Agent harnesses on Ollama: OpenClaw / Hermes Agent / Goose / Pi

Parent: Ollama integrations and ecosystem · Published reference · snapshot 2026-10-14 · skill ollama-expert/references/ollama-integrations-ecosystem.md

Also known as: Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi

↓ Facts as markdown all context files

ollama launch hermes` installs the Hermes CLI agent if absent, lets the user choose a local or cloud model, configures the Ollama provider against the local OpenAI-compatible endpoint, optionally connects a messaging gateway, then starts Hermes chat. Later changes use `hermes model`; messaging uses `hermes gateway setup`. (from docs.ollama.com)

These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.

Definitions

On macOS and Windows, the Hermes Desktop installer is the recommended upstream installation path. It installs the desktop app and Hermes Agent together. If you prefer the command line, `ollama launch hermes-desktop` remains the explicit Ollama-managed path and uses the same Hermes configuration, sessions, skills, and memory as the CLI. [source]
Hermes Desktop is a native AI assistant app by Nous Research. It provides a desktop chat interface for Hermes Agent, an AI agent that can work with models, run tools, manage projects, use memory and skills, and connect to messaging gateways. [source]
OpenClaw is a personal AI assistant that runs on your own devices. It bridges messaging services (WhatsApp, Telegram, Slack, Discord, iMessage, and more) to AI coding agents through a centralized gateway. [source]
Hermes Agent is a self-improving AI agent built by Nous Research. It features automatic skill creation, cross-session memory, and 70+ skills that it ships with by default. [source]
Oh My Pi (OMP) is a terminal coding agent with IDE-style tools built in. It combines chat, project context, structured code edits, language server support, debugging tools, browser access, plugins, and subagents in one terminal workflow. [source]
Widespread misconfiguration of popular AI assistant OpenClaw means many instances are exposed to the public-facing internet, SecurityScorecard has warned. [source]
The report concludes that OpenClaw is not an isolated case but a leading indicator of a broader security challenge facing agentic AI. [source]

How it works

OpenClaw talks to Ollama's native API (/api/chat), not the OpenAI-compatible /v1 endpoint. Its docs say not to use the /v1 URL because it breaks tool calling and models may print raw tool-call JSON as text. Set `baseUrl` to `http://host:11434` (no /v1); `baseUrl` is the canonical key, `baseURL` is accepted. [source]
Added to the mix is that because OpenClaw agents are designed to act with legitimate authority, malicious activity can appear normal and, as a consequence, delay detection and increase potential impact. [source]

How-to and procedures

Goose has a built-in Ollama provider. For models without native tool calling, an experimental tool shim (enabled with `GOOSE_TOOLSHIM=true goose session`) uses an interpreter model, default `mistral-nemo` (`ollama pull mistral-nemo`), to turn the primary model's output into tool calls. The docs say behavior and configuration may change. [source]
Run `ollama launch` in a terminal and choose OpenClaw or Hermes Agent. [source]
Run `ollama launch hermes-desktop` again to switch models later. [source]

