English documentation · runtime 7.24.4 · SDK 2.6.7. Content is maintained with runtime development; see each guide's scope and review date.

AX Code Documentation

Status: Active Scope: public, current-state Last reviewed: 2026-10-09 Owner: AX Code maintainers

AX Code is a coding-agent runtime for reviewable, reversible work: every session is recorded as a structured event log with file snapshots, candidate implementations can be verified against your repository’s own checks, and nothing merges automatically. Why AX Code explains what that optimizes for, who it is for, and what it deliberately does not claim.

The root README is the shortest path to install and launch AX Code. Use this hub when you need to configure a workflow, understand a runtime boundary, or integrate AX Code with another system.

Internal process records (PRDs, ADRs, specifications, roadmaps, implementation plans, competitive reviews, and audit working files) live under .internal/, not in this public tree — see “Documentation boundaries” below. The public pages are the ones linked here.

Public website: ax-code.app/en/docs/. For existing installations, start with Upgrade and recovery. Source access and public downloads are separate; see Source access.

Choose by task

For opt-in shared terminal navigation through an external MCP client, see Share a live TUI through MCP.

For the experimental browser bridge (read pages, optionally click and fill), see WebMCP browser bridge.

For standalone questions about selected files through AX Trust caching, see Fixed-context questions.

Getting started

• Standard and Business — open-source Standard, free commercial use, planned proprietary Business, and public-access status.

• Why AX Code — what AX Code optimizes for, its intended audience, how it differs, and what it does not claim.

• Start Here — product mental model and the shortest paths by use case.

• Installation and Runtime Channels — supported platforms, packages, updates, and contributor launchers.

Runtime guides

• Wiki evidence visualization — open a local graph of recorded Wiki page/source relationships, or export offline HTML. Includes radial, force, and page-focus screenshots from AX Code’s own wiki.
• Conversation Recap — catch up on recent work with /recap and configure idle banners.
• Audio Notifications — system sounds and opt-in spoken alerts when a run needs your attention.
• Execution Evidence — graph, compare, replay, risk, rollback, branch, trace, and audit export.
• Verified Multi-Model Changes — the council and arena workflow end to end.
• Sandbox Mode — isolation modes, protected paths, network controls, and precedence.
• Cloud Operations Mode — plan/approve/apply workflow, approval tokens, and the cloudops agent posture.
• Autonomous Mode — unattended execution, approvals, headless use, and safeguards.
• Headless CLI — the one-shot ax-code run command for scripting, CI, and agent callers.
• TUI Languages and Setup — interface and conversation languages, first-time setup.
• TUI Animations — opening and ending pairs, terminal capabilities, and rendering fallbacks.
• Loop Mode and Scheduled Tasks — recurring prompts, durable schedules, and long-run limits.
• Long-Running Operations — supervised service examples, recovery semantics, and operational checks.
• Execution Modes — agent, hybrid, council, and arena behavior.
• Multi-Model Routing Best Practices — split premium reasoning and lower-cost support work without silently weakening correctness-sensitive tasks.
• Model recovery — exact model selection, explicit fallback configuration and diagnostics.
• Auto-Route — specialist routing and optional complexity routing.
• Model Effort — thinking levels and provider-specific behavior.
• Performance — coding tool profiles and local request-timing diagnostics.
• Memory usage — memory profiles, cache retention, and background output limits.
• Local Evidence Cache — optional SQLite and RocksDB evidence reuse, qualification, and rollback.
• Lifecycle Hooks — hook events and bundled policy packs.
• Run Report — workspace usage, activity, model/tool breakdowns, and per-session reports.

Providers

• Supported Providers and Models — provider IDs, credentials, and model discovery.
• Free-Tier API Quickstart — compatible no-cost paths, constraints, and safe evaluation.
• Custom and Gateway Providers — OpenAI- and Anthropic-compatible endpoints.
• MTPLX and oMLX Setup — source-checkout local presets, discovery, tools, and authentication.
• Tiel peer summary (20 September 2026) — current Tiel / Cyber-Tiel vs MTPLX native-API numbers; 194.88 completion default, 249.01 peak decode.
• AX Code/OpenCode Client Retest — six client/runtime combinations, active MTP, first/repeat timing, and code checks.
• Local Inference Measurements — dated AX Code/OpenCode and runtime testing with timing and comparability limits.
• AX Engine Model Selection — local model ranking and memory guidance.

Integrations

• MCP Integrations — trust, permissions, resources, and server security.
• ACP — the Agent Client Protocol happy path for IDE hosts.
• AX Wiki — source-backed repository knowledge and CI workflow.
• VS Code Integration — editor commands, settings, and workflows. Install from the marketplace listing.

SDK and service boundaries

• @defai-digital/ax-code-sdk — first-party TypeScript and JavaScript embedding.
• Native SDK Transport — gRPC-shaped desktop/native boundary and fallback behavior.
• HTTP and OpenAPI Compatibility — server mode and generated clients for other languages.
• OpenAPI snapshot — authoritative HTTP route and schema contract.

Architecture and reliability

• Semantic Layer — graph and LSP provenance, audit, and replay boundaries.
• Local Engine Architecture — why AX Code uses an AX Engine sidecar.
• Runtime Stability — cancellation, crash, stream, timeout, and TUI reliability contracts.

Reference

• Configuration JSON Schema — generated configuration input contract.

• TypeScript Native Support — source compiler, language service, diagnostics, and compatibility.

• Skill and Plugin Catalog — bundled skills, project skills, plugins, and evals.

• Isolation Policy Packs — machine-readable policy examples.

• Release Verification — canonical minisign public key and verification command.

• Windows Runtime Integrity — signatures, installed-file verification, and antivirus evidence.

• Security Policy — threat model, credential storage, and supported versions.

Documentation boundaries

docs/ contains public guidance for behavior that exists in a released runtime. Planning material, release targets, implementation policy, and temporary analysis do not belong here:

Every public Markdown page should declare its status, scope, last-reviewed date, and owner near the top. Prefer links to generated contracts or implementation sources over copied route lists and other high-drift snapshots.

Maintenance checklist

Before a release or a substantial documentation change:

1. Verify commands, defaults, flags, provider IDs, and runtime labels against their implementation.
2. Update the narrowest authoritative guide instead of repeating the same behavior in several front-door pages.
3. Run pnpm run test:scripts to catch broken local links, orphaned pages, and missing page metadata.
4. Update the approved export manifest and run node script/export-public-docs.mjs --website /path/to/ax-code.app, then run the website check, build and Workers smoke checks. Website builds use the committed snapshot and do not require source access. New public pages and assets must be explicitly approved in the manifest.
5. Keep proposals, project-development notes, release targets, and historical decision records under .internal/, not in the public navigation.