Interface live
The terminal agent, the browser chat with the artifact beside it, and the Mac app.
39,141 lines, 29.7% of the code
- zorp-web
- web
- zorp-desktop
Everything in the stack, in one place
One open-source project, from the chat window down to training your own model. Planned pieces are marked as planned.
Use it
- Terminal agent
- Browser chat
- Mac app
Investigate
- Validate
- Investigate
- Co-write
- Deliver
- Critique gate
- Panel review
Know
- Local memory
- Web search
- Skills
- Voice input
- MCP tools
Measure
- Harness eval suite
- Public benchmarks
- BudgetBench
Train
- Pretraining
- Serve your checkpoint
- Fine-tuningplanned
- RLStudio post-trainingplanned
The open stack
One harness runs any open model. Everything else sits on top of it.
Pick a layer to see what it does, whether it ships today, and how much of the codebase it is.
Frontier, when you must
Your walls, offline
Investigate live
The research agent: tracks, the Kill Threshold, the evidence record, and the aryabhatta discovery layer.
17,266 lines, 13.1% of the code
- zorp-track
- erbga
Evaluate live
The harness suite that gates every change, compatibility runs and public benchmarks. BudgetBench lives in its own repository.
7,201 lines, 5.5% of the code
- zorp-eval
Memory live
Local recall and memory over your own conversations. Text goes to a loopback address or nowhere.
2,800 lines, 2.1% of the code
- zorp-recall
- zorp-loopback
Adapt and train partial
Pretraining in developer mode on Apple Metal through MLX today. Fine-tuning and RLStudio post-training are planned.
3,026 lines, 2.3% of the code
- zorp-train
Tools live
Web search through Tavily or your own SearXNG, Claude Code compatible skills, and local voice.
4,516 lines, 3.4% of the code
- zorp-search
- zorp-skill
- zorp-voice
Harness live
The core and the agent: any OpenAI-compatible model, tools, approvals, sessions and MCP, in one CLI and API.
57,875 lines, 43.9% of the code
- zorp
- zorp-agent
- zorp-mcp
- zorp-stub
The model live
Any open-weight model, or one you build. Swap it and nothing above it changes.
No code of ours. This is the part you choose.
Your hardware live
Your laptop, your workstation, your servers. Offline if you want it to be.
No code of ours. This is the part you choose.
Where the 131,825 lines go
The problem
The market is built to rent you intelligence, not to help you own it.
The default path is a frontier API. Easy to start, metered by the token, and every call sends your data out of the building.
The meter never stops
Every call, forever. The bill grows with your success.
Your data leaves
Each request ships your context to someone else's servers.
One vendor owns the dial
Price, limits and the model change without your say.
Own the whole stack. Rent only the thin slice on top.
Why now
The models got good, and the tooling grew up.
Small open-weight models clear the bar for most everyday work, and the tools around them became a stack you can own.
Evaluation
Tuning
Memory
Serving
Open-weight model
Good enough for most work
Classification, extraction, routing, drafting.
Open building blocks
Serving, memory, tuning and evaluation, ready to use.
Your schedule
Adopt a better open model when you choose to.
The strategy
Hybrid by design: local for the heavy lifting, frontier when you must.
The question is not open versus closed. It is which workload runs where.
Data sensitivity, higher at the top
Local
Sensitive, routine. Keep it in your walls.
Local, tuned
Sensitive and hard. Adapt a local model, never send it out.
Local
Routine, low stakes. The clear home for open-weight models.
Frontier, selectively
Hard, and safe to send out. Worth renting.
Task difficulty, harder to the right
Runs local. Keep it in your walls.Runs local, tuned. Adapt a local model, never send it out.Runs local. The clear home for open-weight models.The frontier, selectively. Worth renting.
One rule decides it
Keep the work local unless it is beyond a local model and the data is safe to send out. Sensitive data stays home however hard the task.
The open-source case
You are not buying a model. You are owning an interchangeable stack.
Swap in a better open model when one ships. Nothing above it changes.
Your app unchanged
The harness one interface
Any weights, swappable
Open model AOpen model BYour own model
No lock-in
The model is a component, not a vendor.
Open and inspectable
Open weights you can test and change.
A CLI and an API
The way developers and researchers already work.
Where it stands today: the harness, the research agent, local memory and pretraining in developer mode ship now. Fine-tuning and RLStudio post-training are planned. Zorp is pre-alpha.
The memory layer
Give it a memory that stays home.
A model with no memory starts cold every time. Zorp indexes your earlier conversations on your machine and quotes the relevant ones back into a new turn.
On your machine, loopback only
- Earlier conversationsyour own history
- Local indexSQLite, local embeddings
- Quoted into the turnverbatim, labelled
- The modelstarts warm
No remote embedding provider, and no fallback to one
Make it yours
Run it, adapt it, or train your own. On your hardware.
You are not stuck at any rung.
Run live
Use an open model
Run open weights as they are. Enough for most work.
Adapt planned
Fine-tune and post-train
Tune it on your data and shape behaviour with RLStudio.
