Anymodel

AnyModel Development Skill

You are working on AnyModel — a universal AI coding proxy that lets you use any AI model through one interface. This skill contains everything you need: architecture, conventions, deployment rules, and debugging guidance.

Product Identity

• Name: AnyModel (capital A, capital M — never "anymodel" or "Any Model" in user-facing text)
• Package: anymodel on npm
• Website: https://anymodel.dev
• GitHub (proxy): https://github.com/anton-abyzov/anymodel
• Client bundle: cli.js in this repo/package is the shipped Claude Code-compatible TUI
• Author: Anton Abyzov (@aabyzov on X, @AntonAbyzov on YouTube)
• License: MIT

Architecture

AnyModel client (cli.js) --> anymodel proxy (:9090) --> OpenRouter / Ollama / OpenAI-compatible

Three components:

1. Proxy (proxy.mjs) — HTTP server intercepting /v1/messages, sanitizes requests, translates formats, routes to providers
2. CLI (cli.mjs) — entry point for starting proxy, connecting, managing presets
3. Bundled client (cli.js) — modified Claude Code client with violet diamond character and AnyModel branding

Providers

Presets (update version in KNOWLEDGE-BASE.md when changing)

When adding a new preset: add to MODEL_PRESETS in cli.mjs, update help text, update KNOWLEDGE-BASE.md, update site/index.html, update README.md.

Proxy Sanitization (proxy.mjs)

The sanitizeBody() function is the heart of AnyModel. It makes any model work with Claude Code's request format.

What it does (in order):

1. Strips Anthropic-only fields: betas, metadata, speed, raw output_config, context_management; preserves output_config.effort internally for compatible OpenAI forwarding
2. Preserves thinking — reasoning models (DeepSeek R1) need this for chain-of-thought
3. Preserves cache_control for OpenRouter (Anthropic models support it), strips for Ollama/OpenAI
4. Clamps max_tokens to minimum 16 — Claude Code sends max_tokens: 1 for probes, OpenAI/GPT rejects anything below 16
5. Fixes tool schemas:
   • Missing input_schema entirely -> adds a minimal object schema
   • Empty properties: {} -> keeps it empty and adds additionalProperties:false
   • Missing type field -> adds "type": "object"
   • Recursively fixes nested schemas (anyOf, oneOf, allOf, items)
   • Never injects or strips _unused; real _unused params must round-trip
6. Strips Anthropic-only tool fields: cache_control, defer_loading, eager_input_streaming, strict
7. Normalizes tool_choice: converts string format to object format
8. Local providers (Ollama, LMStudio, llama.cpp): capability-aware tool passthrough — since v1.12.0 the proxy no longer blanket-strips tools. Behavior is controlled by OLLAMA_TOOLS env var (auto default, on, off). In auto mode, the proxy tries tools and caches per-model capability; on "does not support tools" errors it retries without tools and remembers that model as no-tool-support. Tools are then compressed and budget-trimmed via providers/tool-compressor.mjs to fit local context windows (core tools Bash/Read/Write/Edit always kept). Translation: Anthropic tool definitions → OpenAI function format → Ollama message.tool_calls[] → translated back to Anthropic tool_use content blocks with stop_reason: "tool_use". Streaming supported.
9. tool_choice always stripped for Ollama — Ollama doesn't support tool_choice; proxy removes it unconditionally.
10. Auto-retry without tools — if a provider returns "No endpoints found that support tool use" / "does not support tools", proxy retries with tools removed AND caches model as no-tool-support.

Why this matters

Without sanitization, MCP tools break on non-Anthropic models. Claude Code sends 80+ tool definitions including MCP servers (Slack, Figma, Gmail, etc.) with every request. The proxy ensures these schemas are valid for the target model, so MCP works seamlessly through any provider — including local Ollama/LMStudio/llama.cpp models that support function calling (Qwen 3/Coder, Llama 3.1+, Mistral, DeepSeek R1, Gemma 4).

