We're building an open-source platform to make AI agents cheaper and more reliable.
Here's why,
AI agents are everywhere. OpenClaw alone has created a massive wave of personal agents running locally on people's machines. Mac Minis with high RAM are on 2 to 6 weeks backorder because people are buying them to run agents. Inference spending jumped from $9.2B to $20.6B in a single year.
The problem
Agents are great at general tasks. Summarize this, draft that, answer this question. But the moment you ask them to do something specialized, things fall apart.
The success rate on specific real-world tasks is around 50%. On complex professional workflows like legal review or banking operations, it drops to 24%. When an agent fails, it doesn't fail quietly. It retries, pulls context from everywhere, burns through tokens trying to brute-force its way to an answer. The user either gets a mediocre result or gives up.
And here's the paradox: despite cost per token dropping 280x since 2022, total enterprise AI spending surged 320% in 2025. It's cheaper per token, but people consume so many more tokens that the bills are getting bigger, not smaller.
Most users discover this too late. They have no idea which agent costs what, which action triggered a spike, which model is eating their budget. They're running agents completely blind.
What we're building
Manifest is an open-source platform that makes AI agents more efficient and less expensive. We're startick to attack this from two angles.
First, cost observability. You connect your OpenClaw agent and you see in real time what each agent, each action, and each model is costing you. Per agent breakdown, per action breakdown, per model breakdown. Budget alerts before things spiral. No prompt collection, no content stored. Everything runs locally, fully self-hostable.
Second, automated model routing. Most agents default to the most expensive model for everything, even when a smaller, cheaper model would do the job just as well. We're building a system that automatically directs each task to the right model. Same result, fraction of the cost.
We want to bring the delegation cost as close to zero as possible.
Why open source
We believe the infrastructure layer around AI agents should be open. Cost observability shouldn't require sending your data to someone else's server. Model routing shouldn't be a black box. And agent owners should have full control over their telemetry, their data, and their spending.
Manifest is fully open source and self-hostable. We connect through OpenTelemetry with one ingest key per agent. Simple setup, full transparency, no lock-in.
Where we are now
We just shipped the cost observability tool this week. We already have our first users.
The model routing is next and we're building it right now.
We're currently at Skydeck, UC Berkeley's accelerator.
We think the future of AI agents isn't about making models bigger or smarter. It's about making the infrastructure around them more intelligent. Knowing which model to use for which task. Knowing how much something costs before it costs it. Knowing when to do the work yourself and when to hand it off.
That's what's missing. That's what we're building.
Try it
If you're running OpenClaw agents, give Manifest a try. It takes a few minutes to set up and we need your feedback to make it better.
And if you're not an OpenClaw user, give us some energy anyway. Star the repo, upvote on Product Hunt, or just share this with someone who runs agents. Every bit helps.
We're building an open-source platform to make AI agents cheaper and more reliable.
Here's why,
AI agents are everywhere. OpenClaw alone has created a massive wave of personal agents running locally on people's machines. Mac Minis with high RAM are on 2 to 6 weeks backorder because people are buying them to run agents. Inference spending jumped from $9.2B to $20.6B in a single year.
The problem Agents are great at general tasks. Summarize this, draft that, answer this question. But the moment you ask them to do something specialized, things fall apart.
The success rate on specific real-world tasks is around 50%. On complex professional workflows like legal review or banking operations, it drops to 24%. When an agent fails, it doesn't fail quietly. It retries, pulls context from everywhere, burns through tokens trying to brute-force its way to an answer. The user either gets a mediocre result or gives up.
And here's the paradox: despite cost per token dropping 280x since 2022, total enterprise AI spending surged 320% in 2025. It's cheaper per token, but people consume so many more tokens that the bills are getting bigger, not smaller.
Most users discover this too late. They have no idea which agent costs what, which action triggered a spike, which model is eating their budget. They're running agents completely blind.
What we're building Manifest is an open-source platform that makes AI agents more efficient and less expensive. We're startick to attack this from two angles.
First, cost observability. You connect your OpenClaw agent and you see in real time what each agent, each action, and each model is costing you. Per agent breakdown, per action breakdown, per model breakdown. Budget alerts before things spiral. No prompt collection, no content stored. Everything runs locally, fully self-hostable.
Second, automated model routing. Most agents default to the most expensive model for everything, even when a smaller, cheaper model would do the job just as well. We're building a system that automatically directs each task to the right model. Same result, fraction of the cost.
We want to bring the delegation cost as close to zero as possible.
Why open source We believe the infrastructure layer around AI agents should be open. Cost observability shouldn't require sending your data to someone else's server. Model routing shouldn't be a black box. And agent owners should have full control over their telemetry, their data, and their spending.
Manifest is fully open source and self-hostable. We connect through OpenTelemetry with one ingest key per agent. Simple setup, full transparency, no lock-in.
Where we are now We just shipped the cost observability tool this week. We already have our first users.
The model routing is next and we're building it right now.
We're currently at Skydeck, UC Berkeley's accelerator.
We think the future of AI agents isn't about making models bigger or smarter. It's about making the infrastructure around them more intelligent. Knowing which model to use for which task. Knowing how much something costs before it costs it. Knowing when to do the work yourself and when to hand it off.
That's what's missing. That's what we're building.
Try it If you're running OpenClaw agents, give Manifest a try. It takes a few minutes to set up and we need your feedback to make it better.
And if you're not an OpenClaw user, give us some energy anyway. Star the repo, upvote on Product Hunt, or just share this with someone who runs agents. Every bit helps.