Train live
Pretrain your own
A tokenizer and a model from scratch, in developer mode on Apple Metal through MLX.
Data stays in
The data that adapts or trains the model never leaves your machine to do it.
Prove it holds up
A leaderboard will not tell you if a small model clears your bar.
Test it under the token, context and memory budget you will actually run with.
Accuracy by memory strategy and context budget
Budget2K4K8K16K32K
Truncation0.290.20–0.400.000.270.18–0.360.000.270.18–0.360.000.330.23–0.430.000.330.24–0.430.00
Summary0.300.22–0.400.030.240.16–0.330.070.260.17–0.350.050.290.20–0.400.070.310.23–0.410.03
RAG0.300.20–0.400.000.310.23–0.420.000.350.25–0.450.000.340.24–0.430.000.310.22–0.420.00
The best strategy changes with the budget, and quality does not rise steadily with more context. The intervals overlap: this is a pilot, not a ranking.Only summarization broke its budget, on up to 7% of items. A leaderboard at one budget would miss it.
Budget-tiered evaluation is formalized in BudgetBench (Rao and Jaggi, arXiv 2609.13149).
Serve it fast
Local does not mean slow. Tuned inference is fast and cheap to run.
Zorp talks to any OpenAI-compatible server, so a quantized model on Ollama, LM Studio or vLLM becomes your endpoint. Then the harness has to make each call count.
An agent finishes only if every call does
21%
of attempts finish
Quantize
Fits commodity hardware and speeds up each call.
Batch and cache
Throughput climbs, cost per run falls.
No round trip
No hop to someone else's data center.
Agentic security
An agent you run offline cannot leak what it never sends.
Where an agent runs is the strongest control on it.
Bounded reach
It reaches only what your machine exposes.
Approvals
Tool calls pass approval rules you set.
A logged trail
Tool calls are stored with the conversation.
Nothing leaves
With a local model there is nowhere to send it. Web search is off until you turn it on.
For the answers that must hold up
Stop asking for an answer. Ask for an investigation.
For questions that have to hold up, Zorp's research agent keeps a record as it goes.
A prompt
Nothing kept
- Question
- Answer
Fluent in seconds, and silent on what pointed the other way.
An investigation
A record to answer to
- Question
- Registered intent
- Sources
- Evidence
- Checkpoint
- Answer
Every claim traces back to the record, checkable by someone who was not there.
Answers are cheap. Evidence is not.
Commit before you look
Say what would prove you wrong, before any evidence exists.
The hypothesis, the metric and a kill threshold are committed to git and hashed first. Only a person sets the threshold, and only a person can move it.
tracks/kafka-migration/prereg.toml
- hypothesis
- metric
- kill threshold
committed and hashed, before any evidence exists
- attempt 1within threshold, recorded
- attempt 2conflicts with attempt 1, recorded anyway
- attempt 3crosses the threshold
Measured, not claimed
Does the agent's confidence mean anything? We check.
Each forecast is recorded before its attempt runs, so it cannot be a postdiction.
Does a stated confidence mean anything?
Well calibrated below 0.96 and overconfident above it, across 151 forecasts recorded before each attempt ran. Both bands sit inside the 0.10 tolerance a person set from this curve.
What the answer is accountable to
Every claim resolves to a row, not a paragraph.
Each investigation writes to six tables. Measurements are typed values, so a claim in the report points at a stored row.
tracksOne investigation. Its question, its hypothesis, its status.
- preregistrationsThe committed hypothesis, metric, and kill threshold, with the hash that makes them tamper-evident.
- experimentsEvery attempt, including the ones that conflicted and the one that ended the run.
- metricsTyped key-value measurements, such as accuracy: 0.87.
- checkpointsEach human yes or no, and when it was given.
- validationsRedundancy and feasibility scores, each with the citation that justified them.
What you get back
Own it once, and stop paying to think.
Renting meters you forever. Owning costs the hardware once, then power.
Rent or own, on your numbers
Renting$900a month
Pays back in4.7 months
Saved over 24 months$16,640
Cost
An asset you run at the cost of power.
Sovereignty
Data, context and memory stay with you.
Control
You choose the model and when it changes.
Pace
Adopt open advances the day they land.
The one move
Move one heavy workload onto the open stack. Measure it.
Do not boil the ocean.
Pick one workload
High volume and routine. A summarizer or a classifier.
Run it on your stack
An open model on hardware you own, for that one workload.
Measure the gap
Cost per run, quality against your bar, data that left.
What you learn
How much of your frontier spend was never necessary, and what owning the rest would be worth.
The through-line
Own the model, own the stack, own the evidence.
The frontier has its place. For the rest, the future is open, local and yours to run.
- Any open model, or one you train
- Memory and evaluation, on your machine
- A research agent and an evidence record
- Offline-capable, on your infrastructure
- Made for developers, researchers and scientists
- The frontier, only where it earns its place
Adapted from the Own Your Intelligence briefing by Arjun Jaggi. Sources: Zorp, zorp.dev and github.com/aviskaar/zorp. BudgetBench, Rao and Jaggi, arXiv 2609.13149.