Q&A: How Headroom went from side project to enterprise infrastructure.

Interview by Tim Mitchell.

Every lengthy agent session leaves behind a token path no person requested for: restated file contents, duplicate instrument outputs, a complete log file learn finish to finish for the one line that mattered. Tejas Chopra discovered that path whereas burning by way of his personal API credit on side initiatives. 

Six months later, the compression layer he constructed to repair it has turn out to be one of many fastest-growing instruments in agentic AI.

Headroom picked up a point out in The Register in June 2026 and gained sufficient momentum to pull Chopra out of Netflix and into constructing Headroom Labs full time. The pitch is straightforward to state and onerous to construct: compress an agent’s context aggressively, however preserve the unique in reserve so the mannequin can at all times ask for it again.

Before the total dialog:

• The origin: constructed out of frustration together with his personal token payments whereas working Claude and Codex side initiatives on a private API key.
• The mechanism: reversible compression, that means the mannequin will get a compressed context plus a instrument name it might probably use to retrieve the unique if the compression left one thing out.
• The second: a Register writeup landed in the identical week GitHub Copilot’s pricing change pushed some customers from round $1,000 to $4,000 for a similar utilization, spiking token value nervousness proper when Headroom bought visibility.
• The scale: someplace between 200 and 300 billion tokens saved by Chopra’s personal depend, a quantity he stopped monitoring as soon as utilization took off, as a result of Headroom runs fully on the person’s machine by design.

We spoke with Tejas about how reversible compression truly holds up beneath load, the place it has failed in manufacturing, and what he’s constructing subsequent for a world the place brokers hand context to different brokers as a substitute of writing all of it down first.

Where Headroom got here from

Tim Mitchell: “For individuals who have not come throughout Headroom, what was the origin right here? Was there a selected second you thought “context wants a compression layer,” or did it construct up progressively?”

Tejas Chopra: “I run numerous side initiatives by myself API key, and I stored burning by way of my subscription tokens and falling again to pay-as-you-go. That’s after I began taking a look at the place the tokens had been truly going. I exploit Claude primarily, and Codex, and their dashboards will let you know that you simply spent X {dollars} in a session, however they do not let you know what drove it. Was the immediate unhealthy? Was the information I queried too giant? None of that was seen.

The second it actually broke was asking a GPU utilization query in opposition to a log file and working out of context window inside two questions. One line in that log file mattered. Why did the mannequin want to learn the entire thing? That’s the place it began. I constructed a means to intercept site visitors between Claude Code and the Claude fashions, which turned out to be a genuinely onerous proxy to construct, however as soon as I had it, I may see how a lot was flowing by way of that had nothing to do with what I’d truly requested.

That first model solely labored for API keys. Most individuals run on subscription-based Claude, which makes use of a special endpoint, so I had to construct separate paths for API keys, subscriptions, streaming, non-streaming, Vertex, and Bedrock. Even for one supplier, that is numerous combos, and everybody’s setup is a bit of completely different. Headroom began as a means to collapse all of that into one integration layer.

It started as my very own Claude drawback and expanded into Codex, then into different individuals who needed the identical compression method. The onerous half was proving accuracy, as a result of LLMs are non-deterministic. The core concept that separates Headroom from most of what is out there’s reversible compression: we compress the context however preserve the unique in reserve, and we inform the mannequin, when you want the supply, here is the instrument name to get it. In the overwhelming majority of instances the mannequin asks for that supply not often. But the choice is at all times there, and that is what lets us stand behind accuracy. Most compressors available on the market are lossy. They strip knowledge and offer you a means again to it solely whenever you rebuild it your self. We combine immediately into the Claude Code and Codex harness itself, and getting that integration proper takes extra engineering effort than the compression logic does.”

The traction spike

Tim Mitchell: “You began building this final December, so six or seven months in the past. The previous couple of weeks specifically have been an actual second of recognition. What modified?”

Tejas Chopra: “Honestly, basically little modified on our finish. Most of the core product was constructed 5 months in the past. I’ve spent the time since fixing bugs. It took off as a result of a number of issues landed in the identical week: The Register wrote about us, and inside three or 4 days, GitHub Copilot modified its pricing mannequin and folks went from paying round $1,000 to $4,000 for a similar utilization. Token value nervousness spiked at precisely the second we bought visibility, and that mixture is what drove it.

Stars and adoption are nonetheless rising, however we genuinely do not know by how a lot, as a result of we do not observe any utilization metrics. Headroom runs fully domestically, by design. Before this wave hit, we might measured someplace between 200 and 300 billion tokens saved. Since then, we’ve not checked. It could possibly be a trillion tokens saved, it could possibly be much less. We need individuals to belief that it is working on their very own machine, so we skip amassing knowledge by way of the open-source product fully.

The purpose we open-sourced it goes again to the mission: making AI sustainable and scalable for a future the place 100 million brokers may be working on knowledge concurrently. Getting there means growing the quantity of intelligence per token. Compression is one lever. We’re engaged on others.”

Tim Mitchell: “And the plan from right here? Staying open supply, or is there an organization forming round this?”

Tejas Chopra: “I’ve began Headroom Labs and left Netflix to construct this full time. We’re extending Headroom for enterprise use instances and are already in conversations with a number of enterprises. I’ve a CTO with me, additionally ex-Netflix, and a bunch of open-source contributors, a few of whom we’re planning to carry on full time.

We’ve additionally raised a small pre-seed spherical, which we’re saying this week.

I’m excited and scared in equal measure, which is an effective signal. This is why you decide a occupation like this: you need to get up wanting to do the work, and receives a commission for it. I’m chasing my very own model of IKIGAI, like everybody else.”

