# Run the agent loop on a local or self-hosted model

> Run the agent loop on a local or self-hosted model: 0 companies fully do it, 24 with a limit, 48 carry the tag. Every listing in Agent harnesses with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-harnesses.local-model/
- Markdown: https://www.anchorterminal.com/niches/agent-harnesses.local-model/index.md (~3,150 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-harnesses.local-model/index.min.md (~680 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-harnesses.local-model/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 24 | 48 | 8 | 5 |

## What counts as done

The harness completes a small multi-file edit with the local model, and no request leaves the machine.

Measured, for our own tests: every tool call is executed rather than printed as text.

How agents ask for it: use my llama-server with the agent; run the coding agent on Qwen locally; no cloud, everything stays on this machine; point it at LM Studio

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agent Zero (Agent Zero, s.r.o.)**: The setup has a local model server as a provider and the agent runs in its own container, but whether the container's own update or telemetry traffic leaves the machine was not established.
- **Aider (Aider AI LLC)**: Ollama, LM Studio and OpenAI-compatible APIs are documented and edits use text formats rather than tool calls, but the update check pings by default until --no-check-update.
- **AXLR (Underpass AI)**: local_models lists llama-server and vLLM entries that must return native tool_calls, and OPENROUTER_API_KEY becomes optional, but the engine updater and catalog fetches still reach api.github.com and the registry.
- **Clawboo (Sanreds)**: Since 0.6.0 Clawboo Native runs on a local Ollama with no key and marks models that cannot use tools, but Claude Code, Codex and Hermes teammates still call their own vendors.
- **Cline (Cline Bot Inc.)**: Ollama and LM Studio are listed providers and -P selects them, but telemetry is on by default and must be turned off with telemetryOptOut before the first task.
- **Crush (Charmbracelet, Inc.)**: provider add ollama --type ollama --base-url and model add register local models, and any OpenAI-compatible server works, but usage metrics are sent by default until CRUSH_DISABLE_METRICS=1 or DO_NOT_TRACK=1.
- **Pi (Earendil Inc.)**: models.json takes a compatible endpoint with a baseUrl and api type, but the pages read name no Ollama or llama.cpp preset, and the install ping and version check leave the machine unless PI_OFFLINE is set.
- **GitHub Copilot CLI (GitHub, Inc.)**: A BYOK provider registry file and local Ollama models are supported, but a Copilot sign-in may still be required and client telemetry is on until --disable-client-telemetry.
- **goose (Agentic AI Foundation (Linux Foundation))**: GOOSE_PROVIDER with GOOSE_PROVIDER__HOST points at a custom endpoint and Ollama is a provider, but Ollama unloading models closed MCP transports (acc-18) and usage data is sent until turned off.
- **Junie CLI (JetBrains s.r.o.)**: Custom model profiles take an Ollama localhost baseUrl or any OpenAI-compatible or Anthropic endpoint, but the headless page says a Junie token is required, and the anonymous statistics default is unstated.
- **Kilo Code CLI (Kilo Code Inc. (the terms are a contract with Anaconda, Inc. on behalf of itself and its affiliates, including Kilo Code Inc.))**: Any provider/model with your own key or a local model needs no Kilo account, but PostHog telemetry with the conda client ID is on until turned off.
- **Kimi Code CLI (Moonshot AI (the global Kimi terms name NOVASCENT PRIVATE LIMITED, Singapore))**: `openai`-type providers take any `base_url`, so llama-server or LM Studio work, but anonymous telemetry and the update preflight are on by default until `KIMI_DISABLE_TELEMETRY=1` and `KIMI_CODE_NO_AUTO_UPDATE=1` are set.
- **Letta (Letta)**: `--backend local` with Ollama, LM Studio and llama.cpp keeps agent state on the machine, but telemetry and the auto-updater stay on until `LETTA_CODE_TELEM=0` and `DISABLE_AUTOUPDATER=1`.
