an experiment on giving your @openclaw bot a real crm to operate on.
for a while now, while building plugins and skills, we kept running into the same practical limitation.
when we needed to organize leads, track conversations, assign follow-ups, move deals across stages, or simply understand where things stood, we were forced to rely on external tools that weren’t built for agents.
they solve the problem for humans, but they don’t map cleanly to how agents execute workflows.
so we built the crm layer we wanted to use ourselves.
a lightweight, open system that behaves like a traditional crm, but is designed so your @openclaw agent can operate it directly via api.
introducing OpenClaw CRM
→ generate a skill file tailored to your instance
→ plug it into your OpenClaw config
→ your agent can create and update contacts, manage deals, move pipeline stages, assign tasks, attach notes, and query structured records as part of its workflow
→ the solution also includes a built-in ai layer.
while connected to your OpenClaw bot, it can analyze activity, surface context, and act on structured data instead of just storing it.
under the hood:
→ 19 structured endpoint categories covering core crm logic
→ 40+ rest endpoints across contacts, deals, tasks, notes, stages
→ fully self hostable, so you control the data and infrastructure
we’re already running it inside our own processes, and we’ll keep refining it as those processes evolve.
Our x402 skills are now available on @openclaw.
We’re excited to share that 3 of our x402 skills are now available on @openclaw for everyone to download.
You can get them by sharing the links belwo with moltbot or by downloading from Clawdhub.
What skills are available?
→ X Search (x402)
clawdhub.com/TzannetosGianni…
AI powered X/Twitter search for real time trends, breaking news, sentiment analysis, and social media insights. Use it when you want to search Twitter/X for topics, hashtags, viral content, or public opinion.
Costs: $0.05 USDC per request via the x402 protocol on Base.
→ Find People (x402)
clawdhub.com/TzannetosGianni…
Open source intelligence (OSINT) tool for researching individuals, including professional backgrounds, career timelines, due diligence, competitive intelligence, and investor research. Use it when you need to research people, verify credentials, or gather professional information.
Costs: $0.15 USDC per request via the x402 protocol on Base.
→ Alpha Finder (x402)
clawdhub.com/TzannetosGianni…
Market oracle for mainly PM market intelligence, including Polymarket and Kalshi research, probability assessments, market sentiment analysi.
Use it when you want to research prediction markets, analyze betting odds, or spot market inefficiencies.
Costs: $0.03 USDC per request via the x402 protocol on Base.
An experiment on benchmarking and evaluating LLM outputs
x402 is the core of our experimenting around a lot of things like PolyAI, our plugins and more.
However, we spent the past week focusing on a different question:
How do you actually evaluate an LLM’s output?
In most teams, the workflow looks like this:
Create a prompt → run it on different LLMs → get an output → manually judge it → tweak the prompt → repeat.
This works, until you try to scale it.
A few problems show up quickly.
1/ Is the final output the only thing worth evaluating?
In agentic workflows, failures usually happen earlier, in reasoning, tool usage, state handling, or execution flow.
This gets even harder when agents need to run in isolated environments, like in cases such as PolyAI, where shared state or leaked context can invalidate results.
2/ How do you evaluate consistency across models?
If something breaks on Sonnet 4.5, it doesn’t mean it broke the same way on GPT-5.2. Manually inspecting outputs across models doesn’t scale and hides where things actually fail.
3/ When something works, do you know why?
Was it the prompt, the agent path, or luck? Can you reproduce it reliably?
So we tried a different approach.
We added tracing and basic evaluation into the workflow. The goal wasn’t to remove humans from the loop, but to move human judgment to a level where it can scale.
Instead of judging only the final output, we observed full runs, reasoning steps, tool calls, failures, and execution paths, across isolated environments, using monitoring and tracing via @langfuse, alongside our @opencode instances.
Patterns started to appear.
→ Agents generated content but never saved it.
→ They referenced tools that didn’t exist.
→ They returned plans instead of results.
→ Sometimes they stopped entirely.
The surprising part wasn’t the failures, but how repeatable they were.
With proper isolation and monitoring, we can now compare models running the same tasks and see where cheaper models behave almost identically to expensive ones. In practice, testing multiple models and combinations can reduce costs by 10x to 100x for specific tasks.
Some findings:
> Open-source models can outperform paid ones for narrow skills
> Agents doing the same job fail in different, repeatable ways
> Prompting gets better, instead of guessing, you can see where it breaks
At this point, it’s less about picking “the best model” and more about understanding where each model, agent, or prompt fits. Cost, reliability, and output quality become measurable.
There’s also a side effect: private benchmarking.
This kind of visibility creates internal insights most teams don’t have. We already know companies do this quietly, and that information has real value.
We’re still early. For now, we run agents, observe behavior, change one variable, and run again.
If you’re running autonomous workflows, we love to hear where your biggest pain points are and see if we can help refine the process.
(below an example we did on optimising a marketing workflow for content creation)
some skills / features people are asking for on X about @openclaw (screenshot_
using our plugin we also identified common problems / complaints people have on clawdbot. Focused on installation, instability and bugs.
pretty cool we found all that in less than < 1'
what skills are missing from @openclaw that you wish it had and would actually help you?
Prediction market related queries are spiking on our X Searcher.
From total snowfall and FED rate cuts to politics around Ali Khamenei, users are searching for high-value information that helps inform their predictions while saving meaningful research time.
After comparing results with similar tools, we’re confident about the value provided.
Concise, straight to the point, and an exact summary of the information people actually need.
All by leveraging the x402 protocol, settling payments per search.
Simple example below on "How many inches of snow in NYC this weekend"
1
53
1/
x402 Tools Plugin | Agentic tools for @OpenCode using x402 Protocol
We created an npm package that adds two specialized AI agents to OpenCode:
→ x_searcher (0.05 USDC per request): real-time X/Twitter search agent for trends, sentiment analysis, and social media insights
→ find_people (0.15 USDC per request): an OSINT agent for researching individuals, including professional backgrounds and career timelines, with source citations
each tool call triggers a micropayment on Base with no gas fees, so you only pay when you actually use the tools.
No subscriptions, no API key management.
You can check/download the package here: npmjs.com/package/@itzanneto…
1
1
37
Hello world.
we’ve been in stealth for a while, but it’s time to open the gates.
we are x402 Ventures, a group of builders leveraging the x402 payment protocol to create agentic solutions.
our aim is to experiment and build around x402, focusing on solutions users actually need, while openly providing resources for anyone who wants to use them in whatever they’re building.
the team is rooted in AI and has been building agents for the last couple of years.
our goal is to create value through experimentation and by building the next toolkit that solves real problems for users.
we’re looking for builders who want to implement tools, explore new ways to leverage x402, share feedback, and grow the collective by creating a space for discussion and productive feedback.
if you want to join us, we’ll drop a link to our community at the end of this thread.
below are some of the things / experiments we’ve built that are currently live:
1
2
177




