Overview - Fireworks AI Docs

FireConnect is an open-source CLI that routes agentic coding harnesses through Fireworks models. Choose where inference runs:

Direct Fireworks routing (default) — requests go to the Fireworks gateway; sign in with fireconnect login or use a Fireworks API key (fw_...) or Fire Pass key (fpk_...)
Fireworks on Microsoft Foundry — the same models, deployed in your Azure subscription and billed through Azure; use --provider azure and an Azure API key (see Fireworks on Microsoft Foundry)

Install once, sign in, then enable or disable routing per harness without editing config files by hand.

Claude Code

Anthropic-compatible routing with multi-slot model aliases

OpenCode

OpenAI-compatible adapter in opencode.json

Codex

OpenAI Codex CLI via the Responses API

Pi

Pi agent settings and auth

Cursor

OpenAI BYOK settings for Cursor IDE

VS Code

GitHub Copilot Chat custom endpoint

Deep Agents

LangChain Deep Agents Code (dcode) via OpenAI-compatible config

Prerequisites

A Fireworks API key (fw_...) or a Fire Pass key (fpk_...) for direct Fireworks routing
For Fireworks on Microsoft Foundry: a Microsoft Foundry resource, an Azure API key, and at least one model deployment (see Fireworks on Microsoft Foundry and the Microsoft Foundry integration guide)
Node.js 18+ (the installer can install it via Homebrew on macOS or prints upgrade instructions elsewhere)
At least one supported harness installed locally

Install

curl -fsSL https://raw.githubusercontent.com/fw-ai/fireconnect/main/install.sh | bash

The installer:

Clones the CLI to ~/.fireconnect/cli, runs npm install --omit=dev for its runtime dependency (cross-keychain, used for secure API-key storage), and adds a fireconnect launcher to ~/.local/bin
Does not sign you in or write harness settings — run fireconnect login then fireconnect <harness> on after it finishes

After install, sign in and enable a harness:

fireconnect login        # guided sign-in (browser or paste a key)
fireconnect claude on    # Claude Code (default starting point)
fireconnect opencode on  # OpenCode
fireconnect codex on       # Codex
fireconnect pi on          # Pi
fireconnect cursor on      # Cursor IDE (quit Cursor first)
fireconnect vscode on      # VS Code + GitHub Copilot Chat
fireconnect deepagents on  # LangChain Deep Agents Code

You do not have to run login first — any key-needing command like fireconnect claude on runs the same sign-in flow inline, then finishes the job.

Upgrade

If you installed via the curl installer, pull the latest release in place:

fireconnect upgrade

Re-run the install script if FireConnect was not installed with git (for example, a manual checkout). When a newer version is available, FireConnect prints a one-line update banner on startup with upgrade instructions. The check runs in the background and does not block your command. Check your installed version:

fireconnect --version

FireConnect stores API keys in the OS keychain when available (via cross-keychain), not as plaintext in harness config files.

fireconnect login     # browser sign-in (creates a machine key) or paste an existing key
fireconnect logout    # clear the stored key (keychain entry + config ref)
fireconnect status    # show sign-in state and where the key is stored

~/.fireconnect/config.json stores a reference such as {keychain:fireworks-api-key} or {env:FIREWORKS_API_KEY} — never a literal key. Claude Code uses apiKeyHelper to fetch the key at runtime; Codex, OpenCode, Pi, and Deep Agents use environment references plus a shell profile hook that exports FIREWORKS_API_KEY="$(fireconnect key export)". Fire Pass users can paste a fpk_... key during login or on. FireConnect detects the key type and applies the correct defaults.

Fireworks on Microsoft Foundry

Route OpenCode, Codex, and Pi through a Microsoft Foundry deployment instead of the Fireworks gateway. Usage is billed through Azure and counts toward MACC where applicable.

FireConnect + Microsoft Foundry

Configure --provider azure, deployment names, per-harness config, and switching back to direct Fireworks

Microsoft Foundry portal setup

Enable Fireworks on Foundry, PayGo / PTU modes, billing, and troubleshooting

Quick start:

export AZURE_API_KEY=<your-azure-api-key>

fireconnect configure \
  --provider azure \
  --base-url https://<resource>.services.ai.azure.com \
  --api-key $AZURE_API_KEY

fireconnect opencode on --main FW-GLM-5.1

Claude Code, Cursor, VS Code, and Deep Agents do not have Azure routing in FireConnect yet — they always use direct Fireworks when you run on, even if global config has --provider azure. See the full Microsoft Foundry guide for which harnesses are implemented, endpoint normalization, one-off --azure routing, turning Foundry off, and what each harness writes.

CLI design

FireConnect uses harness-first syntax: fireconnect <harness> <command> Bare harness names run on (for example, fireconnect claude is the same as fireconnect claude on).

Global commands

fireconnect login        # Sign in — browser (creates a key) or paste a key you have
fireconnect logout       # Clear the stored key (keychain entry + config ref)
fireconnect status       # Show sign-in state and where the key is stored
fireconnect configure    # Set the provider (Azure/Foundry) and Anthropic key for router mode
fireconnect demo         # Race your provider vs Fireworks GLM 5.2 Fast on the same prompt
fireconnect upgrade      # Pull the latest FireConnect from GitHub (curl install)
fireconnect uninstall    # Disable all harnesses, restore configs, remove CLI
fireconnect help         # Show help
fireconnect --version    # Print the installed CLI version (-V also works)

Providers

FireConnect supports two inference backends. The CLI provider flag is shown in parentheses.

