OpenRL · Monoid Labs

Agent RL is a data race. We are opening it to every lab.

Frontier agents improve through reinforcement learning on vast numbers of real, long-horizon sessions, collected by the few companies that operate them. Sovereign AI labs, universities and open model builders have the methods and, increasingly, the compute. What they lack is the experience to train on. OpenRL is an open, consented corpus of real agent sessions and runnable environments, so every lab can hill-climb on the same problems.

Contribute your sessions

The Monoid app finds sessions from the agents you use, redacts them on your machine, and asks you to approve each one.

Install the Monoid app for macOSApple silicon and Intel

Windows Linux

Apply for data access Read the data protection protocol

Fig. 1 · Agentic coding, held-out tasks

The gap is data, not method

Illustrative

Closed frontier labs
Open labs today
Open labs + OpenRL
Same RL recipe

Every lab climbs the same curve with the same RL recipe. Closed labs sit far to the right because their products collect the sessions. OpenRL pools consented sessions so open labs move up that curve, ten times more data per step.

Illustrative data behind Fig. 1
Lab	Real sessions	pass@1 (%)
Frontier lab A	50M	71.3
Frontier lab B	20M	66.5
Frontier lab C	7.9M	66.5
Sovereign lab	794	12.6
University group	2.0k	14.1
Open-weight builder	5.0k	16.8
Open labs + OpenRL (×10 data)	20k	26.7
Open labs + OpenRL (×100 data)	200k	44.5
Open labs + OpenRL (×1,000 data)	2.0M	59.8

Who it is for

Sovereign AI labs

National and regional programmes building their own models need agent training data they do not have to license from a competitor.

Universities

Academic groups can study RL for agents on real tasks, and publish results that other groups can reproduce.

Open model builders

Open-weight projects can close the gap on agentic work with trajectories and verifiers that are shared, versioned and audited.

Motivation

Agent capability is limited by data that few people can see.

Monoid Labs is a research lab working to democratize agent capabilities. We think the most direct way to do that is to make high-quality interaction data available to everyone who studies agents, with the consent of the people who produced it.

Public examples

Real work is scarce

Benchmarks and synthetic tasks cover a narrow slice of what developers ask of agents. Real sessions contain ambiguous requests, long tool chains, failed attempts, recoveries and corrections.

Who holds the data

Access is concentrated

The companies that operate popular agents see this data at scale. Universities, independent researchers and open model projects largely do not, and cannot reproduce results that depend on it.

One open corpus

Shared data compounds

A versioned corpus with verifiable environments makes results comparable across groups, and lets the wider community hill-climb on the same problems instead of rebuilding data in isolation.

Sources

Sessions from the agents developers already use.

The Monoid app reads the session logs these tools keep on disk. It needs no plugins or API keys and does not change how you work. We add tools as their log formats stabilize.

Supported

Claude Code

~/.claude/projects

Supported

Codex

~/.codex/sessions

Supported

Cursor

workspaceStorage/*.vscdb

Supported

OpenCode

~/.local/share/opencode

Planned

Gemini CLI

Google

Planned

Windsurf

Cognition

Planned

Cline

Planned

GitHub Copilot

GitHub

Contributing

How contributing works.

Contributing takes a few minutes, and nothing leaves your machine until the final step. The Monoid app guides you through each stage.

Install the Monoid app
Available for macOS, Windows and Linux. You can scan and review your own sessions without creating an account.
Guided setup
Select projects
The app lists the tools and projects it found. Only the projects you select are considered; everything else is excluded by default.
Opt-in per project
Redact and review
Secrets, personal information and identifying paths are removed locally. You read each redacted session and discard any you would rather keep.
Runs on your machine
Contribute
Submit a single batch, or have the app queue new sessions for weekly review. Any contribution can be withdrawn later.
Withdrawable

Install from the terminal

$curl -fsSL https://openrl.monoidlabs.ai/install.sh | sh

Monoid · OpenRL setupv0.1

✓Connect tools
✓Choose projects
3Redact locally
4Review
5Contribute

Redacting 403 sessions on this Mac

312 / 403

Nothing has left this device.

Claude Code

214 sessions

Codex

58 sessions

Cursor

131 sessions

OpenCode

not installed

Found	Count	Action
Secrets & tokens	37	removed
Emails & phone numbers	112	replaced
Names & usernames	41	replaced
Home paths & hosts	2,904	normalized

Next: read every session before it’s sentReview →

Data protection

We assume every session contains something sensitive.

