What is IronClaw?
IronClaw is an open-source AI agent from NEAR AI that runs background tasks such as inbox triage, morning briefings and KPI reports, with your API keys and passwords locked away from the language model. The GitHub README calls it a Rust reimplementation inspired by OpenClaw and an agent OS focused on privacy, security and extensibility. Credentials sit in an encrypted vault and are injected only at the host boundary for approved endpoints, so neither the model nor its tools see the raw values.
You can install IronClaw locally on Windows, macOS or Linux, or deploy it in one click on NEAR AI Cloud, where each instance boots inside a Trusted Execution Environment. NEAR AI released IronClaw 1.0 on 27 July 2026, adding a single checkpoint called the guard for every action, continuous checkpointing and shared memory across the CLI, web, Slack and Telegram. The project is active in October 2026, with 12.6k GitHub stars and dual Apache-2.0 or MIT licensing.
Key Features
- Encrypted credential vault: IronClaw stores secrets in an encrypted vault and injects them only for approved endpoints at the host boundary. The model never sees raw keys, which limits the damage from prompt injection or a malicious skill.
- WebAssembly tool sandbox: IronClaw runs each untrusted tool in its own WebAssembly container with capability-based permissions and resource limits. A compromised tool cannot reach files or endpoints it was not granted.
- TEE hosting: IronClaw hosted instances boot inside Trusted Execution Environments on NEAR AI Cloud. The official site says not even the cloud provider can see your data, which matters for regulated or sensitive workloads.
- Leak detection and allowlists: IronClaw scans outbound traffic in real time and blocks requests that appear to contain secrets. Network allowlisting also restricts tools to pre-approved endpoints, closing a common exfiltration route.
- Prompt injection defences: IronClaw applies pattern detection, content sanitisation and policy rules with block, warn, review and sanitise levels. Sensitive actions in version 1.0 need explicit approval before they run.
- Routines and heartbeats: IronClaw runs cron schedules, event triggers and webhook handlers in the background. Scheduled heartbeat runs push results to a chosen channel, such as a daily briefing in Slack.
- Persistent memory: IronClaw keeps memory across conversations using hybrid full-text and vector search. Version 1.0 also checkpoints continuously, so long tasks resume after an interruption instead of starting again.
- Self-built tools and MCP: IronClaw builds and sandboxes new tools on request for any service with an API, then reuses them later. It also connects to Model Context Protocol servers for extra capabilities.
- Multi-channel access: IronClaw works through a terminal REPL, web gateway, HTTP webhooks, Slack, Telegram and Signal. All channels share the same memory and safety rules in the 1.0 release.
Pros
- Security designed in: IronClaw combines a secret vault, Wasm sandboxing, leak detection and TEE hosting in one runtime. Teams worried about agents leaking API keys get several layers of protection without building them.
- Free and open source: IronClaw is dual-licensed under Apache-2.0 or MIT, and the code is on GitHub. You can audit it, self-host it on your own hardware and avoid subscription costs entirely.
- Free hosted tier: IronClaw Starter costs $0 a month in October 2026 and includes one TEE agent instance with $5 in credits. Non-technical users can try a hosted agent without setting up servers.
- Strong published benchmarks: IronClaw 1.0 scored 93.5 percent on PinchBench and 88.6 percent on ClawBench using deepseek-v4-flash, according to NEAR AI. The company also reports fewer OfficeQA errors than OpenClaw and Hermes.
- Ready-made automations: IronClaw ships recipes for inbox triage, meeting prep, uptime alerts, release tracking and invoice parsing. Integrations with Gmail, Google Drive, Sheets, GitHub and Discord cover common office work.
- Flexible deployment: IronClaw supports multi-tenant setups with shared tools and skills, or single-tenant deployments with full isolation. Organisations can match the deployment model to their own security policy.
Cons
- Full protection needs NEAR AI Cloud: IronClaw’s TEE isolation applies only to hosted instances on NEAR AI Cloud. Self-hosted users keep the Wasm sandbox and vault but lose the enclave guarantee, as Toolradar also notes.
- Technical setup for self-hosting: IronClaw local installs use scripts or a Rust build, and source builds need Rust 1.96 or later plus Node.js 22. Developers will cope, but non-technical users are better served by the hosted plans.
- Upgrade path not ready: IronClaw 1.0 cannot yet be applied as a direct upgrade to existing instances, according to the launch post. Early adopters must deploy new agents at agent.near.ai to get the new architecture.