Provider flag	Where inference runs	API key	Supported harnesses
`fireworks` (default)	Fireworks gateway	`fw_...` or `fpk_...`	Claude Code, OpenCode, Codex, Pi, Cursor, VS Code, Deep Agents
`azure`	Fireworks on Microsoft Foundry	Azure API key	OpenCode, Codex, Pi

Set the default with fireconnect configure --provider fireworks or --provider azure. For Foundry, also pass --base-url with your resource endpoint. Harness on commands use the configured provider unless you pass --azure or per-command --base-url / --api-key overrides.

Per-harness commands

Each CLI harness (claude, opencode, codex, pi, deepagents) supports:

fireconnect <harness> on — Route through the configured provider
fireconnect <harness> off — Restore your previous config
fireconnect <harness> status — Show provider, auth, and models
fireconnect <harness> model list — Browse Fireworks serverless models (direct Fireworks routing only)
fireconnect <harness> model select — Pick a serverless model interactively (direct Fireworks routing only)
fireconnect <harness> model reset — Reset models to defaults
fireconnect <harness> help — Harness-specific help

Claude Code also has fireconnect claude usage (estimate usage cost from a session log). Each IDE harness (cursor, vscode) supports:

fireconnect <harness> on — Configure direct Fireworks routing (Azure not implemented for IDE harnesses yet; quit the IDE first)
fireconnect <harness> off — Restore your previous config (quit the IDE first)
fireconnect <harness> status — Show provider, auth, and registered models (read-only; IDE can stay open)
fireconnect <harness> model list — Browse Fireworks serverless models (read-only)
fireconnect <harness> model add <id> — Add a Fireworks model to the IDE picker
fireconnect <harness> model select — Pick a model interactively (Cursor: pass --mode)
fireconnect <harness> model reset — Reset fireconnect-managed models to defaults
fireconnect <harness> help — Harness-specific help

VS Code also supports fireconnect vscode on --router for FireRouter (Anthropic models via https://router.fireworks.ai/v1/messages). Pi supports fireconnect pi on --router for the same routing pattern. Cursor stores settings in SQLite (state.vscdb). Commands that write to that database (on, off, model add, model select, model reset) require Cursor to be fully quit first. VS Code stores the API key in state.vscdb as well, so on and off require quitting VS Code; model add, model select, and model reset only edit chatLanguageModels.json, which VS Code hot-reloads. With Fireworks on Microsoft Foundry (--provider azure), pass your Foundry deployment name to on with --main instead of using model list or model select. Run fireconnect help for the overview, or fireconnect claude help (and similarly for other harnesses) for harness-level options.

API key resolution

When a command needs credentials, FireConnect resolves them based on the active provider. Direct Fireworks routing (--provider fireworks)

Explicit --api-key
OS keychain (via fireconnect login)
Global ~/.fireconnect/config.json reference
FIREWORKS_API_KEY environment variable

Claude Code additionally reads harness-local keys from ~/.claude/settings.json when FireConnect is already enabled there. Fireworks on Microsoft Foundry (--provider azure)

Explicit --api-key
Global ~/.fireconnect/config.json
AZURE_API_KEY environment variable

Recommended models

These serverless model short IDs apply to direct Fireworks routing and expand to full Fireworks paths automatically. With Fireworks on Microsoft Foundry, pass your deployment name instead (for example, FW-GLM-5.1 or FW-MiniMax-M2.5) via --main.

Short ID	Best for	Notes
`glm-latest`	All-around use, agentic tasks	Version-tracking router; strong reasoning, 1M context.
`glm-fast-latest`	Latency-sensitive agentic use	Default for `main`, `opus`, and `fable` slots in Claude Code. Version-tracking router on the high-speed Fast serving path (100+ tok/s), at a higher per-token price. 1M context.
`glm-5p2-fast`	Latency-sensitive agentic use	Same as `glm-fast-latest` but pinned to GLM 5.2 rather than version-tracking. 1M context.
`glm-5p1`	General use (lighter)	Default `sonnet` slot in Claude Code. Good balance of speed and quality.
`deepseek-v4-flash`	Background / fast tasks	Default `haiku` and `subagent` slots in Claude Code. Lowest latency.

Fire Pass keys

Fire Pass keys (fpk_...) default all slots to glm-fast-latest. The model browser shows Fire Pass-supported routers: glm-latest, glm-fast-latest, glm-5p2-fast, kimi-fast-latest, and kimi-k2p7-code-fast. Fire Pass keys work with direct Fireworks routing only. Use an Azure API key with --provider azure.

Codex does not support Fire Pass keys yet. Use a standard Fireworks API key (fw_...) with Codex.

Migration from earlier syntax

FireConnect uses harness-first commands. If you have older docs or scripts from pre-0.5.0 releases, update them:

Before	After
`fireconnect on`	`fireconnect claude on`
`fireconnect off`	`fireconnect claude off`
`fireconnect status`	`fireconnect claude status`
`fireconnect list`	`fireconnect claude status`
`fireconnect set --main <id>`	`fireconnect claude on --main <id>`
`fireconnect reset`	`fireconnect claude model reset`
`fireconnect on --harness opencode`	`fireconnect opencode on`
`fireconnect model list`	`fireconnect <harness> model list`

Source

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

Claude Code

OpenCode

Codex

Pi

Cursor

VS Code

Deep Agents

​Prerequisites

​Install

​Upgrade

​Sign in and API keys

​Fireworks on Microsoft Foundry

FireConnect + Microsoft Foundry

Microsoft Foundry portal setup

​CLI design

​Global commands

​Providers

​Per-harness commands

​API key resolution

​Recommended models

​Fire Pass keys

​Migration from earlier syntax

​Source

Prerequisites

Install

Upgrade

Sign in and API keys

Fireworks on Microsoft Foundry

CLI design

Global commands

Providers

Per-harness commands

API key resolution

Recommended models

Fire Pass keys

Migration from earlier syntax

Source