Using Claude Code with Sail

Run Claude Code on GLM-5.3 via Sail: one command, no Anthropic account.

Claude Code speaks the Anthropic Messages API to whatever endpoint ANTHROPIC_BASE_URL points at. Sail serves that API for GLM-5.3 (system prompts, tool calling, and streaming included), so pointing Claude Code at Sail is just environment configuration. The command below does it in one step.

Want to keep Claude Code on its current model and send selected work to Sail instead? Install the Sail plugin for coding agents.

Quickstart (terminal)

Install Claude Code if you haven’t, then launch it routed to Sail:

npm install -g @anthropic-ai/claude-code            # once
curl -fsSL https://cli.sailresearch.com/install.sh | sh   # once
sail auth login                                     # once
sail claude

That starts claude on GLM-5.3. Nothing on your machine is modified: the routing applies to that one session, so your Claude Code setup and your Anthropic credentials are left alone. If you prefer not to store a key, set SAIL_API_KEY in your environment and skip sail auth login.

Three commands, three scopes

Command	What it affects
sail claude	This session only. Does not write anything.
sail claude on	Every Claude Code that reads your user settings, including a bare claude and the VS Code extension.
sail claude off	Undoes sail claude on.

You never need sail claude on to use sail claude. Turning it on is only for reaching the entry points you do not launch yourself, and sail claude keeps working after you turn it off.

One exception to “every Claude Code”: routing set in a repository’s own .claude/settings.json takes precedence over your user settings, so a session started there keeps using that repository’s routing. Run sail claude on --project inside such a repository to route it too.

Trusted workspaces (no permission prompts)

For trusted repos, containers, or VMs where you want long-running agent tasks without clicking through Claude Code permission prompts, use:

sail claude --trusted

--trusted launches Claude Code in bypass-permissions mode. This avoids Claude Code’s auto-mode safety classifier path, so it also avoids failures like safety classifier temporarily unavailable when Sail model requests are otherwise healthy. Use it only where you are comfortable letting Claude Code run tools without prompting.

This works inside a Sailbox too: Sail sets IS_SANDBOX=1 on Sailbox commands by default, which Claude Code recognizes as a sandbox, so bypass-permissions mode is allowed there even for root (the default user for Sailbox commands when the image does not set one). Setting IS_SANDBOX yourself on a command replaces the default for that command. A Sailbox created before this marker existed gains it after an upgrade.

Pass flags through to claude after --:

sail claude -- --resume

VS Code extension

The VS Code extension isn’t launched from your shell, so it doesn’t receive per-invocation environment variables. Persist the routing instead:

sail claude on    # writes ~/.claude/settings.json
sail claude off   # removes exactly what `sail claude on` added

sail claude off removes only the keys sail claude on wrote. Settings that changed while it was on, whether you changed them or Claude Code did, are kept.

Extension sessions read that env block, but the extension’s own pre-launch login check does not. If you don’t have a saved Anthropic login, also add the routing to VS Code’s user settings (Preferences: Open User Settings (JSON)) and reload the window. sail claude on prints this snippet:

"claudeCode.disableLoginPrompt": true,
"claudeCode.environmentVariables": [
  { "name": "ANTHROPIC_BASE_URL", "value": "https://api.sailresearch.com" },
  { "name": "ANTHROPIC_AUTH_TOKEN", "value": "<your SAIL_API_KEY>" }
]

Reload the extension after configuring. Use sail claude on --project to scope the routing to one repository. It writes your API key into .claude/settings.local.json (Claude Code’s personal, not-shared project file) and adds that file to .claude/.gitignore so it stays out of commits. Claude Code applies project settings only after you trust the folder in its first-run prompt.

Choosing models

GLM-5.3 is the default. The /model picker lists five Sail-hosted rows:

Picker row	Model id	Also selectable with
GLM-5.3 via Sail (default)	zai-org/GLM-5.3	/model opus
DeepSeek V4 Pro via Sail	deepseek-ai/DeepSeek-V4-Pro-0813	/model sonnet
GLM-5.3-Flash via Sail	zai-org/GLM-5.3-Flash	/model haiku
Kimi K3 via Sail	moonshotai/Kimi-K3	/model fable
DeepSeek V4.1 Flash via Sail	deepseek-ai/DeepSeek-V4.1-Flash	/model <id>

You can switch models in these ways:

• Mid-session: open /model and pick a row, or type /model <id> for any id in the table. The choice lasts for this session only: Sail pins the startup model, so the next launch starts on the configured main model again. Use the two options below to persist a different one.
• At launch: --model <id> starts on that model. Any id from the model catalog works, not only the five rows. A model outside the lineup is added to the picker for that session.
• Persistently: sail claude on --model <id> pins it for every Claude Code that reads your settings.

sail claude --model moonshotai/Kimi-K3

GLM-5.3-Flash handles Claude Code’s background (haiku-tier) work by default. --background-model <id> changes that. Subagents use the main model either way.