- **mini-swe-agent (No legal entity named. The MIT licence names Kilian A. Lieret and Carlos E. Jimenez)**: Any LiteLLM model works, but a local model needs `MSWEA_COST_TRACKING=ignore_errors` and the docs say that disables spend management.
- **MiniMax Code (Nanonoble Pte. Ltd.)**: `mcode provider add` takes any `openai-completions` or `anthropic-messages` base URL, but the pages read do not say whether telemetry or update checks stay on.
- **Mistral Vibe CLI (Mistral AI (RCS Paris 952 418 325))**: Any OpenAI-compatible endpoint works through the `llamacpp` preset or `api_base`, but telemetry and update checks are on by default and must be disabled in `config.toml`.
- **OpenAI Codex (OpenAI OpCo, LLC)**: `--oss` with `oss_provider = ollama|lmstudio` and custom `model_providers.<id>.base_url` run local models, but analytics and feedback collection are on until `analytics.enabled = false`.
- **OpenCode (Anomaly Innovations, Inc.)**: Ollama is probed at `127.0.0.1:11434` and any OpenAI-compatible `baseURL` works; acc-24 passed five tasks on LM Studio, but the default `notify` update check still leaves the machine.
- **OpenHands (All Hands AI)**: LM Studio, Ollama, vLLM and SGLang are documented with a recommended local model, but the page warns local models often fail tool calls and PostHog telemetry is on by default.
- **OpenMausBot (Supamaus Software Private Limited)**: Engines can point at any OpenAI-compatible endpoint per the README, but usage analytics are on by default.
- **Qwen Code (Qwen team, Alibaba Group (no legal entity is named in the repository or the docs))**: `model.baseUrl` takes an OpenAI-compatible server and acc-25 ran Qwen3.6 35B locally, but `privacy.usageStatisticsEnabled` defaults on and must be set false.
- **Mirror CLI (Reflection AI)**: `--api openai-chat-completions` with `--base-url` or `OPENROUTER_BASE_URL` reaches a local server, but analytics go to Statsig by default until turned off in `/settings` or `config.json`.
- **Unstoppable (Unstoppable Domains, Inc.)**: The support index says Ollama can run models on your own computer, but the page was not read and the app's accounts and credits still sign in.
- **Vibe-Trading (HKUDS)**: Ollama is supported, but market data pulls leave the machine by design and the CLI's update command calls PyPI.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Antigravity CLI (Google LLC)**: The models page lists a Local models section in its navigation, but the rendered page gave only the hosted model list and the section was not read.
- **DeepSeek Harness (dsh) (DeepSeek (the LICENSE names DeepSeek, and no legal entity is named in the repository or the docs))**: The pages read name only the native DeepSeek adapter; whether an OpenAI-compatible local endpoint can be configured was not found.
- **dsh-superagent (DeepSeek) (DeepSeek (the LICENSE names the DeepSeek Harness Team; DeepSeek Harness's terms name Hangzhou DeepSeek Artificial Intelligence Co., Ltd.))**: The model comes from the DSH profile; whether a local endpoint works was not established for DSH either.
- **Droid CLI (The San Francisco AI Factory, Inc.)**: The CLI reference points to a Custom Models (BYOK) page for other models, which was not read, so local endpoint support is unknown.
- **HarnessRouter (Lumentree Corporation)**: Model calls are at the provider's price or on your own keys; whether the self-hosted Community Edition reaches a local model was not read.
- **Paperclip (Paperclip Labs, Inc.)**: Custom model provider connectors are listed in the docs index but the page was not read; CLI adapters inherit the harness's model.
- **Prime Agent (Prime Intellect, Inc.)**: `models.json` takes provider overrides, but no local or OpenAI-compatible endpoint setup was found in the pages read.
- **Qoder CLI (BRIGHT ZENITH PRIVATE LIMITED)**: docs.qoder.com answered 502 from the fetch proxy on three attempts and a DNS failure through WebFetch, so no page was read.
- **Sculptor (Imbue, Inc.)**: The Pi harness lists providers from its own backend; whether a local endpoint is among them is not in the pages read.

