Inside the coding workflow
AX Code is an agent harness: it connects a model to project tools, session records and review workflows. The value is what you can inspect, verify and recover after an agent edits your code.
This walkthrough uses an illustrative task. Commands below describe real AX Code interfaces; replace example session IDs with your own. No model is running in this page.
Open AX Code in your repository and use /connect to choose a supported local runtime, cloud API, coding CLI or private GPU endpoint. Choose your sandbox and permission settings before granting access to unfamiliar code. Ordinary sessions do not automatically enforce your project’s test suite.
ax-code --sandbox read-onlyAsk the agent to find where empty input is accepted and identify the relevant tests. Review its findings, enable the write scope you intend to allow, and request a focused fix. A reusable skill can supply your team’s investigation or review instructions.
Find why whitespace-only input is accepted. Show the relevant code and tests before making changes.
Run the repository’s checks and inspect the diff. Use the session graph to see recorded steps and tool calls. A recorded check failure remains a failure; the agent’s completion message is not proof that a patch is correct.
git diff
ax-code graph <session-id>Compare two sessions to review changes in execution and risk signals. List available rollback points before choosing a recovery step. Recovery depends on retained snapshots and cannot undo external side effects such as sent messages or remote deployments.
ax-code compare <session-a> <session-b>
ax-code rollback <session-id> --list安裝支援的 CLI 並登入，再使用下方的供應商 ID 將它連接至 AX Code。介接器會沿用 CLI 的本機登入狀態與工作階段。
| Coding CLI | Provider ID | Local command |
|---|---|---|
| Claude Code | claude-code | claude |
| Codex CLI | codex-cli | codex |
| Grok Build CLI | grok-build-cli | grok |
| Kimi Code CLI | kimi-cli | kimi |
| Muse Code CLI | muse-cli | muse |
Install and sign in to the vendor CLI first, then connect its provider. For example:
ax-code providers login codex-cliFor Muse Code CLI, install muse, sign in, then connect it to AX Code:
muse login
ax-code providers login muse-cliUse AX Code’s skills and repository checks to structure the task, then review the recorded session and resulting diff. Capabilities and visibility depend on the adapter and CLI; integration alone does not establish higher coding accuracy or make every nested action visible.
CLI provider setup and requirements ↗Choose Claude Code, Codex CLI, Muse Code CLI, Grok Build CLI or Kimi Code CLI as a provider, then combine it with supported cloud APIs, local models and private GPU endpoints. Use Council to gather independent reviews from multiple connected providers, or Arena to compare candidate solutions. Available members depend on your connections and model capabilities. Council is advisory; you review the results and decide what to keep.
Multi-model workflow guide ↗Connect a private GPU cloud or your own GPU server through an OpenAI-compatible endpoint. In /connect, choose Private GPU cloud → Custom provider, enter the endpoint URL and API key, then select a discovered model. The server must expose model discovery and support the model capabilities needed for coding tasks.
Arena implement mode runs candidate solutions in isolated Git worktrees and evaluates them against repository checks. Verified candidates rank ahead of candidates that fail verification. You review the evidence and choose what to keep; AX Code does not automatically merge a winner.
Read the documented execution and arena capabilities ↗AX Code replaces blind code edits with a dual-layer semantic intelligence architecture:
Direct integration with active language servers across TypeScript, Rust, Go, Python, and Ruby. The lsp tool provides definition jumping, reference finding, call hierarchies (incomingCalls / outgoingCalls), and aggregated compiler diagnostics (diagnosticsAggregated). When an agent edits code, live typecheck and compiler errors are fed directly back into the execution loop for immediate self-correction.
Populated via ax-code index, the persistent code graph provides repository-wide structural intelligence without burning LLM context window. Query callers and callees, compute exact blast radius impact before making changes (impact_analyze), scan for async race conditions (race_scan), and catch resource lifecycle leaks (lifecycle_scan).
Large codebases require architectural grounding, not just grep. AX Code includes a native repository wiki compiler (@automatosx/ax-wiki) accessed via ax-code wiki:
Compiles repository architecture, module responsibilities, conventions, and data flows into a clean Markdown knowledge base under ax-wiki/, linking directly to verified source symbols.
ax-code wiki update compares source hashes against the wiki manifest to regenerate only pages affected by recent git commits, keeping documentation in sync without expensive full rebuilds.
Preserve human knowledge with <!-- AX-WIKI:PROTECTED:START --> markers. Architecture decisions, production caveats, and security invariants survive automated updates.
Compiles compact wiki cards (.ax-code/wiki-cards.md) and injects knowledge pointers into AGENTS.md, ensuring the agent understands module boundaries and constraints before proposing changes.
Produces responses and tool-use decisions. Qwen and DeepSeek models are examples. Their speed and quality depend on the exact model, serving setup and task.
Connects the model to tools and manages the execution loop. AX Code, Codex, DeepSeek Harness and Hermes Agent belong in this comparison. DeepSeek Harness is a separate project from DeepSeek’s model API.
Reusable task instructions and optional supporting resources. Skills guide the agent; they do not replace the runtime or guarantee correct output. AX Code supports project skills using SKILL.md.
Evaluate tool boundaries, session evidence, recovery, customization and verification using the same task. Shared features such as skills or local models are not evidence that one tool is uniquely capable.