Examples and snippets

```bash theme={"system"} — `ollama launch hermes-desktop -- --force-build` … [source]
```shell theme={"system"} — `ollama launch pi --model qwen3.5:cloud` … [source]
```bash theme={"system"} — `ollama launch openclaw` … [source]
```bash theme={"system"} — `ollama launch openclaw --config` … [source]
```bash theme={"system"} — `ollama launch openclaw --model kimi-k2.5:cloud` … [source]
```bash theme={"system"} — `ollama launch openclaw --model kimi-k2.5:cloud --yes` … [source]
```bash theme={"system"} — `ollama launch pi` … [source]
```shell theme={"system"} — `ollama launch pi --config` … [source]
```bash theme={"system"} — `ollama launch hermes-desktop` … [source]
```bash theme={"system"} — `ollama launch hermes-desktop --model <model>` … [source]
```bash theme={"system"} — `hermes gateway setup` … [source]
``` — `Connect a messaging platform? (Telegram, Discord, etc.)` … [source]
```bash theme={"system"} — `curl -fsSL https://www.nvidia.com/nemoclaw.sh | \` … [source]
```bash theme={"system"} — `openclaw configure --section web` … [source]
```bash theme={"system"} — `openclaw configure --section channels` … [source]
```bash theme={"system"} — `openclaw gateway stop` … [source]
```bash theme={"system"} — `npm install -g @earendil-works/pi-coding-agent` … [source]
```bash theme={"system"} — `curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash` … [source]
```bash theme={"system"} — `nemoclaw my-assistant connect` … [source]

Measurements and reference values

The STRIKE researchers identified 28,663 unique IP addresses hosting exposed OpenClaw control panels across 76 countries using live internet-wide reconnaissance. Of those, 12,812 instances were flagged as vulnerable to remote code execution, with 63% of observed deployments classified as exploitable. [source]

Problems, failure modes and limitations

Hermes docs list gemma4:31b as the only model in their table with reliable tool calling (others chat-only); models without tool support cannot take actions. Long silence on first response is prefill of the large fixed prompt, documented as expected, not a hang. [source]
As well as RCE, OpenClaw instances are at risk of indirect prompt injection, whereby an attacker sends a target a message or places hidden text on a website containing malicious instructions. When the agent reads the message, it will faithfully follow these instructions – often without its owner being aware. [source]
HERMES_API_TIMEOUT=1800 # 30 minutes — generous for slow local models [source]

Comparisons and alternatives

Manual Hermes setup uses `hermes setup` -> Custom Endpoint with base URL `http://localhost:11434/v1`, or `model.provider: custom` and `model.base_url` in ~/.hermes/config.yaml. Using `--provider ollama` without `providers.ollama.base_url` fails with "provider 'ollama' has no endpoint configured" rather than falling back to OpenRouter. Hermes sets stream read timeout to 1800s for local endpoints; `HERMES_API_TIMEOUT=1800` goes in ~/.hermes/.env; `OLLAMA_KEEP_ALIVE=24h` avoids the 5-minute unload. [source]
Ai AI Coding Ai-Agents Alerting Anaconda Android API Architecture Aws AWS Amplify Backup Bash Cheatsheet Cloud Coding Community CookBook Cursor Data Database DeepLearning Dev Devops DGX Spark Digital Detox Docker Documentation Embeddings Filofax Flutter Git Gitea GitHub Go Golang Gpu Grafana Hardware Hermes Hosting Hugo Images Infrastructure Integration JavaScript K8S Knowledge-Management Kubernetes LabelStudio Latex Linux Llama.cpp Llamacpp Llm Logging Logseq Machine Learning Mainroad Markdown Melbourne Microservices Minio Mmdetection Monitoring Node.js NVidia ObjectDetection Observability Ob [source]
If you'd rather drive Hermes's own wizard instead of `ollama launch hermes`, install it directly: [source]
OpenClaw, formerly known as Clawdbot and Moltbot, is an agentic artificial intelligence framework designed to run continuously and act on behalf of users. The software allows AI agents to execute commands, interact with external services, integrate with messaging platforms and operate with broad system-level permissions. It has become increasingly popular among developers, enterprises and individual users experimenting with autonomous assistants capable of performing real-world tasks rather than simply generating responses. [source]

Facts and statements

As of the docs mirror, `ollama launch` supports Claude Code (`ollama launch claude`), Codex (`ollama launch codex`), OpenCode (`ollama launch opencode`) and Copilot CLI (`ollama launch copilot`); the list has grown past the four launch-day tools (Droid, DeepSeek Harness `dsh` and others). Check `ollama launch --help` for the live list. [source]
