Local AI Advisor

AI tools

Coding agents, IDE assistants and chat apps are clients: they connect to a local runtime through an API and shape how heavy your workload is.

Four layers between you and the hardware

Each hop has to be compatible. The advisor checks every one of them.
  1. 1AI toolThe app you use (Claude Code, Cline, Open WebUI…). It builds prompts and tool calls but never runs the model.
  2. 2Provider / APIThe protocol between tool and runtime: OpenAI-compatible, Anthropic-compatible or Ollama API.
  3. 3RuntimeThe inference server (Ollama, llama.cpp, LM Studio, MLX-LM, vLLM…) that loads and executes the model.
  4. 4ModelThe open weights at a chosen quantization and format (GGUF, MLX, safetensors).
  5. 5HardwareYour GPU / unified memory and its bandwidth, which set the speed limits for everything above.

Tool × runtime compatibility

The best connection path from each tool to each runtime, computed from the tool's documented connections and the APIs each runtime exposes. Hover a cell for the full path.
Connection level from each AI tool (rows) to each runtime (columns)
ToolOllamallama.cppMLX-LMLM StudiovLLMSGLangJanLocalAILemonade
Claude CodeCommunity-supportedCommunity-supportedRequires bridgeCommunity-supportedCommunity-supportedRequires bridgeRequires bridgeCommunity-supportedCommunity-supported
Codex CLIOfficialCommunity-supportedCommunity-supportedOfficialCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supported
OpenCodeOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
AiderOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
ContinueOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
ClineOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
Roo Code (discontinued)Community-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supported
Pi Coding AgentOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
VS Code (Copilot Chat, bring-your-own model)OfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
ZedOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
Open WebUIOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
LM Studio (built-in chat)UnsupportedUnsupportedUnsupportedOfficialUnsupportedUnsupportedUnsupportedUnsupportedUnsupported
API onlygenericOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
Custom appgenericOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficialOfficial
Other / not suregenericCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supportedCommunity-supported
Legend:OfficialCommunity-supportedRequires bridgeExperimentalUnsupported

“Requires bridge” means the tool only speaks the Anthropic API and the runtime only the OpenAI-compatible one, so a translating proxy such as LiteLLM sits in between. “Community-supported” paths work and are documented by the runtime or community, but not by the tool’s vendor.

The API layer

Every runtime exposes one or more of these protocols; every tool speaks at least one.

  • OpenAI-compatible API

    The de-facto standard (/v1/chat/completions, sometimes /v1/responses). Exposed by nearly every runtime: llama.cpp, Ollama, LM Studio, MLX-LM, vLLM, SGLang, Jan, LocalAI.

    Exposed by: Ollama, llama.cpp, MLX-LM, LM Studio, vLLM, SGLang, Jan, LocalAI, Lemonade

    OpenAI API reference(opens in a new tab)verified 2026-09-30high confidence

  • Anthropic-compatible API

    The Messages API (/v1/messages) used by Claude Code. Served locally by Ollama (v0.14+), llama.cpp's llama-server, LM Studio (v0.4.1+), Lemonade, vLLM and LocalAI — or via a bridge such as a LiteLLM proxy.

    Exposed by: Ollama, llama.cpp, LM Studio, vLLM, LocalAI, Lemonade

    Claude Code LLM gateway protocol(opens in a new tab)verified 2026-09-30high confidence

  • Ollama API

    Ollama's native API (/api/chat, /api/generate) with model management. Used natively by Open WebUI, Continue, Cline, Aider and VS Code.

    Exposed by: Ollama

    Ollama API(opens in a new tab)verified 2026-09-30high confidence

All tools

Prompt sizes and calls per task are approximate, editorial characterizations of typical sessions — not measurements.

  • Claude Code

    Anthropic

    Coding agentAgentic loop

    Terminal coding agent. Speaks the Anthropic Messages API; local models work through runtimes that expose an Anthropic-compatible /v1/messages endpoint (ANTHROPIC_BASE_URL).

    Base prompt
    ≈18K tokens
    Calls per task
    ≈25
    Min. context
    32K tokens
    Tool calling
    Heavy

    Connections

    • Anthropic-compatible APIany runtime exposing itCommunity-supported

      Local use is documented by Ollama, LM Studio, llama.cpp and Lemonade via ANTHROPIC_BASE_URL; Anthropic does not officially support local models. Some features (prompt caching, token counting) may be unavailable.

    Documentation(opens in a new tab)last verified 2026-09-30
    Claude Code LLM gateway docs; Ollama (v0.14+), LM Studio (v0.4.1+) and llama.cpp Anthropic-compatible endpoints(opens in a new tab)verified 2026-09-30high confidence

    Anthropic states it does not support routing Claude Code to non-Claude models.