Coding sessions routinely include credentials, client code and personal details. Redaction therefore happens on the contributor’s machine, before anything is shared, and is checked again at every later stage.

session 7f3a · claude code

userdeploy is failing again, can you check the env?

toolRead(<PROJECT_ROOT>/.env)

·STRIPE_KEY=<SECRET:stripe_key>

·DATABASE_URL=postgres://<USER>:<SECRET:password>@<IP>:5432/billing

assistantThe URL is missing sslmode. I'll add it and ping <EMAIL> with the change.

toolEdit(<PROJECT_ROOT>/db.ts) +1 −1

2 secrets

3 personal details

2 paths

Safeguards, in order

On the contributor's device: Detectors redact secrets, personal information and identifying paths before anything is shown for review.
By the contributor: Each redacted session is read and approved by the person who produced it.
On receipt: A second, stricter scan runs on our side. Anything it flags is quarantined and excluded.
Before release: Every dataset version is sampled and audited by people before publication.

Local redaction

Raw sessions never reach our servers, so there is no raw copy for us to lose.

Secrets and credentials

API keys, tokens, private keys, passwords and connection strings, found by pattern and entropy detectors.

Personal information

Names, emails, phone numbers, addresses, IPs, usernames and hostnames become typed placeholders.

Contributor control

Nothing is included until you select it, and you approve every session before it is sent.

Withdrawal

Withdrawn sessions are deleted from our store and excluded from every later release.

Documented rules

Each category we detect, and the placeholder it becomes, is written down and open to correction.

Data access

Access for researchers.

The corpus is released in versions to researchers at universities, non-profits and companies under a research license. Access is granted on application. Each release includes a data card describing its contents, how it was collected, and its known limitations.

Trajectories: Prompts, model turns, tool calls, observations and diffs, in order.
Environments: A repository snapshot, the task, and a verifier that can be rerun.
Outcomes: Tests passed, edits kept or reverted, and user corrections, usable as reward signals.
Provenance: Tool, model family and language. Never who contributed or where.

Apply for access Log inThe first release follows the pilot.

trajectory.jsondraft schema

{
  "id": "traj_01JC8Z…",
  "source": "claude-code",
  "env": {
    "repo": "<PROJECT_ROOT>",
    "snapshot": "sha256:9f2c…",
    "language": ["typescript"],
    "verifier": "npm test -- db.spec.ts"
  },
  "task": "Fix failing deploy: missing sslmode",
  "steps": [
    { "role": "user", "text": "deploy is failing again…" },
    { "role": "tool", "name": "Read", "args": { … } },
    { "role": "tool", "name": "Edit", "diff": "+1 −1" }
  ],
  "outcome": {
    "tests": "pass",
    "edits_kept": true,
    "user_correction": false
  }
}

Questions

Common questions.

What exactly is uploaded?+

Only the redacted sessions you approve: the conversation, tool calls, outputs and diffs, with secrets, personal information and identifying paths replaced by placeholders. Raw logs stay on your machine.

What if redaction misses something?+

No detector is perfect, which is why there are four safeguards. You review every session, we scan again on receipt and quarantine anything flagged, and each release is audited before publication. If you find a miss, withdraw the session and tell us so the rule can be fixed.

Can I contribute sessions from my job?+

Only if you have the right to share that work. We cannot know your employer's policy, so every project is excluded until you select it. If in doubt, contribute personal and open-source work only.

How do I withdraw a contribution?+

From the Monoid app or the web dashboard. Withdrawn sessions are deleted from our store and excluded from later releases. Copies downloaded from earlier releases cannot be recalled, which is why review happens first.

Who can use the data?+

Researchers whose application is approved, under a research license that forbids attempts to re-identify contributors.

Who runs OpenRL?+

Monoid Labs, a research lab working to democratize agent capabilities. We build the Monoid app and the redaction pipeline, and coordinate releases of the corpus.

Get involved

OpenRL depends on contributors, researchers and funders.

Contributors

If you use a supported agent, install the Monoid app and contribute sessions from projects you have the right to share.

Install the Monoid app

Researchers

Apply for access to the corpus, or tell us which environments and signals would make it more useful for your work.

Apply for data access

Funders and partners

We are seeking research grants, compute and institutional partners to support collection, auditing and open releases.

partners@monoidlabs.ai

Monoid Labs

A research lab.

Democratizing agent capabilities.

OpenRL is a project of

Monoid

monoidlabs.ai

hello@monoidlabs.ai

Agent RL is a data race. We are opening it to every lab.

Sovereign AI labs

Universities

Open model builders

Agent capability is limited by data that few people can see.

Real work is scarce

Access is concentrated

Shared data compounds

Sessions from the agents developers already use.

How contributing works.

Install the Monoid app

Select projects

Redact and review

Contribute

We assume every session contains something sensitive.

Local redaction

Secrets and credentials

Personal information

Contributor control

Withdrawal

Documented rules

Access for researchers.

Common questions.

OpenRL depends on contributors, researchers and funders.

Contributors

Researchers

Funders and partners