The table above shows the default rows. The rows are Claude Code’s opus, sonnet, haiku, and fable model aliases plus one custom row, and Sail points each one at a different Sail model. The main model always takes the opus row, which the picker lists first, and the background model takes the haiku row. The remaining Sail models fill the other rows, so no model appears twice. For example, with --model moonshotai/Kimi-K3, Kimi K3 moves to the first row and GLM-5.3 moves to the sonnet row.

sail claude and sail claude on write Claude Code’s availableModels setting so fresh local sessions show only Sail-backed picker rows, with the main model first so the picker’s Default row resolves to it. Claude Code concatenates non-managed availableModels from user, project, local, and temporary settings, though, so rows from another settings scope may still appear. Rows that name non-Sail models (such as claude-* IDs) still route to Sail and will not be served there. Remove them from the other settings file or enforce the picker with managed settings. sail claude on prints a warning when it can see those extra rows.

Choosing a completion window

Claude Code runs on Sail’s default completion window, asap, unless you choose another one. Lower-priced windows trade latency for cost:

sail claude --completion-window balanced

• Accepts asap or balanced.
• Covers every request the session sends: the main model, background work, and subagents.
• sail claude on --completion-window balanced persists the window for every Claude Code that reads your settings, such as a bare claude or the VS Code extension. Run sail claude on without the flag to return to the default.
• sail claude sets each session up from its own flags, so a window saved with on does not carry over. Pass --completion-window to sail claude as well.
• balanced suits long agentic sessions: turns can take longer, at lower token prices. See Pricing.
• The /model picker lists asap prices whichever window you choose.

flex is unavailable in Claude Code because its Messages API requests wait for each turn. Use balanced for lower-cost sessions. If an earlier setup saved flex, run sail claude on --completion-window balanced to replace it, or sail claude off to remove Sail routing.

The flag sets Claude Code’s ANTHROPIC_CUSTOM_HEADERS to X-Sail-Completion-Window: <window>, the header Sail reads a completion window from when a request body names none.

Commit attribution

sail claude on sets Claude Code’s top-level attribution.commit to a Sail co-author trailer naming the resolved model:

Co-Authored-By: GLM-5.3 via Sail <[e-mail]>

Commits and PRs the agent makes are then attributed to the Sail-served model instead of Anthropic’s default model-derived trailer. A --model override names that model in the trailer instead (so a non-default setup isn’t misattributed to GLM-5.3). attribution.pr is emptied (no PR-body attribution line). sail claude off puts your original attribution back. This only affects the persisted path: a plain sail claude session does not touch attribution.

Manual configuration

If you’d rather configure Claude Code yourself, set the routing environment by hand (sail claude env prints the full block, including picker display metadata):

export ANTHROPIC_BASE_URL="https://api.sailresearch.com"   # bare host: Claude Code appends /v1/messages
export ANTHROPIC_AUTH_TOKEN="$SAIL_API_KEY"                # bearer auth (recommended); ANTHROPIC_API_KEY also works
export ANTHROPIC_MODEL="zai-org/GLM-5.3"               # pins startup over a saved /model choice
export ANTHROPIC_DEFAULT_OPUS_MODEL="zai-org/GLM-5.3"                    # one Sail model per alias row
export ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-ai/DeepSeek-V4-Pro-0813"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="zai-org/GLM-5.3-Flash"             # haiku = background work
export ANTHROPIC_DEFAULT_FABLE_MODEL="moonshotai/Kimi-K3"
export ANTHROPIC_CUSTOM_MODEL_OPTION="deepseek-ai/DeepSeek-V4.1-Flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL_NAME="GLM-5.3 via Sail"              # picker labels; _DESCRIPTION too
export ANTHROPIC_DEFAULT_SONNET_MODEL_NAME="DeepSeek V4 Pro via Sail"
export ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME="GLM-5.3-Flash via Sail"
export ANTHROPIC_DEFAULT_FABLE_MODEL_NAME="Kimi K3 via Sail"
export ANTHROPIC_CUSTOM_MODEL_OPTION_NAME="DeepSeek V4.1 Flash via Sail"
export CLAUDE_CODE_SUBAGENT_MODEL="zai-org/GLM-5.3"     # pins subagents too; overrides frontmatter + any inherited value
export ANTHROPIC_DEFAULT_OPUS_MODEL_SUPPORTED_CAPABILITIES="effort,thinking"                  # custom model IDs match no known
export ANTHROPIC_DEFAULT_SONNET_MODEL_SUPPORTED_CAPABILITIES="effort,thinking,xhigh_effort"   # pattern, so effort + extended
export ANTHROPIC_DEFAULT_HAIKU_MODEL_SUPPORTED_CAPABILITIES="effort,thinking"                 # thinking would be silently
export ANTHROPIC_DEFAULT_FABLE_MODEL_SUPPORTED_CAPABILITIES="effort,thinking,xhigh_effort"    # disabled without these
export ANTHROPIC_CUSTOM_MODEL_OPTION_SUPPORTED_CAPABILITIES="effort,thinking,xhigh_effort"
export API_TIMEOUT_MS="3000000"
export API_FORCE_IDLE_TIMEOUT="0"                          # off: else a 5-min stream-idle watchdog (on for custom base URLs) aborts long-queued turns
export ENABLE_TOOL_SEARCH="false"                          # tool search needs tool_reference expansion Sail doesn't serve
# Clear leftovers from other provider setups. Selectors outrank
# ANTHROPIC_AUTH_TOKEN, and stale custom headers would be sent to Sail:
unset CLAUDE_CODE_USE_BEDROCK CLAUDE_CODE_USE_VERTEX CLAUDE_CODE_USE_FOUNDRY \
  CLAUDE_CODE_USE_MANTLE CLAUDE_CODE_USE_ANTHROPIC_AWS ANTHROPIC_CUSTOM_HEADERS
# Optional: choose a completion window (the default is asap).
# export ANTHROPIC_CUSTOM_HEADERS="X-Sail-Completion-Window: balanced"
claude

