DeepSeek in Claude Code: Excession gateway setup
Configure the Claude Code CLI to use DeepSeek V4.1 Flash Abliterated through Excession's Anthropic-compatible endpoint. This configuration passed a local client integration test; live inference and coding tools were outside that test.
Understand the compatibility boundary
This is a third-party compatibility setup. Anthropic's gateway documentation says it does not support routing Claude Code to non-Claude models, and DeepSeek is not a Claude model. This guide does not imply Anthropic endorsement or complete Claude Code feature compatibility.
Set the URL, key and model
Install the CLI using the official Claude Code instructions. Create an Excession API key in the dashboard and load it into your terminal's EXCESSION_API_KEY environment variable using your usual secret manager. Check your funds and current model pricing on the API page.
In Bash or Zsh, run:
: "${EXCESSION_API_KEY:?Load your Excession API key first}"
export ANTHROPIC_BASE_URL="https://api.excessionlabs.ai"
export ANTHROPIC_AUTH_TOKEN="$EXCESSION_API_KEY"
unset ANTHROPIC_API_KEY CLAUDE_CODE_OAUTH_TOKEN
claude --safe-mode \
--model "deepseek-v4.1-flash-abliterated" \
--tools "" \
-p "Reply with READY."The base URL is the host without /v1, because Claude Code appends /v1/messages. The ANTHROPIC_AUTH_TOKEN variable carries your Excession key and sends it as Authorization: Bearer …, as Anthropic's gateway connection documentation describes. These exports apply to this terminal and programs launched from it.
The smoke check disables tools with --tools "" and uses --safe-mode to keep local customizations out of the first request. A successful live run should return a short answer. Unlike our mock test, it uses metered API access on your account.
Open an interactive session
claude --model "deepseek-v4.1-flash-abliterated"Run this in the same terminal, then use /status to confirm the base URL and credential source. Our fixture did not exercise reading, tool calls or edits. Test the ones you need on a disposable example before you permit project changes.
Troubleshoot the connection
- Authentication error: confirm the token is an Excession key and available to this process. Never paste it into an issue or log.
- 404 or wrong endpoint: use
https://api.excessionlabs.aias the base URL; appending/v1here can duplicate the path. - Wrong model: pass the full model ID with
--model. Built-in Claude aliases refer to other models. - Exports seem ignored: inspect
/status, settings and shell launchers. Our host had a wrapper forcing another gateway, so validation used the official executable directly. - Funds or rate limit error: inspect your dashboard and API error. A mock configuration test cannot validate account entitlement.
- Text succeeds but tools fail: inspect non-secret request metadata and versions. Anthropic's compatibility guide describes the protocol surface. A text answer does not verify every feature.
What we tested
On 4 November 2026 we ran the actual Claude Code 2.1.318 and OpenCode 1.19.5 clients against an isolated localhost fixture with fake credentials. Each sent the exact model ID, a Bearer credential and a streaming request to its expected endpoint. Both displayed LOCAL_FIXTURE_OK from the fixture and exited successfully.
This was a mock client integration test with no live DeepSeek inference. The fixture supplied the answer, and no model generated it. The test validates configuration, request routing and text-stream parsing. It does not establish production availability, model quality, billing, long conversations, or tool-call and file-edit compatibility.
Inspect the recorded request metadata and results or download the local validation script. Run the script with --claude /path/to/claude --opencode /path/to/opencode, using official client executables. It starts a loopback endpoint, uses fake keys and temporary client configuration, and does not read a real API key. Package setup and client startup may contact public package or model metadata services.