claude_code
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| DEEPSEEK_API_KEY | Yes | DeepSeek API key, passed to the Claude Code subprocess as ANTHROPIC_API_KEY. Can alternatively be stored in ${XDG_CONFIG_HOME:-$HOME/.config}/claude-codex-bridge/deepseek-api-key with 0600 permissions. | |
| CLAUDE_BRIDGE_BASE_URL | No | Anthropic-compatible base URL used by Claude Code. | https://api.deepseek.com/anthropic |
| CLAUDE_BRIDGE_CLAUDE_BIN | No | Path to the Claude Code executable; defaults to `claude` (Windows: `claude.cmd`). | claude |
| CLAUDE_BRIDGE_PROJECT_DIR | No | Default working directory for tasks; defaults to process.cwd(). | process.cwd() |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server bridges to Claude Code (headless, DeepSeek-backed) for this project. Use it as follows:
Delegate vs. do it yourself: use claude_run for substantial work (new features, multi-file changes, anything needing iterative test-and-fix). For small, obviously-scoped edits (single file, a config tweak, read-only analysis) just do it directly yourself — don't delegate trivial work.
Before the first claude_run call for a fresh task, check whether the git working tree is clean (git status --porcelain). If it is not, stop and ask the user how to proceed instead of running Claude Code against a dirty tree.
Never trust result_text alone. After each call, verify the change yourself (git diff, run the project's tests/build) before deciding whether it meets the request.
If it doesn't meet the request, call claude_run again with the same resume_session_id and a specific description of what's wrong. Cap this at around 4 rounds total; if still unresolved, stop and report the situation to the user instead of continuing to retry.
Once satisfied, make a single local git commit summarizing the change (include the returned session_id in the commit body, so the native Claude Code transcript can be found later via ~/.claude/projects//<session_id>.jsonl). Never push to a remote automatically — that stays a manual, human decision.
Only interrupt the user when the task is fully done, or when you're stuck. Don't report progress after every intermediate round.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| claude_runA | Run Claude Code (headless, DeepSeek-backed) to write/edit code and run commands in the project. Blocks until Claude Code finishes and returns its result. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 1 tool
There is only one tool, so there is no possibility of selecting the wrong one. Its purpose (headless Claude Code execution) is stated clearly and unambiguously.
With a single tool named claude_run, there is no opportunity for convention mixing. The name is readable and follows a noun_verb pattern that fits the server's identity.
A single blocking run tool is a thin surface for a coding-agent server; there is no way to list runs, check status, or cancel. It is defensible for a minimal wrapper but leaves the server feeling under-scoped.
The tool blocks until completion, so there is no async execution, status polling, cancellation, or output streaming. For long-running code edits and command execution these are significant gaps that will cause agent timeouts or dead ends.