Notes on these settings:

• Use the bare host. Claude Code appends /v1/messages itself. A /v1 base URL would resolve to /v1/v1/messages and return a 404.
• ANTHROPIC_AUTH_TOKEN is what Sail expects. Sail also accepts ANTHROPIC_API_KEY, which Claude Code sends through x-api-key. Set only one credential variable to avoid ambiguous client precedence.
• Pin ANTHROPIC_MODEL and remap every alias. A saved /model choice (or a --resumed session) can otherwise start on an unmapped claude-* model ID that Sail cannot serve. Each alias row can point at a different Sail model, which is how the five-row picker above is built: Claude Code offers four alias rows plus one custom row, and the _NAME companions label them.
• Pin CLAUDE_CODE_SUBAGENT_MODEL. It’s the highest-priority subagent model source, so a value inherited from your shell or a settings file (e.g. a claude-* ID left over from a prior Anthropic/Bedrock setup) would send every subagent to an unmapped model. Setting it to a Sail model routes all subagents onto GLM.
• Declare _SUPPORTED_CAPABILITIES. Claude Code enables effort levels and extended thinking by matching the model ID against known patterns. A custom ID like zai-org/GLM-5.3 matches none, so /effort and thinking are silently disabled without the declaration. Sail maps both onto each model’s reasoning controls. The DeepSeek and Kimi rows also declare xhigh_effort.
• Unset other providers’ variables. A provider selector such as CLAUDE_CODE_USE_BEDROCK comes before ANTHROPIC_AUTH_TOKEN in Claude Code’s credential precedence, so a leftover export would silently keep routing to that provider instead of Sail. ANTHROPIC_CUSTOM_HEADERS from a prior gateway setup would be sent to Sail on every request. sail claude strips these automatically, and sets ANTHROPIC_CUSTOM_HEADERS only to the Sail completion window header when you pass --completion-window. By hand, the unset line above does the same.

Limitations

• The native Claude Code desktop app reads endpoint routing only from managed (administrator-distributed) configuration. It cannot be pointed at Sail with environment variables or settings.json. Use the terminal or the VS Code extension.
• A signed-in Claude apps gateway session (an enterprise/corporate-SSO deployment) is not part of the normal credential precedence and overrides ANTHROPIC_AUTH_TOKEN, so claude keeps using the gateway and never sends requests to Sail. This only affects orgs running that gateway. Ordinary Pro/Max users don’t have one. If you do, run /logout first, or launch that session with CLAUDE_CONFIG_DIR pointed at an empty directory (note: an isolated config dir won’t see your normal Claude Code settings/MCP servers).
• An administrator-managed (MDM / policy) availableModels allowlist that excludes the Sail model takes precedence. Claude Code replaces the model at startup (with a warning), and a managed allowlist cannot be overridden because managed settings take precedence over the per-session settings sail claude uses. That’s an enterprise policy. Raise it with your admin. (If you set availableModels yourself in your own settings.json, just include the Sail model id, e.g. zai-org/GLM-5.3, or drop the restriction.)
• Sail is throughput-optimized. Long agentic turns can take longer than the Anthropic API. sail claude raises Claude Code’s request timeout and disables its 5-minute stream-idle watchdog (on by default for custom base URLs) so a turn that queues before streaming isn’t aborted mid-flight.
• Subagents run on the main Sail model by default: sail claude pins CLAUDE_CODE_SUBAGENT_MODEL, which overrides a subagent’s frontmatter model: field. That makes custom subagents with a claude-* frontmatter ID work (they route to the main model instead of failing), but it also means per-subagent model choices and --background-model don’t apply to subagents: they all use the main model. Override it yourself if you need a specific Sail model for subagents.
• Background (haiku-tier) work runs on GLM-5.3-Flash by default rather than the main model. Pass --background-model zai-org/GLM-5.3 to keep everything on one model.

Next steps

Models

Browse the catalog for alternative models.

Pricing

Per-token rates by model and completion window.

Completion windows

How the latency-for-price tradeoff works.

Support matrix

What the Anthropic-compatible API supports.

AI Quickstart

Give Claude Code Sail’s docs MCP and workflow skills, including sail-migrate.

Sail for coding agents

Delegate scoped work or request a review from Sail.

Migrate to the Sail API

Move an existing app or agent to Sail.