Using with Claude Code

The easy way

copillm claude

That handles login check, daemon start, agent install/resolve, and env wiring. Done.

Manual wiring

If you’d rather drive claude yourself, copillm env claude prints the matching env block. It auto-detects the latest plain (non--high / -xhigh / -internal) variant per family from your live Copilot model list, pins them to the matching Claude Code alias env vars, and enables gateway discovery:

$ copillm env claude
# Claude Code → copillm
export ANTHROPIC_BASE_URL="http://127.0.0.1:4141/anthropic"
export ANTHROPIC_AUTH_TOKEN="copillm-local"
export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-4.7"
export ANTHROPIC_DEFAULT_SONNET_MODEL="claude-sonnet-4.6"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-haiku-4.5"
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY="1"
export CLAUDE_CONFIG_DIR="$HOME/.copillm/claude/home"

Paste it into a shell, or eval "$(copillm env claude)" to load it into the current shell, then run claude. Use copillm env claude --inline for the legacy single-line form. --shell fish and --shell powershell are also supported.

To persist the same provider wiring into Claude Code’s native settings without launching through copillm, run:

copillm config sync --agent claude

That writes the provider environment into ~/.claude/settings.json and syncs the active profile’s MCP servers into user scope in ~/.claude.json.

What each piece does

• ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN — point Claude Code at the local copillm proxy
• ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL — Claude Code resolves the opus/sonnet/haiku aliases (used by /model selections, claude --model opus, and background haiku-class tasks) to these specific Copilot variants client-side
• CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 — Claude Code v2.1.129+ calls our /anthropic/v1/models endpoint at startup and populates the /model picker with every Copilot model you’re entitled to that supports chat — not just Claude-branded ones, so Gemini and GPT variants appear too. Each appears labelled “From gateway”
• CLAUDE_CONFIG_DIR — points Claude at a copillm-owned config home (~/.copillm/claude/home) instead of your real ~/.claude. This keeps copillm-launched Claude deterministic and isolated (so dev and prod instances never collide), but it means your personal ~/.claude settings, CLAUDE.md, and subagents do not apply to copillm claude. Run copillm config sync --agent claude if you want copillm’s wiring in your real ~/.claude for a direct claude launch.

Override any env var in your shell (e.g. ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4.7-high) to pick a different Copilot variant. The copillm claude launcher does not persist Anthropic preferences; only copillm config sync --agent claude writes native Claude settings.

Context windows and the [1m] alias

Each Copilot model carries three distinct token limits in its upstream metadata:

• max_prompt_tokens — hard ceiling on input tokens in a SINGLE API call, enforced by Copilot server-side
• max_output_tokens — hard ceiling on output tokens in a single call
• max_context_window_tokens — total conversation budget across turns (input + output + cache reads)

Claude Code’s /anthropic/v1/models gateway-discovery validator only reads id and display_name per model — there is no field through which copillm can communicate a numeric context window. Without recognising the model id, Claude Code falls back to a hardcoded 200K per-model max for autocompact purposes, regardless of CLAUDE_CODE_AUTO_COMPACT_WINDOW or the autoCompactWindow setting (which can only reduce the cap, never raise it).

The only client-side marker Claude Code recognises is a literal [1m] suffix on the model id (its binary matches id.toLowerCase().includes("opus") && id.toLowerCase().includes("[1m]") for opus; the sonnet matcher requires a contiguous sonnet[1m] substring that copillm-aliased ids don’t form; no non-Claude vendor has any [1m] matcher). So when copillm sees an opus upstream model with max_context_window_tokens >= 1_000_000, it advertises the id with [1m] appended in the /anthropic/v1/models response and strips the suffix back off before forwarding any request to Copilot. Net effect:

• The /model picker entry for the model carries the [1m] suffix
• Claude Code allocates a 1M-class autocompact budget (effectiveWindow ≈ 980_000)
• Upstream still receives the canonical model id
• Per-request input is still bounded server-side at the model’s max_prompt_tokens — well above the typical fresh delta sent on any single turn thanks to prompt caching

Models with max_context_window_tokens between 200K and 1M, and non-opus models even when they exceed 1M, get no alias: Claude Code has no marker for intermediate tiers, and its 1M matcher is restricted to opus ids in practice.