Coding Harnesses

Connect Claude Code, OpenCode, Codex, Pi, Cursor IDE, VS Code, Copilot, or DeepSeek Harness to Fireworks

Install FireConnect and run fireconnect login once. Then choose your harness below. Each section shows the connect command, the settings FireConnect writes, and how to apply a new --model.

Shared rules

• Auth: fireconnect login once, then fireconnect <harness>. You do not need --api-key on connect when a key is already saved.
• Quit first: fully quit Cursor IDE, VS Code, and the Copilot App before connecting or running off. Restart other harnesses after connect.
• status is read-only while the app is open.
• Models: Pass --model <id> when you connect. Browse IDs with fireconnect model list. See FireRouter for router behavior.
• Web search: Claude Code, Codex, and ChatGPT keep native search. See Web Search for details.

Claude Code

fireconnect login
fireconnect claude                  # appends the catalog to /model
fireconnect claude status

Connection and model changes take effect in a new or resumed session. Exit and resume with claude --resume <id>, or start a new session.

Model picker. FireConnect adds auto, model routers, and Fireworks models to /model. Native Anthropic tiers remain unchanged. --model adds a row; it does not pin the default:

fireconnect claude --model firerouter
fireconnect claude --model glm-5p3-flash

What gets written

• Fireworks key in ANTHROPIC_CUSTOM_HEADERS (X-Fireworks-Api-Key), backup under ~/.fireconnect/ for off.
• Serverless rows in /model; [1m] tag on 1M-context models.
• Optional statusLine unless you already have one.
• Native web search and fetch stay on.

Usage. For the status line and fireconnect claude usage, see Usage and Cost. For a side-by-side model comparison, see the Side-by-Side Demo.

Routers

• Bare firerouter and routes containing a Claude family alias or Anthropic model ID can use your Claude login.
• firerouter/astra requires an OpenAI key configured in Provider Keys.
• See Harness Compatibility for --routing-preference and --anthropic-api-key.

Troubleshooting

• Text-only models fail on pasted images. Use /rewind.
• Claude Code calculates its in-app cost with Anthropic list prices. Use the FireConnect status line for the Fireworks estimate.
• Resume or restart the session after changing models.

OpenCode

fireconnect login
fireconnect opencode
fireconnect opencode status

Restart OpenCode after connecting. With a standard Fireworks key, the default model is auto, stored as fireworks-ai/auto. FireConnect expands short model IDs to full Fireworks paths automatically.

OpenCode model picker after FireConnect, with Auto selected

What gets written

• A fireworks-ai provider using the OpenAI-compatible endpoint at https://api.fireworks.ai/inference/v1
• The default model and serverless /model catalog in ~/.config/opencode/opencode.json
• The API key in that file with mode 0600

FireConnect snapshots the original file under ~/.fireconnect/opencode/. It never changes auth.json. Use --config-path for a configuration stored elsewhere.

fireconnect model list --search glm
fireconnect opencode --model glm-5p2

Without FireConnect, OpenCode can also connect directly: /connect → search fireworks.ai → paste your key → /models to pick.

Codex

fireconnect codex and fireconnect chatgpt share one config. The Codex CLI and the ChatGPT desktop app route through Fireworks with a single command. Quit the ChatGPT app before connecting or running off so its model list refreshes. Native web search stays on.

ChatGPT connectors. Connectors you added before connecting continue to work. You cannot add a new connector while connected.

fireconnect login
fireconnect codex

config.toml updates immediately; exit Codex and codex resume <id> (or start fresh) to load it. With a standard Fireworks key, the default model is auto.

To add FireRouter to the Codex and ChatGPT pickers, reconnect with fireconnect codex --model firerouter.

Codex CLI after connecting with FireRouter enabled; Auto is selected

ChatGPT desktop after FireConnect, with FireRouter selected

What gets written

• Root model_provider and model values in ~/.codex/config.toml
• A [model_providers.fireworks-ai] block using the Responses API
• The API key in config.toml with mode 0600
• A serverless catalog at ~/.codex/fireworks-model-catalog.json, linked through model_catalog_json

FireConnect preserves unrelated TOML settings and snapshots the original under ~/.fireconnect/codex/. Use --config-path for a configuration stored elsewhere.

fireconnect codex status
fireconnect codex --model glm-5p2

When you resume an old session, explicitly select its provider: codex resume <id> -c model_provider="fireworks-ai" (or "fireworks-azure" on the Foundry path).

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 (Claude Code, OpenCode) for MiniMax.

Pi

fireconnect login
fireconnect pi
fireconnect pi status

Restart Pi after connecting if it was already running. With a standard Fireworks key, the default model is auto.

What gets written

• defaultProvider and defaultModel in ~/.pi/agent/settings.json
• The API key in auth.json with mode 0600
• The serverless catalog in ~/.pi/agent/models.json

FireConnect snapshots all three files under ~/.fireconnect/pi/. It tracks the IDs it adds so off removes only those entries. Use --settings-path for a settings file stored elsewhere.

fireconnect model list --search glm
fireconnect pi --model glm-5p2

Cursor IDE

fireconnect cursor configures Cursor IDE. Cursor CLI (agent or cursor-agent) is not supported.

Cursor IDE keeps AI settings in SQLite (state.vscdb). Fully quit Cursor IDE before connecting or running off. In an interactive terminal FireConnect waits for you (or --force to write anyway). status is read-only while Cursor IDE runs.

fireconnect login
fireconnect cursor
fireconnect cursor status

Connecting updates every existing mode in modelConfig to use the Fireworks default. FireConnect never creates modes. It also registers the serverless catalog in the picker. status shows the model assigned to each mode. Reopen Cursor IDE and choose a Fireworks model.