ollama launch hermes` installs the Hermes CLI agent if absent, lets the user choose a local or cloud model, configures the Ollama provider against the local OpenAI-compatible endpoint, optionally connects a messaging gateway, then starts Hermes chat. Later changes use `hermes model`; messaging uses `hermes gateway setup`. [source]
Manual Pi config: ~/.pi/agent/models.json gets an `ollama` provider with base URL http://localhost:11434/v1, api `openai-completions`, key `ollama`; ~/.pi/agent/settings.json sets the default provider and model. Ollama's page names the package `@earendil-works/pi-coding-agent`. [source]
OpenClaw's Ollama provider supports three setups: Cloud + Local (an Ollama host serving local models and, if signed in, :cloud models), Local only, and direct Ollama Cloud at https://ollama.com via the `ollama-cloud` provider with no daemon (model refs `ollama-cloud/<model>`). [source]
Ollama recommends a context window of at least 64k tokens for OpenClaw with local models; the Hermes docs state Hermes requires at least 64,000 tokens for agentic work with tools and that Ollama defaults to 2048. Hermes fix: a Modelfile with `PARAMETER num_ctx 64000` and `ollama create`, then point Hermes at the new model. [source]
OpenClaw has a serious exposure record: SecurityScorecard STRIKE counted ~40,000 internet-reachable instances in early Feb 2026, with reported RCE-exploitable hosts. Treat any OpenClaw on Ollama as needing a loopback-only gateway and sandboxing. [source]
<Note>Previously known as Clawdbot. `ollama launch clawdbot` still works as an alias.</Note> [source]
5. Hermes auto-detects downloaded models, confirm the one you want: [source]
<Card title="Hermes Agent" icon="https://mintcdn.com/ollama-9269c548/VdS6E00liu8o_Gkj/images/launch-icons/hermes-agent.svg?fit=max&auto=format&n=VdS6E00liu8o_Gkj&q=85&s=23dcf989a34d894a97bdb847c3831057" href="/integrations/hermes"> Personal assistant with memory, reusable skills, and messaging. </Card> </CardGroup> [source]
OpenClaw talks to Ollama's native API (/api/chat), not the OpenAI-compatible /v1 endpoint. Three setups are supported: [source]
NemoClaw is NVIDIA's open source security stack for [OpenClaw](/integrations/openclaw). It wraps OpenClaw with the NVIDIA OpenShell runtime to provide kernel-level sandboxing, network policy controls, and audit trails for AI agents. [source]
On the Hermes desktop app, Settings → Providers → Local Models installs and manages a local llama.cpp server for you — model downloads, memory fitting, and context sizing included. See Local Models. This guide is for manual setup: Ollama specifically, CLI-first workflows, or servers you want to run yourself. [source]
You'll set up Hermes Agent running entirely on your own hardware, using Ollama as the model backend. No API keys, no subscriptions, no data leaving your machine. Once configured, Hermes works exactly like it does with OpenRouter or Anthropic — terminal commands, file editing, web browsing, delegation — but the model runs locally. [source]
1. **Install** - If Hermes isn't installed, Ollama prompts to install the Hermes command-line agent. On first desktop launch, Hermes builds its packaged desktop app. 2. **Model** - Pick a model from the selector 3. **Configure** - Ollama configures Hermes Desktop to use your selected Ollama model 4. **Launch** - Ollama opens Hermes Desktop [source]
2. Find **Ollama**, click **Configure** 3. Confirm **API Host** is `http://localhost:11434` and click Submit [source]
<CardGroup> <Card title="OpenClaw" icon="https://mintcdn.com/ollama-9269c548/VdS6E00liu8o_Gkj/images/launch-icons/openclaw.svg?fit=max&auto=format&n=VdS6E00liu8o_Gkj&q=85&s=41b25b9a809d087e9e55b1b7bdb76cae" href="/integrations/openclaw"> Chat with an assistant through your messaging apps. </Card> [source]
<Note>OpenClaw requires a larger context window. It is recommended to use a context window of at least 64k tokens if using local models. See [Context length](/context-length) for more information.</Note> [source]
Cloud + Local A reachable Ollama host, serving local models and (if signed in) :cloud models [source]
Widen the API timeout — set HERMES_API_TIMEOUT=1800 in ~/.hermes/.env (see What You Need). [source]
Long context windows — cloud models offer 100K–1M tokens; local runtimes often default below Hermes' 64K minimum unless you configure them [source]
Chat and coding models Alibaba Model StudioAnthropicAnthropic Vertex plugin referenceArcee AIBasetenAmazon BedrockAmazon Bedrock MantleCerebrasChutesCohereDeepInfraDeepSeekFeatherless AIFireworksGitHub CopilotGMI CloudGoogle (Gemini)GroqHugging Face (inference)LongCatMetaMiniMaxMistralMoonshot AINovitaAINVIDIAOllama CloudOpenAI OpenAI OpenAI setupOpenAI authenticationOpenAI modelsOpenAI runtimes and Codex authOpenAI coverage and costOpenAI image generationOpenAI voice and speechAzure OpenAI endpointsOpenAI advanced configurationOpenCodeOpenCode GoPerplexityQianfanQwenRadiusStepFunSyntheticTeln [source]