- not this job: Amp (Amp Frontier Corporation): The settings reference has no local or OpenAI-compatible endpoint setting; models are Amp-served or the account's provider subscriptions, and every thread goes through Amp's servers.
- not this job: Auggie CLI (Augment Computing, Inc.): --model picks from Augment's hosted model list; no local or OpenAI-compatible endpoint option appears in the reference.
- not this job: Claude Code (Anthropic, PBC): The docs name Anthropic's API, Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and Anthropic-compatible gateways through ANTHROPIC_BASE_URL; no local OpenAI-compatible server or non-Claude model is supported.
- not this job: CLI Manager (solhun (Jeong Kyeonghun)): CLI Manager runs no model loop; the model belongs to the child harness.
- not this job: Conductor (Conductor): Conductor runs other harnesses; the model is theirs.
- not this job: Cursor CLI (Anysphere, Inc.): --model picks from Cursor's hosted list; no local endpoint option appears in the parameters.
- not this job: Devin (Cognition AI, Inc.): Hosted agent on Cognition's models and VMs.
- not this job: Emdash (General Action, Inc.): Emdash runs no model loop; the model belongs to the child agent.
- not this job: Gemini CLI (Google LLC): Gemini models through a Gemini API key, Vertex AI or Code Assist only.
- not this job: Harness (Autonomous, Inc.): Harness runs no model loop; local models belong to the agents it hosts.
- not this job: Harness² (Google LLC): Models are configured on Vertex AI or another cloud provider; the local demo uses deterministic editors with no model.
- not this job: Jules (Google LLC): Gemini models in Google's VMs only, and read-only since 11 November 2026.
- not this job: Kiro CLI (Amazon Web Services, Inc.): Models are Kiro's hosted list selected with `--model`; no custom endpoint or local provider setting is documented.
- not this job: Muse Code (Meta Platforms, Inc.): A Meta credential is required before the model can be called; no local or OpenAI-compatible endpoint is documented.
- not this job: OpenAI Agents API (OpenAI OpCo, LLC): Models are OpenAI's; the loop runs on OpenAI's service.
- not this job: Orca (Lovecast Inc.): Orca runs other CLI agents; the model choice belongs to them.
- not this job: Parallax (Conifer Solutions, Inc.): Reviewers are models from other labs picked by Conifer's router.
- not this job: Paseo (Mohamed Boudra): Paseo runs provider CLIs; model choice is theirs. It sends no telemetry of its own.
- not this job: Rovo Dev CLI (Atlassian Pty Ltd): Models are Atlassian-hosted with credit multipliers; no custom endpoint setting.
- not this job: T3 Code (T3 Tools Inc.): T3 Code drives provider CLIs; model choice is theirs.
- not this job: Traycer (Traycer AI): Traycer drives provider CLIs or its own inference.
- not this job: Verdent (Verdent AI, Inc.): BYOK and BYOA take provider keys and CLIs; no local endpoint is documented.
- not this job: Warp (Warp Agent CLI and the Oz Agent API) (Denver Technologies, Inc. (d/b/a Warp)): Bring your own API key routes through Warp's server-side harness; no local endpoint.

## What stops agents finishing

- chat template mismatch so tool calls are printed, not run (acc-23)
- model unloads between calls and the MCP transport closes (acc-18)
- 1 to 3 tokens a second on small hardware, and the harness falls back to a paid provider silently (acc-29)
- the harness's own telemetry or update ping still leaves the machine

## Around it

Before: model-serving.start-server. After: agent-harnesses.pair-in-terminal.

## Where this comes from

Weight 8 of 100 in Agent harnesses: 5 of 33 accounts (acc-18, acc-23, acc-24, acc-25, acc-29)

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-harnesses.local-model.json, or the MCP tool get_use_case with id agent-harnesses.local-model. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
