CCF — Claude Code Fusion

Run a multi-model fusion panel inside Claude Code, on your own subscription/flat-rate seats — no OpenRouter billing. A panel of models (GLM-5.2 1M, DeepSeek-V4-Pro 1M, optionally Sonnet/Opus, GPT-5.5 Codex) answers a task in parallel; your running Opus session judges their drafts into structured analysis (consensus / contradictions / unique insight / blind spots) and writes the final answer.

It's a local re-implementation of OpenRouter's "Fusion" pattern that bills nothing extra because the panelists run on seats you already pay for.

            ┌── glm-5.2  (z.ai, 1M ctx) ──┐
your task ──┤                              ├──►  Opus judge ──►  synthesized answer
            └── deepseek-v4-pro (1M ctx) ──┘     (this session)

Why

• Diversity beats a single model. A second/third independent draft catches blind spots.
• No metered API. Panelists run on z.ai / OpenCode subscriptions and your Claude sub.
• Huge context. Both default panelists carry ~1,048,576-token windows (verified by needle-in-haystack — see Context limits).
• Opus stays the author. Panelists never write files; the orchestrator applies the code.

Does it actually help?

We ran a controlled Solo vs Fusion benchmark — 5 real engineering tasks (bug fix, security, refactor, architecture, concurrency), the same Opus orchestrator as both the solo answerer and the fusion judge. Fusion scored 196 vs Solo 192 / 200. A real but honest lift: with a top-tier solo model the panel mostly verifies and insures against single-model blind spots (one task: glm 5/5 spots, deepseek 3/5, gpt 4/5 — the synthesis caught what the weakest missed). The run also caught a real usability bug in our own panel. Full methodology, scores, and bias disclosure: benchmark/RESULTS.md.

Requirements

• Claude Code
• bash, jq, curl, awk, tar (preinstalled on macOS/Linux except jq)
• A POSIX shell on Windows: Git for Windows (native) or WSL
• At least one provider key (z.ai and/or OpenCode). The Sonnet/Opus panelists use your claude CLI.
• For the optional Codex (GPT-5.5) panelist: python3 (stdlib only, used only for the one-time browser login).

Codex / GPT-5.5 panelist (optional)

Add OpenAI's Codex (GPT-5.5) as a panelist on your ChatGPT/Codex subscription — no codex CLI required. Log in through your browser:

~/.claude/fusion/ccf-codex-login          # opens ChatGPT OAuth, stores tokens (chmod 600)
~/.claude/fusion/ccf-codex-login --no-browser   # headless/SSH: prints the URL

This runs the same OAuth 2.0 + PKCE flow the codex CLI uses, writes ~/.codex/auth.json, and fusion-call auto-refreshes the token when it nears expiry. Then enable it:

# add the gpt panelist (provider codex / model gpt-5.5) if not already present, then:
/fusion-config enable gpt
/fusion-status            # gpt should probe OK

Install

macOS / Linux / WSL

curl -fsSL https://raw.githubusercontent.com/brahmsyaifullah/CCF/main/install.sh | bash

Windows — one line in PowerShell. It auto-installs everything missing (Git Bash, jq, Python), refreshes PATH, installs CCF, and runs setup. No manual prerequisites.

irm https://raw.githubusercontent.com/brahmsyaifullah/CCF/main/install.ps1 | iex

Prefer not to touch a terminal? Double-click install-windows.bat. Details + troubleshooting: docs/WINDOWS.md.

Or clone and run locally:

git clone https://github.com/brahmsyaifullah/CCF && cd CCF && ./install.sh

The installer copies the dispatcher, hooks, and slash commands into ~/.claude, creates a secrets.env (chmod 600) for your keys, and idempotently wires two hooks into settings.json (backed up first). It never overwrites an existing secrets.env, panel.json, or providers.json.

Then:

1. Add keys — edit ~/.claude/fusion/secrets.env or run /fusion-config set-key.
2. Restart Claude Code (hooks + commands load at launch).
3. /fusion-status to verify, then /fusion <task>.

Usage

Command	What it does
/fusion-onboard	Guided setup — pick providers, add keys (validated), enable panelists.
/fusion <task>	Run a panel → judge → synthesis pass on demand (works anytime).
/fusion-on · /fusion-off	Toggle default-mode (proactive routing of substantial tasks).
/fusion-status	Show roster, providers, context limits, key presence, reachability.
/fusion-config	Add/edit providers, set/rotate keys, enable/disable panelists.
/fusion-analytics	Text dashboard: run stats, latency, success rate, cost saved vs OpenRouter.
/fusion-benchmark	Run SOLO vs FUSION on 5 coding tasks. Sequential, outputs markdown.
/fusion-benchmark-report	Generate comparison REPORT.md from benchmark results.
/ccf-update	Update CCF from GitHub (preserves your keys + config).

Default-mode ships off — /fusion is always available explicitly.

Onboarding

Two ways, both smooth on macOS, Linux, and native Windows:

1. Let Claude Code do it — just say /fusion-setup (or "set up CCF" with the repo link). The agent installs if needed, checks dependencies, asks which providers + models you want (buttons, no secrets in chat), and walks you through key entry privately. Zero file editing.

2. Run the wizard yourself — cross-platform Python, no bash or jq required for setup:

python3 ~/.claude/fusion/ccf-onboard          # interactive: pick providers → key (hidden) → model
python3 ~/.claude/fusion/ccf-onboard --check  # dependency doctor (OS-specific install hints)

It presents the default panel (glm, deepseek, gpt/Codex) and the 20-provider catalog, takes a key hidden (or browser-login for Codex, or nothing for Opus/Sonnet/local Ollama), validates it, and enables the panelist — idempotent, every write backed up, keys only ever in secrets.env.

