireworks provider to GitHub Copilot Chat’s custom endpoints. This requires Copilot Pro or Enterprise. Quit VS Code before connecting or running off.

FireConnect
Manual setup

fireconnect vscode
fireconnect vscode status

Restart VS Code, then pick a model under Other Models → Fireworks in Copilot Chat. glm-latest and glm-fast-latest are text-only; glm-5p3-flash supports images.

VS Code Chat model picker showing the Fireworks provider with Auto selected and DeepSeek, GLM, and Kimi models — VS Code Chat after FireConnect, with Auto selected under Fireworks

See what FireConnect writes

chatLanguageModels.json

[
  {
    "name": "Fireworks",
    "vendor": "customendpoint",
    "apiType": "chat-completions",
    "apiKey": "${input:chat.lm.secret.fw-...}",
    "models": [
      {
        "id": "auto",
        "name": "Auto",
        "url": "https://api.fireworks.ai/inference",
        "maxInputTokens": 1048575,
        "maxOutputTokens": 131072,
        "vision": true,
        "toolCalling": true
      }
    ]
  }
]

The key is stored separately in state.vscdb, encrypted with Electron safeStorage. On Linux, encryption needs libsecret; without it, VS Code only obfuscates the key and FireConnect warns you.

Platform	`chatLanguageModels.json`	`state.vscdb`
Linux	`~/.config/Code/User/`	`~/.config/Code/User/globalStorage/`
macOS	`~/Library/Application Support/Code/User/`	`~/Library/Application Support/Code/User/globalStorage/`
Windows	`%APPDATA%\Code\User\`	`%APPDATA%\Code\User\globalStorage\`

In Copilot Chat, open Language Models, click + Add Models…, and choose Custom Endpoint. Use https://api.fireworks.ai/inference/v1, your Fireworks key, and a model ID such as glm-latest.Follow the step-by-step screenshots in GitHub Copilot.

Copilot App

OpenAI Chat Completions FireRouter with Provider Keys The GitHub Copilot desktop app keeps its built-in models and adds Fireworks alongside them. Quit the app before connecting or running off.

FireConnect
Manual setup

fireconnect copilot-app
fireconnect copilot-app status

Reopen the app and pick a Fireworks model in Settings → Model providers.

See what FireConnect writes

FireConnect writes to ~/.copilot/data.db (SQLite, movable with COPILOT_HOME):

Table	What FireConnect adds
`model_providers`	One `Fireworks` provider of type `openai`, with base URL `https://api.fireworks.ai/inference/v1` and your key as an `Authorization: Bearer` header
`provider_models`	One row per model, with token limits and reasoning efforts (`low`, `medium`, `high`, `max`)

off deletes the provider and its models. Use --db-path for a non-default database.

In Settings → Model providers, add a provider with the values FireConnect uses:

Field	Value
Base URL	`https://api.fireworks.ai/inference/v1`
API key	Your Fireworks key
Models	Fireworks IDs, such as `glm-latest` or `kimi-latest`

The app’s bring-your-own-key path does not support image input, the hover-card Context row, or AI-credits pricing. FireConnect can add any FireRouter ID, but the app cannot send a local Anthropic or OpenAI key. Without a matching Provider Key, FireRouter serves those routes with open models. See Harness Compatibility.

Copilot CLI

OpenAI Chat Completions FireRouter with manual setup The copilot command (@github/copilot) reads plain JSON under ~/.copilot, separate from the desktop app.

FireConnect
Manual setup

fireconnect copilot-cli
fireconnect copilot-cli status

Start a new copilot session. Switch models with copilot --model fireworks/<name> or /model.

Copilot CLI model picker showing Auto selected and Fireworks DeepSeek, GLM, Kimi, and MiniMax models — Copilot CLI model picker after FireConnect, with Auto selected

off restores providers.json from its snapshot, or deletes it if FireConnect created it. Use --providers-path for a non-default location.