Regression guard: test/ollama-tool-passthrough.test.mjs asserts that in auto mode with an uncached model, tools are forwarded (not stripped). See also test/ollama-tools.test.mjs for the capability-cache logic.

Client Identity

• Character: Violet diamond-themed (diamond-head and diamond-feet ASCII art)
• Color: Light violet rgb(147,130,255), ANSI fallback magentaBright
• Branding: "AnyModel" everywhere (not "Claude Code"). Tips say "Ask AnyModel" not "Ask Claude"
• Version: Synced with npm package via prepublishOnly script
• ANYMODEL_MODEL: Env var displayed in client UI showing active model name

When modifying the client: never break the violet identity. The diamond character and "AnyModel" branding are what distinguish this from standard Claude Code.

Branding (reproducible)

cli.js is a 13MB minified bundle, so branding is NOT hand-edited anymore. It is applied declaratively and verifiably:

• scripts/brand-patch.mjs — idempotent applier. Asserts each patch's from appears exactly expect times before touching anything (a mismatch means upstream changed → fail loudly), applies replacements, then re-parses with node --check so a broken bundle can never ship. Some patches are adaptive: they swap a static literal for a JS expression that reads process.env.ANYMODEL_MODEL at render time, so the UI reflects the loaded backend model.
• scripts/brand-patches.json — the declarative manifest (data/code separation).
• scripts/gen-brand-manifest.py — regenerates the manifest. Run it against a PRISTINE bundle (it computes each expect from live anchor counts and fails if any anchor is missing).
• --check gate — node scripts/brand-patch.mjs --check is verify-only (CI anti-regression gate): non-zero exit on drift. Includes a version-tolerant vendor denylist sweep that catches residual user-visible vendor strings the exact from anchors miss (e.g. version-bumped model promos).

User-facing UI text reads "AnyModel" (capital), while the package/command name stays lowercase. Re-apply on every upstream cli.js bump — a bundle refresh wipes all branding.

Environment Variables

Client Discovery (findClient())

When npx anymodel connects, it finds the client in this order:

1. ANYMODEL_CLIENT env var (explicit path)
2. cli.js next to cli.mjs (bundled in npm package)
3. cli.js in current directory
4. Global claude binary (last resort fallback)

Universal Skill Loader (providers/skill-bridge.mjs)

SKILL.md is ONE shared open standard — Claude Code, OpenAI/Codex, Gemini/Antigravity, Cursor and Copilot all read the same <name>/SKILL.md format (YAML frontmatter + Markdown body). Only the discovery path differs. The bundled client only scans .claude/skills, so instead of patching its minified loader, AnyModel bridges at launch time: it discovers foreign-ecosystem skills, symlinks each into a per-session temp .claude/skills shadow, and passes that shadow via --add-dir. Zero format translation — the client's native SKILL.md reader handles everything.

• Roots scanned (under both project cwd AND $HOME, in precedence order): .agents/skills, .codex/skills, .gemini/skills, .agent/skills (Antigravity, singular). The project's own .claude/skills is read for collision precedence.
• ANYMODEL_SKILL_ROOTS (colon-separated) appends/overrides extra discovery roots; relative entries resolve against cwd (a bare relative root would otherwise produce a broken symlink).
• Collision rules: a project-local .claude/skills/<name> wins (foreign same-name is shadowed); among foreign roots the first occurrence of a name wins. Names are compared case-insensitively (macOS/Windows FS). Shadowed duplicates and unlinkable skills are logged, not silently dropped.
• Safety: symlinked skill entries that resolve outside their scanned root are skipped (no exposing ~/.ssh via an untrusted repo's .codex/skills/x link).
• A project-scoped local skill index for context economy lives in providers/skill-catalog.mjs (increment 0016).

Deployment Pipeline (MANDATORY)

Every change MUST follow this pipeline. Never skip steps, never publish without tests passing.

