Claude Code

Use Fireworks AI models in Claude Code with the FireConnect CLI

FireConnect routes Claude Code through Fireworks AI models. See the FireConnect overview for install and CLI basics.

Change models: fireconnect claude on --model <id> (or --opus / --sonnet / …). See Models.

Prerequisites

• Claude Code installed
• A Fireworks API key (fw_...) or a Fire Pass key (fpk_...)
• FireConnect CLI v0.9.5+ (see Install)

Azure routing not implemented yet for Claude Code. fireconnect claude on always configures direct Fireworks, even when global config has --provider azure or you pass --azure. See Microsoft Foundry in FireConnect.

Enable Fireworks routing

fireconnect login
fireconnect claude on

Or pass the key once:

fireconnect claude on --api-key fw_...

Restart Claude Code after enabling, then test with a simple prompt.

After fireconnect claude on, start a new Claude Code session or run /model to pick up model changes. To use a new model in the same session, start a new session or /exit and resume with claude --resume <id>.

Change models

fireconnect model list --search glm
fireconnect claude on --model firerouter               # main only
fireconnect claude on --interactive                    # model mapping wizard
fireconnect claude on --opus glm-fast-latest --sonnet auto-instant
fireconnect claude status

Model flags directly override the named slot. Set every slot you want to control explicitly. Use native as a slot value to leave a slot unpinned so Claude Code chooses its default for that role.

Re-running on without flags preserves your current mapping when FireConnect is already active (including /model changes made inside Claude Code). Reopen the wizard anytime:

fireconnect claude on --interactive

Use --non-interactive to skip the wizard and apply saved preferences or the current profile. --interactive cannot be combined with model flags like --model or --opus.

Use the wizard to choose aliases, or use explicit model flags for a reproducible configuration.

Smart router mixes

Pin Fireworks’ open-model mixes on any slot:

fireconnect claude on --sonnet auto              # default open-model mix
fireconnect claude on --sonnet auto-instant      # latency-first mix

auto appears on every slot in the picker; auto-instant is available on Sonnet. These route among Fireworks open models only — unlike FireRouter, they do not pass through to Claude Opus 5.

What gets written

FireConnect writes the selected mapping to ~/.claude/settings.json. Claude Code authenticates via the X-Fireworks-Api-Key custom header (not apiKeyHelper). The Fireworks key is written to the file with mode 0600.

Why the custom header? The gateway authenticates via X-Fireworks-Api-Key, which wins over any stray ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN — so a leftover Anthropic key can’t silently break routing.

Model IDs and [1m]. Short slugs are accepted everywhere. FireConnect adds a [1m] suffix on 1M-context models for every slot — subagent included — so Claude Code sizes the context window correctly. The gateway still sees the bare model ID; Claude Code strips the tag before sending.

on also:

• Denies Anthropic server-side WebSearch / WebFetch tools the gateway can’t run, and installs the Fireworks WebSearch MCP when your account is eligible
• Installs a status line showing routed models and estimated session cost (see Status line). Your own statusLine is never replaced
• Sends privacy-safe attribution headers (X-Title, HTTP-Referer) where supported

FireConnect saves a backup of your previous provider settings to ~/.fireconnect/claude/. When that backup is available, fireconnect claude off restores it byte-for-byte; otherwise it removes FireConnect-managed settings. Legacy pins like deepseek-v4-flash migrate to deepseek-flash-latest on the next on.

Status line

When you don’t already have a custom statusLine, on installs one that shows which models actually served the session and their estimated cost:

━━━━━━━━━━━━ ━ ━ · $70.39
━ Claude Opus 5 $62.91 98% cache · ━ GLM 5.2 $7.35 96% cache · ━ DeepSeek V4 Flash $0.13 87% cache

The first line is a multi-color bar sized by each backend model’s share of session spend, followed by the session total. The second line names each model with its cost and cache hit rate. FireRouter sessions show the whole mix at a glance — not just the latest model.

The cost uses Fireworks serverless rates for Fireworks models and Anthropic list rates for Anthropic models. Calls without a matched rate remain in usage totals but are not assigned a dollar cost.

off removes the status line. If you already have a statusLine, FireConnect leaves it alone — delete yours and re-run fireconnect claude to opt in.

Web search MCP

When you run fireconnect claude on, FireConnect can install the Fireworks WebSearch MCP for eligible accounts. FireConnect WebSearch integration is Claude Code only today; other harnesses can add the HTTP MCP manually.

When installation succeeds, fireconnect claude on prints Web search → fireworks-websearch (installed). Restart Claude Code, then run /mcp and connect to fireworks-websearch.

fireconnect claude off removes the managed MCP entry and restores your previous ~/.claude/settings.json.

Browsing and picking models

fireconnect model list --search glm
fireconnect model list --refresh
fireconnect claude on --sonnet kimi-latest

fireconnect model list shows serverless endpoints and pricing (cached for one hour; works offline). fireconnect claude status shows your current mapping, every slot including defaults, and available rates per slot.