- Pricing listings disagree: IronClaw’s official page shows Starter at $0 with $5 struck through, while Toolradar lists it at $5 a month. The promotional price may change, so check the plan before relying on free hosting.
- Benchmarks are self-reported: IronClaw benchmark scores come from NEAR AI’s own launch post and a single base model. Independent tests are scarce, so results on your workflows may differ.
- Credits cap heavy use: IronClaw hosted plans bundle fixed credits, from $5 on Starter to $200 on Pro+. Agents that run many routines or long tasks may exhaust credits and incur extra per-token inference costs.
Who is Using IronClaw?
- Security-conscious developers: Running agents that touch production APIs without exposing keys.
- Platform and DevOps teams: Offering company assistants with guardrails around infrastructure.
- Fintech and health teams: Testing agent workflows on sensitive or regulated data.
- Busy professionals: Automating briefings, inbox triage and meeting prep.
Pricing
IronClaw is free to self-host and offers hosted plans at $0, $20 and $200 a month (official site, checked October 2026).
| Plan | Price | What you get |
|---|---|---|
| Self-hosted | Free | Open-source code under Apache-2.0 or MIT; you pay for your own servers and models |
| Starter | $0/month (shown with $5 struck through) | 1 agent instance in a TEE, NEAR AI Inference and $5 in credits |
| Basic | $20/month | Up to 2 agent instances, usage pooling and $20 in credits |
| Pro+ | $200/month | Up to 5 agent instances, early access to advanced models, priority support and $200 in credits |
Toolradar lists Starter at $5 a month with no permanent free tier, which suggests the $0 price is a current promotion.
Token use beyond bundled credits is billed through NEAR AI Inference, so estimate how often your routines run before you pick a plan.
What Makes IronClaw Unique?
Secret isolation is the defining design choice: IronClaw keeps credentials outside the model entirely and injects them only at the host boundary, then layers Wasm sandboxes, outbound leak scanning and TEE hosting on top. Most personal agent frameworks, including OpenClaw, give the agent broader access and rely on the user to vet skills. IronClaw 1.0 also routes every action through one guard, so approval rules and logging apply in the same way across CLI, web, Slack and Telegram.
Compatibilities and Integrations
- Platforms: Windows MSI, PowerShell, shell script for macOS, Linux and WSL, or hosted on NEAR AI Cloud.
- Apps: Gmail, Google Drive, Sheets, Docs, Slides, Calendar, GitHub, Discord, Telegram and Signal.
- Protocols: MCP servers, HTTP webhooks and self-built Wasm tools for any API.
- Models and storage: NEAR AI Inference on hosted plans, OpenAI among configurable providers, and PostgreSQL persistence.
IronClaw Tutorials
Hosted users can have an agent running in minutes, while self-hosting takes a little longer.
- Sign up at ironclaw.com and deploy a Starter instance on NEAR AI Cloud, or install locally from GitHub.
- Add your credentials to the encrypted vault and approve the endpoints each tool may reach.
- Connect a channel such as Slack, Telegram or the web gateway.
- Enable a recipe like the morning briefing or inbox triage, and set its schedule.
- Review approval prompts and logs, then add MCP servers or new tools as needed.
How We Rated It
| Criterion | Score | Why |
|---|---|---|
| Accuracy and Reliability | 4.1 | Strong self-reported benchmarks |
| Ease of Use | 3.4 | Self-hosting needs technical skills |
| Functionality and Features | 4.4 | Routines, memory, tools, channels |
| Performance and Speed | 3.9 | Continuous checkpointing, few tests |
| Customization and Flexibility | 4.4 | Open source, MCP, own tools |
| Data Privacy and Security | 4.8 | Vault, Wasm, TEE, leak detection |
| Support and Resources | 3.5 | Docs, priority support costs |
| Cost-Efficiency | 4.3 | Free self-hosting and Starter |
| Integration Capabilities | 4.2 | Google apps, GitHub, chat apps |
| Overall Score | 4.1 | Average of the nine criteria |
Summary
IronClaw suits developers, security teams and power users who want an always-on AI agent without handing their secrets to the model. Its main limitation is that the strongest enclave protection depends on NEAR AI Cloud, while self-hosting requires technical setup. The hosted plans also cap usage through credits. Start on the free Starter plan with one low-risk routine, such as a morning briefing, before granting access to sensitive systems.

Reviews
Write a Review