# 1. Run all tests
npm test

# 2. Commit and push
git add <changed-files>
git commit -m "description of change"
git push

# 3. Version bump and publish to npm
npm version patch
npm run sync-version    # syncs version to cli.js
npm publish

# 4. Deploy website (only if site/ changed)
vercel --prod

The prepublishOnly script in package.json handles version syncing. After npm version patch, run npm run sync-version to update the version in the bundled client, then npm publish.

Adding a New Provider

1. Create providers/newprovider.mjs following the pattern in existing providers
2. Export: { name, transformRequest, transformResponse, buildUrl, getHeaders }
3. Add detection logic in cli.mjs (PROVIDERS array)
4. Add env var documentation
5. Handle tool schema translation if the provider's format differs
6. Add tests in test/
7. Update KNOWLEDGE-BASE.md
8. Run the full deployment pipeline

Adding a New Preset

1. Add to MODEL_PRESETS in cli.mjs
2. Update the help text in showUsage()
3. Update KNOWLEDGE-BASE.md preset table
4. Update site/index.html preset section
5. Update README.md
6. Run the full deployment pipeline

Debugging Guide

Common Issues

Debugging Proxy Traffic

The proxy logs every request with color-coded output:

• [OPENROUTER] / [OLLAMA] / [OPENAI] — provider prefix
• tools=N — number of tools in request
• stream=true — streaming mode
• Yellow [OLLAMA] Passing N tools to <model> (mode=auto) — tools forwarded (capability-aware passthrough, v1.12.0+)
• Yellow [OLLAMA] Stripping N tools (mode=<x>, model=<m> cached as no-tool-support) — tool removal (either OLLAMA_TOOLS=off or model is in no-tool-support cache)
• Yellow [OLLAMA] Tool optimization: N tools (X tok) → compressed to Y tok — schema compression + budget trimming
• Green 200 — successful response
• Red status codes — errors with body excerpt

Competitive Context

Anthropic Third-Party Cutoff (April 4, 2026)

Claude subscriptions no longer cover third-party tools. AnyModel proxy mode is NOT affected (routes through OpenRouter, never touches Anthropic OAuth). This is a key selling point — AnyModel users don't need a Claude subscription.

URLs and Links

File Structure

anymodel/
  cli.mjs              # CLI entry point
  cli.js               # Bundled client (12MB modified Claude Code)
  proxy.mjs            # HTTP proxy server + sanitization
  package.json         # npm config
  providers/
    openrouter.mjs     # OpenRouter provider (passthrough)
    ollama.mjs         # Ollama provider (OpenAI translation + num_ctx + tool_use roundtrip)
    ollama-tools.mjs   # Tool capability mode + per-model no-tool-support cache
    tool-compressor.mjs # Schema compression + budget trimming for local models
    openai.mjs         # OpenAI provider (bidirectional translation)
    openai-local.mjs   # Factory for LMStudio/llama.cpp (thin wrappers over OpenAI)
    lmstudio.mjs       # LMStudio alias (:1234/v1)
    llamacpp.mjs       # llama.cpp alias (:8080/v1)
    prefix-cache.mjs   # Prefix-aware caching (increment 0004)
    skill-bridge.mjs   # Universal skill loader — foreign skills → temp .claude/skills shadow + --add-dir
    skill-catalog.mjs  # Project-scoped local skill index for context economy (increment 0016)
  scripts/
    brand-patch.mjs    # Idempotent, verifiable cli.js brand applier (--check CI gate)
    brand-patches.json # Declarative brand-patch manifest
    gen-brand-manifest.py # Regenerates the manifest from a pristine bundle
  site/
    index.html         # anymodel.dev homepage
    styles.css / script.js / sitemap.xml / robots.txt
  test/                # 93+ tests
  KNOWLEDGE-BASE.md    # Single source of truth (keep in sync with this skill)

Testing

Run all tests before any commit:

npm test

Tests cover: sanitization, tool schema fixing, provider translation, preset resolution, max_tokens clamping, streaming, error handling. When adding features, add corresponding tests. The test count should only go up.