Skip to main content
FireConnect is an open-source CLI that routes agentic coding harnesses through Fireworks models. Install once, sign in once, then flip any supported harness on or off — no proxy to run, no wrapper to launch. on rewrites the tool’s config; off restores your original file byte-for-byte. Choose where inference runs:
  • Direct Fireworks routing (default): the Fireworks gateway. Sign in with fireconnect login or use a Fireworks API key (fw_...) or Fire Pass key (fpk_...).
  • Fireworks on Microsoft Foundry: models in your Azure subscription, billed through Azure. See Microsoft Foundry.
New here? Follow Quick start below, then open the harness guide for your tool. Try the side-by-side demo to compare models before switching.

​
Quick start

1. Install
2. Sign in
3. Connect a harness
Example output:
4. Restart the tool, then verify
Swap claude for any harness: opencode, codex, chatgpt, pi, cursor, vscode, deepseek. Bare harness names run on — fireconnect claude is the same as fireconnect claude on.

​
Prerequisites

  • A Fireworks API key (fw_...) or Fire Pass key (fpk_...) for direct routing
  • For Foundry: Azure resource, API key, and deployment. See portal setup.
  • Node.js 18+
  • At least one supported harness installed locally

​
Install notes

  • Requires bash and Node.js 18+. The installer clones the CLI to ~/.fireconnect/cli, adds ~/.local/bin/fireconnect to your PATH, and runs the same finalize as fireconnect upgrade. It does not sign you in or touch harness settings.
  • You can pass --api-key or set FIREWORKS_API_KEY instead of login.
Windows: run from Git Bash with the same curl command. Piping through PowerShell corrupts line endings (set: pipefail\r: invalid option name).

​
Upgrade FireConnect

Or re-run the install curl above. From v0.9.0 onward, harness settings stay connected across upgrade. Check version with fireconnect --version. See CLI reference: Migration for renames.

​
Sign in

Fire Pass keys (fpk_...) work during login or on. FireConnect detects the key type and applies the correct defaults. See CLI reference: Sign in options for --with-token, --account, logout --revoke, and configure.

​
Choose your harness

After install and sign-in, open the guide for the tool you use. Each page covers on / off, models, and harness-specific notes.

Claude Code

Six model slots, usage meter, and session status line

OpenCode

OpenAI-compatible adapter in opencode.json

Codex

Codex CLI and ChatGPT desktop app via the Responses API

Pi

Pi agent settings and auth

Cursor

OpenAI BYOK settings for Cursor IDE

VS Code

GitHub Copilot Chat custom endpoint

DeepSeek Harness

DeepSeek’s coding agent (dsh)

​
Models

<harness> is one of: claude, opencode, codex, chatgpt, pi, cursor, vscode, deepseek. Full walkthrough: Models.

​
Default models

Fire Pass keys are detected automatically — no flags needed. Override anytime with flags or the model mapping wizard.

​
Harness support

Notes
  • Foundry: not on Claude Code or DeepSeek Harness.
  • Fire Pass: not on Codex or Foundry. FireConnect rejects --model firerouter with a Fire Pass (fpk_...) key on every harness; use an fw_... account key for FireRouter.
  • FireRouter: Cursor and DeepSeek Harness need workspace BYOK for Anthropic pass-through.
  • Web search MCP: Claude Code auto-install only; any harness can add the HTTP MCP manually. See WebSearch MCP.
  • Cursor / VS Code: quit the IDE before on or off. status is read-only.

​
FireRouter and smart routers

FireRouter routes simple requests to cheaper open models and hard requests to Claude Opus 5. Select it like any other model:
Smart router mixes (preview): pin auto for Fireworks’ default open-model mix, or auto-instant for latency-first routing:
Requires a standard Fireworks key (fw_...), not Fire Pass. See Models and the FireRouter overview.

​
Source

FireConnect is open source: github.com/fw-ai/fireconnect

​
See also