Cursor IDE after FireConnect, with Auto selected

What gets written (under ~/.config/Cursor, ~/Library/Application Support/Cursor, or %APPDATA%\Cursor):

Setting	Location
API key	cursorAuth/openAIKey
Base URL	openAIBaseUrl → https://api.fireworks.ai/inference/v1
Custom models	aiSettings.userAddedModels (tracked for clean off)
Hidden built-ins	aiSettings.modelOverrideDisabled
Per-mode model	aiSettings.modelConfig[<mode>]

Your previous auth state is snapshotted under ~/.fireconnect/cursor/. Use --db-path for a non-default state.vscdb (for example Insiders).

While connected, only Fireworks models work. Built-ins are hidden. off brings them back. Cursor IDE also enforces a server-side allowlist, so not every Fireworks model is selectable even after registering. Some Cursor IDE features may still use the IDE’s own backend depending on plan and version.

Without FireConnect: add a Custom Model with a Fireworks ID, set Override OpenAI Base URL to https://api.fireworks.ai/inference/v1, paste your key.

VS Code

FireConnect adds a Fireworks provider to GitHub Copilot Chat’s custom endpoints. Requires Copilot Pro or Enterprise (free tier only supports Auto). Quit VS Code before connecting or running off; status is read-only while running.

fireconnect login
fireconnect vscode
fireconnect vscode status

Restart VS Code, then pick a Fireworks model under Other Models → Fireworks in Copilot Chat. Vision support comes from the current model catalog. glm-latest and glm-fast-latest are text-only; glm-5p3-flash supports images. Run fireconnect model list to check other model IDs.

VS Code Chat after FireConnect, with Auto selected under Fireworks

What gets written. FireConnect adds a custom endpoint to chatLanguageModels.json. VS Code appends /v1/chat/completions to https://api.fireworks.ai/inference. FireConnect also stores the API key in state.vscdb under chat.lm.secret.fw-*, encrypted through Electron safeStorage:

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\

On macOS, VS Code encrypts the key through the login Keychain. Open VS Code once before connecting. On Linux, encryption requires libsecret; otherwise, Chromium uses obfuscation and FireConnect warns you.

FireConnect snapshots the original JSON under ~/.fireconnect/vscode/. off restores that file and deletes the secret row. Use --vscode-path for a non-default configuration path. For UI setup, see GitHub Copilot.

Copilot App

The GitHub Copilot desktop app routes through Fireworks without changing its built-in models. FireConnect adds a provider alongside them. Quit the app before connecting or running off. status is read-only while the app runs. The app and CLI both store files under ~/.copilot, but they use separate settings and are configured independently.

fireconnect login
fireconnect copilot-app
fireconnect copilot-app status

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

To add FireRouter to the picker, reconnect with fireconnect copilot-app --model firerouter.

Copilot App after connecting with FireRouter enabled; Auto is selected

What gets written to ~/.copilot/data.db (SQLite, movable via COPILOT_HOME):

Table	What FireConnect adds
model_providers	One row, id starting fc-, named Fireworks, type openai
settings_json	baseUrl → https://api.fireworks.ai/inference/v1, wireApi: completions, your key as an Authorization header
provider_models	One row per model: name, display name, token limits, reasoning efforts (low/medium/high/max)

off deletes the fc- provider and its models. The desktop app’s BYOK path does not support image input, the hover-card Context row, or AI-credits pricing. Billing runs through Fireworks. Use --db-path for a non-default database.

Copilot CLI

The copilot command (@github/copilot) reads plain JSON, never the desktop app’s database.

fireconnect login
fireconnect copilot-cli
fireconnect copilot-cli status

Restart copilot after enabling.

Copilot CLI model picker after FireConnect, with Auto selected

What gets written under ~/.copilot:

File	What FireConnect writes
providers.json	One providers[] entry named fireworks + one models[] entry per model (limits, per-model vision, and a reasoning toggle)
settings.json	The selected model. The CLI requires this because BYOK providers do not define a default

Use provider-prefixed model IDs: use fireworks/glm-latest, not bare glm-latest. The CLI rejects a bare model name and silently falls back to its configured model.

Switch models with copilot --model fireworks/<name> or /model. The CLI supports image input according to each model’s catalog entry.

Because providers.json is user-editable, off restores its snapshot or deletes the file if FireConnect created it. Your existing entries remain intact. Use --providers-path for a non-default location.

DeepSeek Harness

DeepSeek’s coding agent (dsh) routes through a custom OpenAI-compatible provider under $DSH_HOME (default ~/.dsh). Restart dsh after connecting or running off.

fireconnect login
fireconnect deepseek
fireconnect deepseek status

With a standard Fireworks key, the default model is auto.

What gets written

• A fireworks provider under llm-pi-ai.providers in settings.yaml
• agent-default-model set to the selected model
• The API key as FIREWORKS_API_KEY in .credentials.yaml with mode 0600

FireConnect snapshots both files under ~/.fireconnect/deepseek/. Use --config-path for a different settings.yaml; the credentials file stays beside it.

fireconnect model list --search glm
fireconnect deepseek --model glm-5p2

Foundry across harnesses

Microsoft Foundry runs Fireworks models in your Azure subscription and bills usage through Azure. It supports OpenCode, Codex, Pi, Cursor IDE, and VS Code. It does not support Claude Code, DeepSeek Harness, Copilot, or model routers.

Pass the Foundry deployment name (FW-GLM-5.2), not a Fireworks short ID. Setup, switching, and restore behavior: Microsoft Foundry.