Anthropic-compatible API¶

POST https://anthropic.keln.ai/v1/messages, the Anthropic Messages schema, streaming and non-streaming, for every catalog model. Built for tools that speak the Anthropic API, Claude Code works against Keln with three environment variables (setup below).

Same account, same key, same models, same prices: only the request shape differs.

curl https://anthropic.keln.ai/v1/messages \
  -H "x-api-key: $KELN_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k3",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Say hello from Keln."}]
  }'

Authentication¶

Send your normal Keln key in the x-api-key header, the Anthropic convention. The anthropic-version header is accepted as clients send it.

Models¶

Use Keln model ids (moonshotai/kimi-k3, deepseek-ai/deepseek-v4-flash, …), the ids from /v1/models, not Anthropic model names. Anything on the rate card is available on this surface.

What's supported¶

Feature	Contract
`messages`, `system`, `max_tokens`	standard Messages semantics
`stream: true`	the full Anthropic event lifecycle: `message_start` → `content_block_start` → `content_block_delta` → `content_block_stop` → `message_delta` → `message_stop`
Thinking models	the reasoning trace comes back as `thinking` content blocks, exactly where Anthropic clients expect it
`tools`, `tool_choice` (incl. forced `{"type": "tool", "name": …}`)	Anthropic `input_schema` tool shape, `tool_use` content blocks out, `tool_result` blocks in, full agent loops work
`usage`	`input_tokens` / `output_tokens` on every response; billing is identical to the OpenAI-shaped surfaces
Everything else	the same routing, validation, and zero-data-retention contracts as the rest of Keln

Using Claude Code¶

Point Claude Code at Keln with environment variables, no config file needed:

export ANTHROPIC_BASE_URL="https://anthropic.keln.ai"
export ANTHROPIC_API_KEY="sk-keln-…"          # your Keln key
export ANTHROPIC_MODEL="moonshotai/kimi-k3"   # any catalog id with the Tools badge

Then run claude as usual.

Notes that matter in practice:

Model choice: agent sessions need tool calling, pick a model with the Tools badge.
Long sessions: repeated context automatically bills at the per-model cached rate; nothing to configure.
ANTHROPIC_SMALL_FAST_MODEL can point at a cheaper catalog model (e.g. deepseek-ai/deepseek-v4-flash) for Claude Code's background tasks.

Web search¶

Send Anthropic's server tool, { "type": "web_search_20250305", "name": "web_search" }, and Keln runs the search and answers in Anthropic's own shape: server_tool_use and web_search_tool_result blocks with the results, citations on the text, and usage.server_tool_use.web_search_requests. Claude Code's WebSearch uses this tool, so it works through Keln as it does against Anthropic. Send the result block back on the next turn and the model keeps its findings. Details and price on Web search.