Anthropic Messages

The Anthropic Messages API (POST {base_url}/v1/messages) with an API key. Use this when you have an Anthropic API key, billed per token; to bill a Claude subscription instead, see claude-subscription, which speaks the same protocol.

The protocol is not Anthropic’s alone. Databricks, OpenRouter, Vercel AI Gateway, LiteLLM, Synthetic and Ollama all serve /v1/messages, as does Amazon Bedrock on its Anthropic-compatible host (https://bedrock-mantle.{region}.api.aws/anthropic, with an API key, not bedrock-runtime, which is SigV4 and /model/{id}/invoke). This backend reaches any of them via base_url, which is why it is named for the protocol rather than for Claude.

Configuration

Quickest Start

meka provider add anthropic --type anthropic-messages --model claude-opus-5

meka provider add prompts for your Claude API key, stores it in the database, and writes the [providers.anthropic] profile. To read the key from a pipe instead of prompting, pass --api-key-stdin.

Config File

meka provider add writes this for you (the key stays in the database, not here):

default_provider = "anthropic"

[providers.anthropic]
type = "anthropic-messages"
model = "claude-opus-5"

effort

meka sends the reasoning-effort control as output_config.effort in the request body. Unlike claude-subscription, no beta header is needed: the parameter is generally available on the direct Messages API. When effort is unset the field is omitted entirely, which is how you ask for Anthropic’s own default. See the effort config reference for the levels.

thinking

adaptive (the default) sends thinking: {"type": "adaptive"} and lets the model set its own budget; budgeted sends the older {"type": "enabled", "budget_tokens": N} form, taking N from the profile’s thinking_budget and falling back to [thinking].budget_tokens; off sends no thinking field. Pre-4.6 Claude models require budgeted.

Supported Models

Any model available through the Claude Messages API; meka forwards the model string verbatim and doesn’t gate which strings are valid. For the current line-up and their retirement dates, see Anthropic’s models overview - meka provider add suggests claude-opus-5 for new Claude profiles.

Custom Base URL

To use a Claude-API-compatible proxy or gateway, set the profile’s base_url. Add it when creating the profile, or change it later:

meka provider add gateway --type anthropic-messages --model claude-opus-5 \
    --base-url https://gateway.example.com/anthropic

meka provider set gateway base_url https://gateway.example.com/anthropic

A trailing /v1 is dropped, since meka appends it per request: publish https://gateway.example.com/anthropic or https://gateway.example.com/anthropic/v1, either works.

Anthropic-compatible endpoints

The model behind the endpoint doesn’t have to be Claude. Ollama, LM Studio and similar runtimes serve local weights over POST /v1/messages, and anthropic-messages reaches them with a placeholder key:

meka provider add local --type anthropic-messages \
    --model 'hf.co/bartowski/Qwen3.8-27B-GGUF:Q8_0' \
    --base-url http://127.0.0.1:11434

Nothing in the request is tuned to Claude unless you ask for it. effort is omitted when unset, so a backend with no reasoning tiers is never handed one, and thinking is whatever the profile says rather than something inferred from the model name - set budgeted if your endpoint only implements the older encoding, or off if it implements neither.

The one setting worth stating is the context window. meka never probes for it, so an unset profile budgets against the 1M default; on a smaller model that means compaction only fires once the backend itself rejects the request:

[providers.local]
context_window = 262144
thinking = "budgeted"   # only if the endpoint rejects the adaptive form

API Details

Endpoint: POST {base_url}/v1/messages

Headers:

• x-api-key: <api_key>
• anthropic-version: 2023-06-01
• content-type: application/json
• accept: application/json
• anthropic-beta: interleaved-thinking-2025-05-14, whenever thinking is on (the default)

System prompt: Sent as a top-level system string.

Tool format: Tools are defined with input_schema:

{
  "name": "read_file",
  "description": "Read the contents of a file at the given path.",
  "input_schema": { "type": "object", "properties": { ... } }
}

Streaming: Server-Sent Events with named event types (message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop, ping).