Reversible compression, and why the eval issues

Tim Mitchell: “The structure is fascinating, and reversible compression looks like the toughest a part of the promise to preserve. How do you assure an agent can at all times get again to the unique knowledge with out defeating the purpose of compressing it within the first place?”

Tejas Chopra: “That’s the best query, and it is why we’re working with a number of corporations on a impartial eval and benchmark technique for compression and token intelligence extra broadly. Models are aggressive about utilizing instrument calls whenever you give them one, so the mechanic is: compress the information, and depart a marker within the compressed textual content saying, if this is not sufficient, here is the instrument name to get the unique. In about 99% of instances, the mannequin solutions from the compressed context and skips the instrument name fully. When it does want the unique, that is one further hop and a few further tokens, however the expectation is that it amortizes throughout a session.

A concrete instance: Snowflake’s inference lead examined Headroom in opposition to Cortex Cocoa, their agent platform, for knowledge queries, and noticed 65 to 70% token compression, in some instances up to 90%, whereas getting again the identical knowledge. Data-intensive pipelines have a tendency to see the largest positive factors from Headroom, typically with out the mannequin ever needing to invoke reversibility. But we nonetheless suppose a correct eval and benchmark is the best means to show that out, which is what we’re constructing with a number of enterprises and corporations engaged on token and context intelligence.”

Where compression has truly damaged issues

Tim Mitchell: “What’s the worst failure mode you have seen, a case the place compression value correctness as a substitute of simply saving tokens?”

Tejas Chopra: “We’ve seen it with a number of the non-reversible methods we ship as defaults, like RTK and LeanCTX. They’ve dropped tokens the mannequin truly wanted, forcing it to retry the identical command with these methods turned off. So you have compressed the flip, but when it now takes two turns to do what one flip used to do, the web saving is shut to zero. That’s uncommon, nevertheless it’s the worst case we have hit.

The larger, extra fixed problem is that if Claude modifications its response construction for any purpose, we’ve to adapt our integration to match. That’s an integration tax we’re blissful to carry, as a result of it retains issues easy on the person’s finish, and not less than these failures fail quick and visibly.”

What’s hiding within the mannequin’s output

Tim Mitchell: “The output-side discount is a more moderen addition. What stunned you about what’s truly sitting in mannequin outputs?”

Tejas Chopra: “Output tokens are billed at roughly 5 occasions the speed of enter tokens, so a dial on the output side is price greater than the equal dial on the enter side. Right now, the one actual lever is the system immediate, telling the mannequin immediately to preserve it temporary. But builders differ loads: some need an evidence, some need it crisp, some skip studying the mannequin’s prose fully. That mismatch is a hidden value that no person’s pricing in.

What we do is study a developer’s type from their very own dialog historical past and use that to tune how the mannequin responds, which ought to unlock higher financial savings with out each engineer having to hand-tune their very own immediate.

This overlaps with a few different open-source instruments working the identical drawback, together with Caveman and a more moderen one referred to as Ponytail. Since Headroom, Caveman, and Ponytail are all open supply, the plan is to see which items are finest of their class and sew the very best of them collectively beneath one context-intelligence harness inside Headroom.”

The roadmap: from single brokers to tens of millions of them

Tim Mitchell: “Looking forward, what is the roadmap for Headroom, and what are you most enthusiastic about?”

Tejas Chopra: “Headroom in the present day solves the issue for brokers working in your laptop computer. The subsequent drawback is agent-to-agent context switch. Right now, if one agent wants to hand context to one other, the primary agent has to use an LLM name to write it out to a markdown file, and the second agent has to use one other LLM name to learn and interpret that file earlier than it might probably act.

With extra brokers within the loop, everybody finally ends up burning a mannequin name simply to perceive context they might have exchanged immediately. Markdown can be lossy: it is flat textual content, and it solely captures a fraction of what the primary agent truly understood.

We’re constructing a context alternate protocol to resolve that with a richer interface between brokers, and increasing it to enterprises working many brokers in opposition to their knowledge. That’s the cut up going ahead: Headroom, the open-source project, retains its personal roadmap and stays free for particular person builders, as a result of staying open supply is core to the sustainability mission. Headroom Labs is the broader guess, at present in stealth, constructing infrastructure for a future the place many brokers are working on and mutating knowledge concurrently, approached from an enterprise angle. Over time we count on to open supply extra from Headroom Labs too, together with work on observability and knowledge mutation.”

What this implies to your personal agent stack

Tejas’ numbers are self-reported and the eval work he mentions continues to be in progress, so deal with the compression percentages as a place to begin somewhat than a benchmark you possibly can cite blind.

What’s tougher to dispute is the form of the issue: token payments have gotten a line merchandise individuals truly scrutinize, and most groups nonetheless have restricted visibility into what’s driving theirs.

A couple of issues price taking from this dialog into your individual setup:

• Check your reversibility story earlier than you compress something. Lossy compression that drops a wanted token nonetheless prices you a retry, and a retry can erase the saving fully.
• Output tokens deserve as a lot consideration as enter tokens. At roughly 5 occasions the price, a verbose system immediate is a costlier behavior than a bloated context window.
• Agent-to-agent handoffs are the subsequent bottleneck. If your structure already routes context by way of markdown recordsdata between brokers, that is the sample Chopra is constructing another to.

Headroom stays free and open supply for particular person builders. Headroom Labs, nonetheless in stealth, is the place the enterprise model of this drawback will get solved subsequent, and Tejas’ roadmap suggests context alternate between brokers will matter as a lot as compression did on this first act. Find Tejas Chopra on LinkedIn.