~/.copilot/providers.json

{
  "providers": [
    {
      "name": "fireworks",
      "type": "openai",
      "wireApi": "completions",
      "baseUrl": "https://api.fireworks.ai/inference/v1",
      "apiKey": "fw_YOUR_FIREWORKS_API_KEY"
    }
  ],
  "models": [
    {
      "id": "glm-latest",
      "provider": "fireworks",
      "wireModel": "glm-latest",
      "name": "GLM (Latest)",
      "maxPromptTokens": 1048576,
      "maxContextWindowTokens": 1048576,
      "maxOutputTokens": 131072,
      "capabilities": { "supports": { "reasoningEffort": true } }
    }
  ]
}

~/.copilot/settings.json

{ "model": "fireworks/glm-latest" }

For a FireRouter route with Claude or GPT models, add a headers map to the provider, such as "headers": { "x-anthropic-api-key": "sk-ant-..." }, and use the router ID, such as firerouter/opus, as the model id and wireModel. FireConnect adds these routes but does not send your provider keys, so use manual setup to send them yourself. The harness setup generates the full file.

Use provider-prefixed model IDs such as fireworks/glm-latest. The CLI rejects a bare glm-latest and silently falls back to its configured model.

DeepSeek Harness

OpenAI Chat Completions FireRouter with manual setup DeepSeek’s coding agent (dsh) runs named profiles, such as tui, web, and headless, under $DSH_HOME, which defaults to ~/.dsh. Restart dsh after changing a profile.

FireConnect
Manual setup

fireconnect deepseek
fireconnect deepseek status

Restart dsh. It starts on auto and goes straight to Fireworks, with no DeepSeek key prompt.FireConnect adds a fireworks provider to ~/.dsh/settings.yaml and to each existing profile’s cordis.patch.yml, which dsh 0.1.7 and later read. off restores both from the snapshot under ~/.fireconnect/deepseek/.

~/.dsh

settings.yaml

.credentials.yaml

profiles/<profile>

cordis.patch.yml

Connected before FireConnect 0.9.8? Run fireconnect deepseek once so dsh 0.1.7 and later stop asking for a DeepSeek key.FireConnect can add any FireRouter ID, but it does not send provider keys from dsh. To use Claude or GPT models in a route, connect a Provider Key or use manual setup.

Create the profile folder once, then add a patch to it. Repeat for each profile you use.

dsh --profile tui --dump-config > /dev/null

~/.dsh/profiles/tui/cordis.patch.yml

- id: llm-pi-ai
  config:
    providers:
      fireworks:
        displayName: Fireworks
        apiKeyEnv: FIREWORKS_API_KEY
        api: openai-completions
        baseURL: https://api.fireworks.ai/inference/v1
        models:
          - id: glm-latest
            name: GLM (Latest)
            reasoning: true
            contextWindow: 1048576
            maxTokens: 131072
- id: agent-default-model
  config:
    provider: fireworks
    model: glm-latest

export FIREWORKS_API_KEY="fw_..."
dsh tui

For a FireRouter route with Claude or GPT models, add headers to the provider, such as x-anthropic-api-key: sk-ant-..., and use the router ID as the model. The harness setup generates the full patch.

Nexus

Connect

Choose models

Operate

Reference

Claude Code

Claude Desktop

Codex CLI, Codex app, and ChatGPT

OpenCode

Pi

Cursor IDE

VS Code

Copilot App

Copilot CLI

DeepSeek Harness

Before you start

Claude Code

Claude Agent SDK

Claude Desktop

Codex

OpenCode

Pi

Cursor IDE

VS Code

Copilot App

Copilot CLI

DeepSeek Harness

Claude Code

Claude Desktop

Codex CLI, Codex app, and ChatGPT

OpenCode

Pi

Cursor IDE

VS Code

Copilot App

Copilot CLI

DeepSeek Harness

​Before you start

​Claude Code

​Claude Agent SDK

​Claude Desktop

​Codex

​OpenCode

​Pi

​Cursor IDE

​VS Code

​Copilot App

​Copilot CLI

​DeepSeek Harness

Before you start

Claude Code

Claude Agent SDK

Claude Desktop

Codex

OpenCode

Pi

Cursor IDE

VS Code

Copilot App

Copilot CLI

DeepSeek Harness