  • Codex CLI

    OpenAI

    Coding agentAgentic loopBuilt-in local support

    OpenAI's terminal coding agent. Built-in --oss mode for Ollama and LM Studio; custom OpenAI-compatible providers via config.toml.

    Base prompt
    ≈9K tokens
    Calls per task
    ≈20
    Min. context
    32K tokens
    Tool calling
    Heavy

    Connections

    • OpenAI-compatible APIOllama, LM StudioOfficial

      codex --oss with oss_provider = "ollama" or "lmstudio".

    • OpenAI-compatible APIany runtime exposing itCommunity-supported

      Custom [model_providers] entry; Codex documents only wire_api = "responses", so the server must implement the Responses API.

    Documentation(opens in a new tab)last verified 2026-09-30
    Codex configuration reference(opens in a new tab)verified 2026-09-30high confidence

  • OpenCode

    SST / Anomaly

    Coding agentAgentic loopBuilt-in local support

    Open-source terminal coding agent with first-class support for any OpenAI-compatible provider, including Ollama, LM Studio and llama.cpp.

    Base prompt
    ≈10K tokens
    Calls per task
    ≈20
    Min. context
    32K tokens
    Tool calling
    Heavy

    Connections

    • OpenAI-compatible APIany runtime exposing itOfficial

      Configure a provider with @ai-sdk/openai-compatible and the local base URL.

    Documentation(opens in a new tab)last verified 2026-09-30
    OpenCode providers (Ollama, LM Studio, llama.cpp listed)(opens in a new tab)verified 2026-09-30high confidence

  • Aider

    Aider (open source)

    Coding agentConversationalBuilt-in local support

    Pair-programming in the terminal with git integration. Uses text edit formats rather than tool calls, so it tolerates models with weak tool calling.

    Base prompt
    ≈4K tokens
    Calls per task
    ≈3
    Min. context
    8K tokens
    Tool calling
    None

    Connections

    • Ollama APIany runtime exposing itOfficial
    • OpenAI-compatible APIany runtime exposing itOfficial

      openai/<model> with OPENAI_API_BASE pointing at the local server.

    Documentation(opens in a new tab)last verified 2026-09-30
    Aider — Ollama / OpenAI-compatible / LM Studio(opens in a new tab)verified 2026-09-30high confidence

  • Continue

    Continue

    IDE assistantInline assistBuilt-in local support

    VS Code / JetBrains assistant: chat, autocomplete, edit and agent mode. Native Ollama, LM Studio and llama.cpp providers.

    Base prompt
    ≈3K tokens
    Calls per task
    ≈4
    Min. context
    8K tokens
    Tool calling
    Light

    Connections

    • Ollama APIany runtime exposing itOfficial
    • OpenAI-compatible APIany runtime exposing itOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    Continue — Ollama / LM Studio / llama.cpp providers(opens in a new tab)verified 2026-09-30high confidence

  • Cline

    Cline

    IDE agentAgentic loopBuilt-in local support

    Autonomous coding agent inside VS Code. Large system prompt with tool definitions; supports Ollama, LM Studio and OpenAI-compatible providers.

    Base prompt
    ≈12K tokens
    Calls per task
    ≈18
    Min. context
    32K tokens
    Tool calling
    Heavy

    Connections

    • Ollama APIany runtime exposing itOfficial
    • OpenAI-compatible APIany runtime exposing itOfficial

      LM Studio and generic OpenAI-compatible providers.

    Documentation(opens in a new tab)last verified 2026-09-30
    Cline — running models locally(opens in a new tab)verified 2026-09-30high confidence

  • Roo Code (discontinued)

    Roo Code

    IDE agentAgentic loopBuilt-in local support

    VS Code agent with modes (architect, code, debug). NOTE: the official extension was shut down on May 15, 2026 — existing installs and the community fork (ZooCode) keep working with Ollama, LM Studio and OpenAI-compatible APIs.

    Base prompt
    ≈12K tokens
    Calls per task
    ≈18
    Min. context
    32K tokens
    Tool calling
    Heavy

    Connections

    • Ollama APIany runtime exposing itCommunity-supported

      Extension discontinued May 2026; maintained only by community forks.

    • OpenAI-compatible APIany runtime exposing itCommunity-supported

      Extension discontinued May 2026; maintained only by community forks.

    Documentation(opens in a new tab)last verified 2026-09-30
    Roo Code docs — shutdown notice (May 15, 2026)(opens in a new tab)verified 2026-09-30high confidence

  • Pi Coding Agent

    Earendil (pi)

    Coding agentAgentic loopBuilt-in local support

    Minimal terminal coding agent with a deliberately small system prompt and tool set. Local models via models.json (OpenAI-, Anthropic- or Google-compatible endpoints) with a dedicated llama.cpp guide.