Do not use the /v1 OpenAI-compatible URL (http://host:11434/v1). It breaks tool calling and models can emit raw tool-call JSON as plain text. Use the native URL: baseUrl: "http://host:11434" (no /v1). [source]
1. **Install** — If Hermes isn't installed, Ollama prompts to install the Hermes command-line agent 2. **Model** — Pick a model from the selector (local or cloud). See currently-supported models [here](https://ollama.com/search). 3. **Onboarding** — Ollama configures the Ollama provider, points Hermes at `http://127.0.0.1:11434/v1`, and sets your model as the primary 4. **Gateway** — Optionally connects a messaging platform (Telegram, Discord, Slack, WhatsApp, Signal, Email) and launches the Hermes chat [source]
<Note>Ollama support in NemoClaw is still experimental.</Note> [source]
1. In Goose, open **Settings** → **Configure Provider**. [source]
<div> <img alt="Goose settings Panel" /> </div> [source]
curl http://localhost:11434/v1/chat/completions \ [source]
Base URL: http://localhost:11434/v1 [source]
base_url: "http://localhost:11434/v1" [source]
Tens of thousands of OpenClaw systems exposed by misconfigurations and known exploits [source]
A new report out today from security rating firm SecurityScorecard Inc. warns that widespread vulnerabilities in OpenClaw deployments have left tens of thousands of internet-facing instances exposed to takeover through misconfigured access controls and known exploits. [source]
While OpenClaw may be rapidly growing in popularity, according to SecurityScorecard’s STRIKE Threat Intelligence team, that growing adoption has been accompanied by systemic security weaknesses. [source]
The report highlights that many OpenClaw deployments are exposed by unsafe defaults and poor deployment hygiene. Out of the box, OpenClaw binds its control interface to all network interfaces, making it accessible from the public internet unless explicitly restricted. [source]
A large proportion of exposed instances were also found to be running outdated versions of the software, despite patches being available for several high-severity vulnerabilities. Only a minority of exposed systems were running the latest release of OpenClaw. [source]
OpenClaw instances were also found to be heavily concentrated within major cloud and hosting providers, suggesting insecure deployment patterns are being reused at scale. During the research period, the number of identified internet-facing instances continued to grow, ultimately exceeding 40,000 exposed deployments. [source]
Infosecurity Magazine Home » News » Researchers Find 40,000+ Exposed OpenClaw Instances [source]
The exposed AI agents could enable threat actors to gain full access to potentially sensitive systems the OpenClaw instance is able to interact with. [source]
Vulnerability exploitation appears likely. SecurityScorecard has already discovered three high-severity CVEs in OpenClaw, with public exploit code available for each. [source]
Some OpenClaw users have also been leaking API keys linked to third-party services via their control panels, further amplifying the impact of instances’ internet exposure. [source]

Related concepts

OpenClaw — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
ollama — is a related of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
Hermes Agent — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
Hermes — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
ollama launch — is a synonym of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
Goose — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
agent harness — is a hypernym of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
openai compatibility — is a dependent of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
localhost — is a problem of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
coding agent — is a hypernym of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
Pi coding agent — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
tool calling — is a dependent of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
NemoClaw — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
ollama signin — is a part of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
num_ctx — is a problem of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
AI Assistant — is a related of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
OLLAMA_HOST — is a part of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
Codex — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
OpenCode — is a instance of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi
gateway daemon — is a part of Agent harnesses on Ollama: OpenClaw, Hermes Agent, Goose, Pi

Children

No children recorded.

← the whole tree · 3D view· how to read this page