Ethical Stack.

OpenClaw

Self-hostable agent gateway that connects chat channels, tools, sessions and memory to configurable model providers.

Score calculation

(4 + 6 + 9 + 1 + 7) ÷ 5 = 5.4

5 of six dimensions rated. Unrated dimensions are excluded from the average.

Full scoring method ↗
Setup details

Gateway and workspace run on the selected host. Tool execution, chat channel, model endpoint, sandboxing and shared-gateway trust boundary are operator-configured.

What fits your needs

Self-hosted gateway on a Mac, Linux host, VPS or other supported environment; inference and channel services may be remote.

Connected tools

Connects supported messaging channels and tools through a persistent local gateway and workspace.

See the evidence
Tradeoffs to expect

Trust boundary

The project treats each gateway as one trust domain; anyone who can operate a shared agent can use its granted tools.

See the evidence
How to switch

Select local or hosted inference

Configure a local endpoint such as Ollama or llama.cpp, or select a hosted provider; model/provider changes alter where prompts are processed.

See the evidence
Practical questions

Tools and approvals

Exec approvals are operator guardrails for accidental commands, not a tenant boundary; mixed-trust users should get separate gateways and OS users.

See the evidence

Local inference

OpenClaw supports local model servers, but local models do not include hosted providers’ safety filters; prompt traffic stays local only when model and embedding routes are local.

See the evidence

Memory and host approvals

Memory is stored as Markdown in the agent workspace and persists until removed; gateway/node host execution defaults to full permissions with ordinary approval prompts off, while sandbox execution defaults to deny. Operators can tighten policy and require approval; prompts are not a general approval gate for every external action.

See the evidence

Energy practices

4/10↑

OpenClaw exposes host/provider choice, model context/output limits and subagent concurrency settings; it publishes no route-specific electricity or energy measurement.

See the evidence

The docs describe where the gateway and model endpoint can run and provide controls over context, generated tokens and subagent concurrency. The reviewed evidence contains no task-level energy result or mapping from remote inference to electricity sources; controls indicate operator practice, not measured reductions.

1 + 1 reporting + 0 electricity + 2 controls = 4/10. This grades documented practices and transparency.

Gateway hardware use and remote provider model, region, facility and electricity mix remain deployment-dependent and unmeasured.

Reporting1/3
OpenClaw documents a gateway hosted by the operator and configurable local or remote inference endpoints, but no route-level environmental report.
Electricity sourcing0/3
Operator-controlled hosting and selectable inference endpoints do not identify the power mix of the host or any remote model provider.
Resource controls2/3
The model manifest exposes context and output-token ceilings, and subagent settings expose concurrency limits; these can bound workload size but are not measured physical savings.

Water disclosure

Water reporting has not been assessed for this choice.

Training data

—Depends on the model↑

The product can route tasks through multiple models or providers, so no single training dataset applies to this agent entry.

See the evidence

The product can route tasks through multiple models or providers, so no single training dataset applies to this agent entry. Model version, provider, and any fine-tuning or post-training sources must be checked for the selected route; a client privacy setting doesn’t confirm training-material permission.

Exact model, version, provider, inherited base, and fine-tuning training data vary by route or are not disclosed in the reviewed product sources.

Your data & privacy

6/10↑

OpenClaw runs its gateway, sessions, workspace and memory on the configured host; prompts still go to the selected model provider unless a local model is used.

See the evidence

Official docs describe OpenClaw as a local-first control plane with local workspace and session history, and support local inference endpoints. Local hosting does not itself keep prompts local when a cloud model is configured, and plugins/channels share the gateway trust boundary.

6/10 is a middle-band judgment: the operator can host gateway state and choose a local model, but remote model, channel, plugin and gateway exposure settings materially alter who processes or can access data.

Gateway exposure, backups, channel providers, external model endpoint, plugins and sandbox settings depend on the deployment.

Openness

9/10↑

OpenClaw’s official repository identifies the project as MIT-licensed open source.

See the evidence

The official OpenClaw repository publishes the code under MIT, enabling inspection, modification, and redistribution subject to the license. The selected model, model host, plugins, and deployment services keep their own terms and may be closed or hosted.

9/10 reflects the publicly available MIT-licensed gateway code; selected model weights, hosted APIs and connected services retain independent terms.

Releases and third-party extensions may add dependencies with separate licenses.

Worker protections

1/10↑

The open-source license doesn’t show working conditions for maintainers or contractors.

See the evidence

The reviewed official repository and security material document software licensing and security practices, but not pay, conditions, worker voice, contractor practices, or independent labor assurance. The gap is about available evidence, not a finding of mistreatment.

1 + 0 + 0 + 0 = 1/10. Documented worker protections; missing disclosure is an evidence gap.

Maintainer employment relationships, contractor labor, and any annotation or moderation supply chain are not established.

Pay & conditions0/3
Reviewed project docs do not publish workforce pay or conditions.
Worker voice & remedy0/3
No worker voice or remedy disclosures appear in the reviewed official project sources.
Evidence & follow-through0/3
No workforce-level independent assurance is disclosed in the reviewed sources.
Workers rubric ↗

Military ties

7/10Not established↑

No military relationship or product-specific restrictions are established by the reviewed project sources.

See the evidence

OpenClaw documents a self-hostable gateway and MIT-licensed software, but the reviewed sources do not identify defense contracts or a binding military-use restriction. Because users choose models and deployment environments, the project entry cannot establish every downstream use.

7 base + 0 safeguard adjustment = 7/10 (0 of 9 safeguard points; every 3 adds 1). No military relationship was established in the cited sources. This ranks above known defense ties without claiming a verified clean record.

Project-funding, government customer and downstream deployment records were not exhaustively established.

Use restrictions0/4
The reviewed project sources don’t confirm a binding military-use prohibition.
Enforcement & oversight0/3
No military-specific review, enforcement or remedy is disclosed in the reviewed sources.
Relationship disclosure0/2
No defense customer or relationship disclosure is identified in reviewed project materials.

Official project materials reviewed did not establish a military contract, customer, or deployment.

Military rubric ↗

BDS targets

Not listed↑

The OpenClaw project is not named in the two reviewed BNC target sources.

See the evidence

The official BNC guide and No Tech campaign name target companies; the OpenClaw project name does not appear in the reviewed lists. This bounded list check does not imply BDS approval or resolve relationships with providers, users, funders, or suppliers.

This is limited to the two official BNC campaign sources; it does not cover every local campaign, supplier, investment, or future change.

All sources

Side by side