    Base prompt
    ≈3K tokens
    Calls per task
    ≈15
    Min. context
    16K tokens
    Tool calling
    Heavy

    Connections

    • OpenAI-compatible APIany runtime exposing itOfficial

      models.json entry with api "openai-completions" and the local baseUrl.

    • Anthropic-compatible APIany runtime exposing itOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    pi coding agent — models.json (repo moved from badlogic/pi-mono)(opens in a new tab)verified 2026-09-30high confidence

  • VS Code (Copilot Chat, bring-your-own model)

    Microsoft / GitHub

    IDE agentAgentic loopBuilt-in local support

    VS Code's built-in chat and agent mode with a bring-your-own local model (Ollama or a custom endpoint). Works without a Copilot plan; inline completions still need a GitHub account.

    Base prompt
    ≈8K tokens
    Calls per task
    ≈12
    Min. context
    16K tokens
    Tool calling
    Heavy

    Connections

    • Ollama APIany runtime exposing itOfficial
    • OpenAI-compatible APIany runtime exposing itOfficial

      Custom endpoint speaking Chat Completions or Responses.

    • Anthropic-compatible APIany runtime exposing itOfficial

      Custom endpoint speaking the Messages API.

    Documentation(opens in a new tab)last verified 2026-09-30
    VS Code — bring your own language model(opens in a new tab)verified 2026-09-30high confidence

  • Zed

    Zed Industries

    IDE agentAgentic loopBuilt-in local support

    Fast native editor with an agent panel; official providers for Ollama, LM Studio, llama.cpp and OpenAI-/Anthropic-compatible servers.

    Base prompt
    ≈7K tokens
    Calls per task
    ≈12
    Min. context
    16K tokens
    Tool calling
    Heavy

    Connections

    • Ollama APIany runtime exposing itOfficial
    • OpenAI-compatible APIany runtime exposing itOfficial
    • Anthropic-compatible APIany runtime exposing itOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    Zed — use a local model(opens in a new tab)verified 2026-09-30high confidence

  • Open WebUI

    Open WebUI

    Chat UIConversationalBuilt-in local support

    Self-hosted ChatGPT-style web interface with RAG, connecting to Ollama or any OpenAI-compatible API.

    Base prompt
    ≈400 tokens
    Calls per task
    ≈1
    Min. context
    4K tokens
    Tool calling
    Light

    Connections

    • Ollama APIany runtime exposing itOfficial
    • OpenAI-compatible APIany runtime exposing itOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    Open WebUI — connecting Ollama / OpenAI-compatible providers(opens in a new tab)verified 2026-09-30high confidence

  • LM Studio (built-in chat)

    LM Studio

    Runtime appConversationalBuilt-in local support

    LM Studio's own chat window. Only works with LM Studio as the runtime.

    Base prompt
    ≈100 tokens
    Calls per task
    ≈1
    Min. context
    4K tokens
    Tool calling
    None

    Connections

    • OpenAI-compatible APILM StudioOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    LM Studio docs(opens in a new tab)verified 2026-09-30high confidence

  • API only

    —

    Direct APIServer / APIBuilt-in local support

    Call the runtime's API directly (curl, SDKs, scripts).

    Base prompt
    None
    Calls per task
    ≈1
    Min. context
    2K tokens
    Tool calling
    None

    Connections

    • OpenAI-compatible APIany runtime exposing itOfficial
    • Anthropic-compatible APIany runtime exposing itOfficial
    • Ollama APIany runtime exposing itOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    OpenAI Chat Completions reference(opens in a new tab)verified 2026-09-30high confidence

  • Custom app

    —

    AutomationBatchBuilt-in local support

    Your own application or automation (LangChain, LlamaIndex, n8n, scripts) using a local endpoint.

    Base prompt
    ≈800 tokens
    Calls per task
    ≈3
    Min. context
    4K tokens
    Tool calling
    Light

    Connections

    • OpenAI-compatible APIany runtime exposing itOfficial
    • Anthropic-compatible APIany runtime exposing itOfficial
    • Ollama APIany runtime exposing itOfficial
    Documentation(opens in a new tab)last verified 2026-09-30
    OpenAI-compatible API(opens in a new tab)verified 2026-09-30high confidence

  • Other / not sure

    —

    Chat UIConversationalBuilt-in local support

    Any other client. We assume it can use an OpenAI-compatible endpoint.

    Base prompt
    ≈1K tokens
    Calls per task
    ≈2
    Min. context
    4K tokens
    Tool calling
    Light

    Connections

    • OpenAI-compatible APIany runtime exposing itCommunity-supported
    • Ollama APIany runtime exposing itCommunity-supported
    Documentation(opens in a new tab)last verified 2026-09-30
    OpenAI-compatible API(opens in a new tab)verified 2026-09-30low confidence