Use an AI CLI You Already Have

Self-hosted only. SynapseOrch Cloud runs no CLIs on your behalf.

If you already pay for Claude, Gemini, ChatGPT or GitHub Copilot, an agent can run on that subscription through its official command-line tool: Claude Code, Gemini CLI, Codex or Copilot CLI. No API key is needed; the CLI is signed in as you.

The CLI's model, and nothing else

Each CLI is an agent in its own right, with its own tools and setup. SynapseOrch uses only its model. On every call it switches off the CLI's own tools, MCP servers, hooks, skills, plugins, extensions and memory files (CLAUDE.md, GEMINI.md, AGENTS.md), your project around it, and its saved chat. Each call runs in an empty folder of its own, deleted afterwards. Your CLI setup is never changed.

The agent's own tools still work. SynapseOrch writes them into the prompt and runs them itself; Codex is given them as real tools through a small bridge, and SynapseOrch runs every call. An answer the CLI reached through a tool of its own is refused.

What a CLI does not let SynapseOrch switch off, said plainly:

The cards in Settings → Models

Each CLI has a card in Settings → Models. It shows whether the CLI is installed and its version, whether it is signed in (and the command to sign in if not), the models you can pick, and where that list came from: From Codex, From Copilot CLI, or Synapse's list for Claude Code 2.1 for a CLI that can't list its own. The answer is kept for ten minutes; Refresh asks the CLIs again.

Custom model names (comma-separated) is still on every card: each name you add becomes a selectable model, listed first, and is passed to the CLI's model flag.

SynapseOrch supplies the agent's own system prompt: to Claude with --system-prompt-file, to Gemini through GEMINI_SYSTEM_MD, and to Codex and Copilot inside the prompt itself.

Claude CLI

Models: cli.claude, or cli.claude.<model>

Uses the claude CLI (Claude Code), through your Claude subscription. The aliases fable, opus, sonnet and haiku always mean the newest model of that kind your plan allows, so they never go stale.

Setup:

1. Install Claude Code: claude.ai/download
2. Run claude in a terminal on the SynapseOrch host and sign in
3. In SynapseOrch, pick a cli.claude model

Models and effort. cli.claude uses the CLI's default model; cli.claude.<model> passes the model with --model. SynapseOrch asks the CLI for JSON output, and sends --effort from the agent's reasoning effort, or the workspace's reasoning_effort setting when the agent has none. off, or no effort set anywhere, sends low. The agent editor shows no Reasoning control for CLI models, so an agent's own value is whatever it was set to on another model.

A model name ending in -thinking, -max or -high, such as cli.claude.claude-opus-4-6-thinking, still works: the suffix is ignored and does not change the effort. The -thinking names are no longer listed.

---

Gemini CLI

Models: cli.gemini, cli.gemini.pro, cli.gemini.flash

Uses the gemini CLI, version 0.64 or newer. .pro and .flash choose the CLI's Pro or Flash model.

Sign in with a key. Google refuses its free Code Assist for individuals sign-in for this use, and says to migrate to Antigravity: set a GEMINI_API_KEY from Google AI Studio in the backend's environment instead, or choose Sign in with Google on an account Google allows.

Setup:

1. Install: npm install -g @google/gemini-cli — see Gemini CLI
2. Run gemini in a terminal on the SynapseOrch host and sign in, or set GEMINI_API_KEY
3. In SynapseOrch, pick a cli.gemini model

---

OpenAI Codex CLI

Models: cli.codex, or cli.codex.<model>

Uses the OpenAI Codex CLI, version 0.163 or newer, through your OpenAI account; SynapseOrch asks you to update an older one. The card lists the models Codex knows (From Codex); that is not the same as what your plan includes, so a ChatGPT sign-in may refuse some of them. cli.codex.<model> passes the model with -m.

Setup:

1. Install: npm install -g @openai/codex
2. Run codex login in a terminal on the SynapseOrch host
3. In SynapseOrch, pick a cli.codex model

Codex runs non-interactively in a read-only sandbox, with none of its own tools; the agent's tools reach it through SynapseOrch's bridge.

---

GitHub Copilot CLI

Models: cli.copilot, or cli.copilot.<model>

Uses the standalone GitHub Copilot CLI — the copilot command — through your Copilot subscription. It is not the older gh copilot extension, which SynapseOrch does not use.

Models:

The card lists the models Copilot CLI offers you (From Copilot CLI), with auto first. A Free or Student plan may allow only auto.

Setup:

1. Install the Copilot CLI — see GitHub's installation guide
2. Run copilot in a terminal on the SynapseOrch host and /login, or sign in with gh auth login
3. In SynapseOrch, pick a cli.copilot model

---

CLI mode limitations

• Slower than direct API calls — each call starts a CLI process
• A fresh CLI session every call — no CLI session is resumed; the conversation so far is sent in the prompt each time
• 5-minute limit — a call that takes longer is stopped
• Tools written into the prompt — except Codex's bridge, the model answers with a tool-call block, which is less reliable than a provider's own tool calling
• No images — attachments are dropped without a warning
• Token counts are estimated from the text, so Usage shows approximate figures
• No streaming — responses are returned after completion
• Rate limits apply based on your subscription plan
• Not suitable for high-throughput production workloads — use direct API keys instead

CLI mode is best for personal use, development, and experimentation where you don't want to manage API keys separately.

Troubleshooting

• … is not installed on this computer (no claude in PATH) — the model names a CLI the backend can't find. Install it where the backend runs, or pick another model.
• Not signed in on a card — run its sign-in command in a terminal on the SynapseOrch host (the card shows it), then press Refresh.
• … is too old for how Synapse runs it — update the CLI: Codex to 0.163 or newer, Gemini CLI to 0.64 or newer.
• … ran its own tool … its answer was refused — the CLI answered through a tool of its own instead of the agent's. SynapseOrch refuses that answer, because it can't see or govern what the CLI's tool did. Ask again; if it keeps happening with one model, pick another.