Fire Pass keys only list Fire Pass routers. FireConnect rejects --model firerouter with Fire Pass (fpk_...) on every harness; use an fw_... key.

FireRouter

Route requests through FireRouter. The default firerouter model ID is more likely to use GLM 5.3 for lower-complexity work and Claude Opus 5 for harder work. See current routing pair.

Choose where FireRouter is used

Use --model for Main or a slot-specific flag for an alias:

fireconnect claude on --model firerouter
fireconnect claude on --opus firerouter

On first setup, FireConnect assigns models to slots you do not specify. Use --interactive to choose alias slots. To control the complete mapping including Main, set every slot explicitly. Use native to leave a slot unpinned.

fireconnect claude on --interactive

# FireRouter on Main; leave every alias unpinned
fireconnect claude on --model firerouter \
  --opus native --sonnet native --haiku native \
  --fable native --subagent native

Custom routing pairs

fireconnect claude on --sonnet firerouter/claude-opus-5/glm-5p3-fast   # custom pair

Use a slash-delimited FireRouter slug on any slot (firerouter/<primary>/<redirect>, …) to change which models FireRouter can pick.

Anthropic credentials are needed only when the slug includes a Claude model (including bare firerouter). A Fireworks-only slug such as firerouter/kimi-k3/glm-5p2-fast needs only the Fireworks API key.

Anthropic auth for pass-through

The default bare firerouter route’s Claude Opus 5 leg needs Anthropic credentials. With FireConnect you usually do not pass a new key:

• Claude subscription or browser OAuth — Claude Code already attaches Anthropic auth on each request; Fireworks forwards it for pass-through.
• Anthropic API key — if Claude Code is configured with ANTHROPIC_API_KEY, that works too.
• Explicit BYOK — pass --anthropic-api-key sk-ant-... on on, or store once with fireconnect configure --anthropic-api-key sk-ant-....
• Workspace BYOK — when enabled on your Fireworks account, no local Anthropic key is required.

See which models served the session

After assistant messages run, check:

1. Status line (installed by on unless you already have a custom statusLine) — lists backend models that served the session (for example Claude Opus 5 or GLM 5.3) and estimated session cost. Before a billed response, it may show FireRouter rather than a resolved backend. See Status line.
2. fireconnect claude status — shows the configured slot mapping and which slots use FireRouter. It does not show per-request routing decisions.

Tune cost vs quality

Set routing preference with --routing-preference on on (1–5 or a named level):

fireconnect claude on --model firerouter --routing-preference 4
# max-intelligence (1) · more-intelligence (2) · balanced (3) · more-savings (4) · max-savings (5)

See Routing preferences.

Omit the flag to use FireRouter’s default. Re-running on without --routing-preference clears a preference that FireConnect previously wrote.

Usage and live meter

Claude Code’s /model picker shows Anthropic list prices. Use fireconnect claude usage to estimate session cost from Fireworks serverless rates and Anthropic list rates.

On a TTY, usage opens a session picker (last 3 days by default), then a live cost meter. Tab: agents pane. Esc: session list. q: quit.

fireconnect claude usage                 # picker → live meter
fireconnect claude usage --days 7        # picker lookback only (1–365)
fireconnect claude usage --session <id>  # start on one session
fireconnect claude usage --last-n 5      # snapshot, no picker (--days ignored)
fireconnect claude usage --plain         # plain text
fireconnect claude usage --json          # JSON
fireconnect claude usage --verbose       # request-level rows and rate details

For a tmux split (Claude Code left, meter right):

fireconnect claude live
fireconnect claude live --session <id>

Requires tmux. Neither command changes harness settings.

Scripting:

fireconnect claude status --json
fireconnect claude usage --last-n 5 --json

Troubleshooting

Text-only models and images

Claude Code cannot mark a model as non-vision. Pasting an image on a text-only slot (for example glm-fast-latest or deepseek-flash-latest) can break the session.

Recover with /rewind, then avoid images on that slot or map it to a vision model:

fireconnect claude on --sonnet kimi-fast-latest

fireconnect claude on warns when your mapping includes text-only models. fireconnect claude status labels each slot vision or text-only.

Pricing estimates

Claude Code’s session cost uses Anthropic list prices, while Fireworks bills at serverless rates. Use fireconnect claude status, fireconnect model list, and the billing dashboard for actual spend.

CLI reference

fireconnect claude on         # Route Claude Code through Fireworks
fireconnect claude off        # Restore your previous provider
fireconnect claude status     # Provider, auth, and model mapping
fireconnect claude usage      # Session cost meter
fireconnect claude live       # tmux split with live meter
fireconnect claude demo       # Race two models on a prompt
fireconnect claude help       # Harness-specific help

Run fireconnect claude help for all options.

Turn off Fireworks routing

fireconnect claude off

This restores your previous ~/.claude/settings.json from the backup saved in ~/.fireconnect/claude/.

Uninstall

To remove FireConnect from your machine entirely (all harnesses):

fireconnect uninstall

Source

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