ClawCodex

What does this agent do, and when should you use it?

ClawCodex is a Python implementation of a Claude Code-style coding agent with both an interactive terminal interface and scriptable CLI entry points. Its default interface is a TypeScript Ink TUI that launches and controls a Python agent-server over an NDJSON pipe, while headless work runs through `clawcodex -p`. The agent loop streams responses, invokes filesystem and shell tools, persists sessions, and extends slash commands through `SKILL.md` files. It documents support for Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, OpenAI-compatible vendors, and local Ollama, vLLM, and SGLang servers. It also includes permission modes, MCP tooling and resources, plan mode, scheduled tasks, and `/eco` Bash-output compression; adopters should account for its default interactive Full Access mode and dependence on external model services.

When a user starts clawcodex, the Ink TUI sends prompts, @ file references, and slash commands to the Python agent-server; clawcodex -p "..." runs a one-shot task instead. The agent loop receives streamed model responses from the configured provider and model, calls tools including Read, Write, Edit, Glob, Grep, Bash, WebFetch, WebSearch, AskUserQuestion, TodoWrite, and Agent, then returns tool results to the model until the task finishes. clawcodex login stores provider credentials, a default model, and optional TAVILY_API_KEY configuration in ~/.clawcodex/config.json, while sessions and permission settings can persist locally. /eco deterministically compresses Bash output before it enters model context while retaining recoverable full output, and clawcodex mcp serve exposes ClawCodex tools as an MCP stdio server.

1. A Python or full-stack developer working in a local Git repository who wants a terminal agent to inspect source, edit files, run tests, and execute shell commands.
2. An engineering team that needs one-off coding tasks in CI, scripts, or automation through `clawcodex -p` and JSON or NDJSON input/output.
3. A team that wants to switch among Anthropic, OpenAI, DeepSeek, Gemini, or local Ollama/vLLM/SGLang models without changing the core tool workflow.
4. A developer debugging failing suites or long-running commands who wants `/eco` to reduce Bash-log tokens sent to the model while preserving the raw output.
5. A repository maintainer defining custom, tool-restricted slash commands in `.clawcodex/skills/<skill-name>/SKILL.md`.
6. A terminal user who needs an agent session to check a build, deployment, or maintenance prompt on a cron schedule or through `/loop`.

What are this agent's strengths and limitations?

How do you install or deploy this agent?

You need Python 3.10+, network access, and an API key for at least one provider or a supported subscription login. Recommended installation:
pipx install clawcodex-cli
Then run:
clawcodex login
Choose a provider and configure credentials interactively, then verify with:
clawcodex --help
For a source checkout:
git clone https://github.com/agentforce314/clawcodex.git && cd clawcodex
uv venv --python 3.11 && source .venv/bin/activate
uv pip install -e ".[dev]"
The interactive Ink TUI additionally needs Node.js 18+ and a built ui-tui/dist; the headless -p path does not need Node.

How do you use this agent?

After login, run clawcodex in the target project to open the interactive TUI, or use clawcodex tui explicitly. Reference a file with @src/cli.py, inspect available tools with /tools, or invoke a project skill. For one-shot work, run clawcodex -p "Summarize src/cli.py"; to select a model, run clawcodex --provider anthropic --model claude-sonnet-4-6 -p "Find TODOs". Add TAVILY_API_KEY to the configuration env block to use WebSearch. Interactive sessions start in Full Access, so use /permissions to reduce access before work, or use --permission-mode plan for headless runs.

How does this agent compare with similar options?

FAQ