gremlord

gremlord is an MIT-licensed, open-source CLI tool written in Go that wraps Claude Code in a local router, letting developers run the unmodified Claude Code harness against Anthropic, OpenAI, xAI, Ollama, vLLM, OpenRouter, or any OpenAI-compatible endpoint. It installs in one command, requires no daemon, and keeps Claude Code auto-updating on its own schedule. The project was formerly named "agentic" and migrated to the gremlord name; the latest release as of September 2026 is v0.1.6-gremlord.

What It Is

gremlord sits between Claude Code and model providers by intercepting ANTHROPIC_BASE_URL — an officially supported gateway hook — and acting as a local proxy. Anthropic traffic passes through byte-faithfully; all other providers (OpenAI dialect, xAI, Ollama, vLLM, OpenRouter) go through full request/stream translation. The result is the same Claude Code TUI, tools, and auto-update behavior, but with the model layer swapped out and every token metered, priced, and checked against configurable budgets.

How the Router Works

The architecture is deliberately minimal: a static Go binary with no persistent daemon. The first gremlord session binds the router port and serves all concurrent sessions; when it exits, another running session takes over within seconds. Model names are user-defined aliases in config.yaml, and anything starting with claude- passes through to Anthropic untouched.

Key routing mechanics include:

• Two-dialect support: Anthropic passthrough and full OpenAI-dialect translation for all other providers
• LLM-classifier tier routing: --model auto uses a cheap classifier (e.g., haiku) to assign each turn to a deep, standard, or light tier — sticky for the whole turn, no mid-flight model flips
• Task specialist overrides: A fixed label set (implementation, sql_data, debugging, code_review, architecture, security_review, critical_review) adds a second routing dimension without an extra classifier call
• Size-aware routing: Requests are checked against each candidate model's declared context window before routing; a request too large for all configured models is refused before reaching any provider
• Context scaling: Each model's real context_window (and optional effective_context) is declared so Claude Code's auto-compact fires at the right moment for 32K local models and 400K cloud models alike

Spend Tracking and Budget Gates

Every routed API token is logged to a local SQLite database with pricing applied at request time. gremlord cost --by model breaks spend down by model, profile, or session. Hard-stop budgets (daily, weekly, monthly) are configurable globally or per profile; when a cap is hit, the router refuses the next request with a message in the TUI — in-flight responses are never cut. A live statusline registered by gremlord setup shows session and daily spend in real time.

Subagents, Sessions, and CLI Delegation

• Subagents on any model: gremlord agents sync writes one subagent definition per configured model alias into ~/.claude/agents/, making any alias selectable by name (e.g., subagent_type: "gremlord-qwen") with full routing and budget enforcement
• Session peer discovery: gremlord peers matches Claude Code sessions by both auto-derived session name and project directory, surfacing ties rather than guessing
• CLI delegation: A cli provider type can hand whole tasks to a locally installed Codex or Grok CLI under the user's own subscription login; gremlord never reads or proxies the OAuth token

Update: v0.1.6-gremlord

The latest release (v0.1.6-gremlord, published September 10, 2026) fixes cost --receipt to honor the --since flag. The project was renamed from "agentic" to "gremlord" in this release cycle; gremlord setup and gremlord agents sync handle migration from ~/.agentic to ~/.gremlord, copying config and cost history while leaving the originals intact. Recent additions include the gremlord eval paired model evaluation feature (experimental), Auto Goal loop detection, deferred tool loading re-enablement, and estimator calibration for translated model token counts.

Tradeoffs to Know

Non-Anthropic models work through translation, but Claude Code's prompts and tool patterns are tuned for Claude — the project's own documentation notes they are "clunkier in the main loop" and better suited as cheap workhorses for background tasks and subagents. Specific gaps include no cache_control breakpoints on OpenAI-dialect backends, display-only thinking blocks, unavailability of Anthropic server tools (web search, code execution) on translated models, and token counting that deliberately overestimates by ~15% to prevent context overflow. The tool is explicitly positioned for a single developer, not a team gateway — for team deployments, the README points to LiteLLM as a more mature option.