Using Zhipu GLM (and Other Anthropic-Compatible Endpoints)

Alfred's Anthropic provider reads ALFRED_BASE_URL (or ANTHROPIC_BASE_URL) to point the Messages API at any compatible endpoint. Zhipu's GLM models expose exactly that interface, so switching is a two-variable change with no code edits.

How it works

src/providers/anthropic.ts constructs the Anthropic SDK client with:

new Anthropic({
  apiKey: config.apiKey ?? process.env.ANTHROPIC_API_KEY,
  baseURL: config.baseUrl ?? process.env.ALFRED_BASE_URL ?? process.env.ANTHROPIC_BASE_URL,
})

When ALFRED_BASE_URL is set to a GLM endpoint the SDK sends every /v1/messages call there instead of api.anthropic.com. The rest of Alfred — tool loop, hooks, skills, ledger — is unchanged.

Environment setup

# The GLM Anthropic-compatible endpoint
export ALFRED_BASE_URL=https://open.bigmodel.cn/api/anthropic

# Your Zhipu API key (reused in the ANTHROPIC_API_KEY slot)
export ANTHROPIC_API_KEY=<your-zhipu-api-key>

Why ANTHROPIC_API_KEY?

The SDK always sends the key in the x-api-key header regardless of the base URL. Zhipu's Anthropic-compatible endpoint reads that same header.

One-shot query with GLM

alfred --model glm-5.1 "Summarise the last five commits in this repo."

Or set the model globally to avoid repeating the flag:

export ALFRED_MODEL=glm-5.1
alfred "List all exported functions in src/tools/index.ts"

Available GLM model ids

Alfred's pricing table (src/cost/tracker.ts) includes three GLM models:

Model id	Input (per 1M)	Output (per 1M)	Cache read	Cache write
glm-4.5	$0.60	$2.20	$0.11	$0.75
glm-4.6	$0.60	$2.20	$0.11	$0.75
glm-5.1	$0.60	$2.20	$0.11	$0.75

These are approximately 5× cheaper than claude-sonnet-4-6 on input tokens.

Autonomous run with GLM

export ALFRED_BASE_URL=https://open.bigmodel.cn/api/anthropic
export ANTHROPIC_API_KEY=<your-zhipu-api-key>
export ALFRED_LEDGER_SECRET=$(openssl rand -hex 32)

alfred run \
  --model glm-5.1 \
  --verify "bun test" \
  --feature-list feature_list.json \
  --budget-usd 1.00

The --budget-usd flag is especially useful with GLM because the cost estimator uses the pricing table above — setting a dollar cap prevents runaway spending on large feature lists.

Role routing with GLM

Alfred supports per-role model overrides (ADR 0005). You can mix providers within a single run by assigning different roles to different models:

# Use GLM for fast editor work; keep the default model for architecture
export ALFRED_MODEL_EDITOR=glm-5.1
export ALFRED_MODEL=claude-sonnet-4-6

alfred run --verify "bun test"

Role env variables:

Variable	Role	Purpose
ALFRED_MODEL_ARCHITECT	architect	Planning and high-level reasoning
ALFRED_MODEL_EDITOR	editor	Applying file edits
ALFRED_MODEL_SUBAGENT	subagent	Delegated subtasks

Roles not set fall back to ALFRED_MODEL (or the --model flag).

Alternative: ANTHROPIC_BASE_URL

If you prefer not to set ALFRED_BASE_URL you can use the SDK's own env variable:

export ANTHROPIC_BASE_URL=https://open.bigmodel.cn/api/anthropic

ALFRED_BASE_URL takes precedence when both are set.

Pointing at any compatible endpoint

The same pattern works for any service that implements the Anthropic Messages API:

# Local proxy or self-hosted model
export ALFRED_BASE_URL=http://localhost:8080

# Another cloud provider's Anthropic-compatible gateway
export ALFRED_BASE_URL=https://gateway.example.com/anthropic

alfred --model my-custom-model "Hello"

Model id mismatch

If the endpoint rejects your model id, check the provider's documentation for the exact string they expect. Alfred passes the value of --model (or ALFRED_MODEL) verbatim to the model field in the API request.

Streaming and token counting

Some compatible endpoints do not implement streaming (client.messages.stream) or token counting (client.messages.countTokens). Alfred falls back gracefully — streaming failure is caught and re-thrown as a ProviderError; token counting is optional and only used for context-window management.