The runtime dispatcher itself uses bash + jq (curl). On Windows that means Git Bash or WSL; install.ps1 auto-installs jq via winget. Onboarding (the Python wizard) needs neither.

Providers & panelists

Default roster (panel.json):

• Judge / writer: Opus (your Claude session) — never demoted.
• Panel (enabled): glm (z.ai GLM-5.2) · deepseek (OpenCode-Go DeepSeek-V4-Pro).
• Available (disabled): sonnet, opus (your sub via claude CLI), deepseek-flash, north-code (free, non-zero-retention), gpt (Codex GPT-5.5 via your OpenAI subscription).

Provider catalog

CCF ships a catalog (~/.claude/fusion/catalog.json) of 21 providers with correct OpenAI/Anthropic-compatible endpoints, recommended current models, and docs links — disabled by default, so you enable only what you have keys for:

OpenAI · Anthropic · Google Gemini · DeepSeek · xAI Grok · Mistral · Moonshot · Kimi-Code · MiniMax · Groq · Cerebras · Qwen · OpenRouter · Together · Fireworks · Novita · HuggingFace · Xiaomi MiMo · Ollama (cloud) · Ollama (local — keyless, air-gapped, zero-retention) · Codex (OpenAI subscription — GPT-5.5 via your own Codex CLI auth).

Enable one via /fusion-onboard, or /fusion-config add-provider --from-catalog <name>. The catalog is read-only reference (refreshed on update); your live providers.json is never touched by it. Keyless/self-hosted endpoints (Ollama local) need no key — ideal for sensitive code.

The complete registry lives at models.dev — 121+ providers, every model with context + cost. The curated catalog is a starter subset; for the full list use ccf-models:

~/.claude/fusion/ccf-models providers            # list all 121+ providers
~/.claude/fusion/ccf-models models <provider>    # a provider's models (context, $/Mtok)
~/.claude/fusion/ccf-models add <provider> <model> [panelist]   # register + add panelist

Panelist system prompt

Every panelist inherits panel.json's default_system_prompt (role + output contract: answer directly, propose-don't-act, flag risks with severity). Override one with a per-panelist system_prompt, or set the default to "" to disable. fusion-call injects it correctly per transport (openai system message / anthropic top-level system).

Context limits

Verified 2026-06-16 by needle-in-haystack (needle retrieved at 90% depth in a full ~1M-token prompt):

Panelist	Provider	Ceiling	tok/char	Usable prompt
glm	z.ai GLM-5.2	1,048,576 tok	~0.264	≤ ~3.9M chars
deepseek	OpenCode-Go DeepSeek-V4-Pro	1,048,565 tok	~0.295	≤ ~3.55M chars
sonnet/opus	claude sub	~200K tok	~0.25	≤ ~760K chars
gpt	Codex GPT-5.5	~200K tok	~0.25	≤ ~760K chars

The dispatcher rejects oversize prompts before upload (with the exact char budget) instead of letting the API return a silent empty response. Large 1M calls take ~40–60s.

Web search (optional)

Set TAVILY_API_KEY in secrets.env to let panelists search the web during a task. Add --search when calling a panelist:

~/.claude/fusion/fusion-call --search deepseek "What changed in the latest React release?"

Panelists use standard function calling to request searches; CCF calls Tavily locally and feeds results back (up to 3 rounds). Currently supported on the openai transport (DeepSeek). Get a key at tavily.com.

Security

• No keys in this repo. secrets.env is gitignored; only secrets.env.example is tracked.
• Keys live in ~/.claude/fusion/secrets.env (chmod 600), referenced by key_env name only.
• Zero-retention gating. Free/pay-per-token providers are flagged zero_retention=false; the /fusion sensitivity gate refuses to send proprietary/customer code to any panelist marked sensitive_ok=false. Check /fusion-status — non-zero-retention providers show FALSE!.
• Rotate a key anytime with /fusion-config set-key.

Updating

Notify-only by default: a throttled (~daily), fail-silent SessionStart hook checks GitHub and, if a newer version exists, prints [CCF] update available …. Apply with:

/ccf-update            # or: bash ~/.claude/fusion/ccf-update.sh
~/.claude/fusion/ccf-update.sh --check    # report only
~/.claude/fusion/ccf-update.sh --force    # reinstall latest

Updates refresh code only (dispatcher, hooks, commands, .dist templates). Your secrets.env, panel.json, and providers.json are preserved; new reference fields are reported for manual merge. Skip the notifier at install with --no-update-hook.

Uninstall

rm -rf ~/.claude/fusion ~/.claude/commands/fusion*.md ~/.claude/commands/ccf-update.md
# then remove the two CCF hook entries from ~/.claude/settings.json
# (a timestamped settings.json.ccf-bak.* backup was made at install)

How it works

File (installed under ~/.claude/)	Role
fusion/fusion-call	Dispatcher. openai / anthropic / claude-cli transports. File-based body (1M-token safe past ARG_MAX), per-provider timeout, oversize-context guard, clear ERR: reporting.
fusion/providers.json	Transport registry: endpoint, auth, key_env, max_tokens, request_timeout, max_context_tokens, tok_per_char, zero_retention.
fusion/panel.json	Roster + active flag. enabled panelists form the default panel; sensitive_ok gates proprietary code.
fusion/fusion-hook.sh	UserPromptSubmit hook — default-mode reminder when active.
fusion/ccf-check-update.sh	SessionStart hook — throttled update notifier.
fusion/secrets.env	Your keys, chmod 600, never committed.
commands/*.md	The slash commands.

License

MIT — see LICENSE.