Skip to content
OpenCorvus
EN中

Long-horizon agent harness · MIT · Self-hosted

Work that finishes,and gets better.

Long-horizon work carried by combined expert squads, evidenced at every handoff, and revised from your feedback.

Download · 0.0.64-beta

All platforms and formats

Windows

macOS

Linux

Served from the GitHub Release · 0.0.64-beta

View sourceRead the docsExpert Squads
$ git clone https://github.com/yangheng95/opencorvus
$ cd opencorvus && bun install
$ bun run --cwd packages/opencorvus build
$ bun packages/opencorvus/src/index.ts doctor
$ cd /path/to/your/repo
$ bun "$OPENCORVUS_SOURCE" serve

→ workbench at http://127.0.0.1:7878/ui/

Product story

Set the goal. Review the key results.

From user feedback to a version you can demo: see how goals, specialist handoffs and acceptance connect a complete workflow.

English · under 2 minutes · narrated with captions · illustrated workflow, not a recorded run

Measured result

Mission raises the same Luna to 34.00%

100 AutomationBench cases scored by strict pass criteria: the unassisted model first, then the result after full OpenCorvus execution.

AutomationBench · Strict pass rate

OpenCorvus Mission Base

openai/gpt-5.6-luna
Original GPT-5.6 Luna8.07%Without OpenCorvus Mission orchestration
OpenCorvus Mission Base34.00%Same Luna model · 100 cases
Evaluated cases
100
Current frozen sample
Absolute lift
+25.93 pp
percentage points
Versus original Luna
4.21×
strict-pass multiple

Different-sample context

Official held-out results

Not a cross-sample rank
  • Gemini 3.7 Flash High30.44%
  • Claude Opus 5 Max26.94%
  • GPT-5.6 Terra Max21.00%
  • GPT-5.6 Sol Max19.63%

Reference values come from the supplied official held-out comparison. They do not use the same sample as this 100-case frozen run, so they provide scale context only, not a model ranking.

Long-horizon

Where long work breaks

Three failures, and what answers each.

  • It stops short

    Requirements carry acceptance and non-goals; a lost process is recovered, not restarted; a terminal Task reopens on your next message.

    Terminal stateReopens

  • The result is not usable

    Typed artifacts with provenance, host observations separate from any agent's summary, and fact-check, integrity and visual review as named stages.

    Checked byNamed stages

  • It never gets better

    Tell a squad what you actually wanted; it drafts the revision, you accept it, and the receipt undoes it.

    Installs onYour acceptance

Composition

Squads, combined

The longest work is not one team working longer. It is several, each owning a stage.

Case

Turn DeBERTa into a complete research program

Model, data, CUDA experiments, a live product, figures, paper, review, and repository — one Mission, with a complete squad owning every stage.

6squads44named roles

Original Mission requirements

  1. Acquire DeBERTa v3 Base ABSA v1.1, re-investigate current sources, and search for or synthesize traceable training data before training.
  2. Provision a complete CUDA-only runtime with no CPU training path; record train/test performance for every innovative design and iteration; build an auto-updating training monitor and inference website; keep improving the model against an explicit baseline.
  3. Use the best experiment's exact model architecture and design rationale to produce reproducible, publication-ready figures.
  4. Research related literature and write a complete ACL-style short paper of at least four pages.
  5. Deeply review and proofread the manuscript, eliminate factual errors, and make its organization concise and informative.
  6. Create a well-organized Git repository, execute the work through Mission decomposition, and deliver a reviewed GitHub push.

Required final deliveries

  • Traceable ABSA dataset
  • CUDA-only training runtime
  • Best model checkpoint
  • Live monitor & inference site
  • Architecture & experiment figures
  • Reviewed ACL paper
  • Published Mission repository
View the audited CUDA experiment artifact on GitHub

Six high-level deliveries

  1. 01Model & dataDeep ResearchA verified DeBERTa v3 Base ABSA v1.1 source, current ABSA evidence, and a sourced plan to find, clean, or synthesize training data.
  2. 02CUDA training systemAdvancedA CUDA-only runtime, iterative baseline and candidate training, an experiment ledger, and a live training-monitor and inference website.
  3. 03Architecture evidenceData Analysis & Business InsightsBest-run comparisons, architecture and design diagrams, and reproducible figures bound to the exact winning checkpoint.
  4. 04ACL short paperResearch StudioA concise, informative ACL-style short paper of at least four pages, grounded in related work and the best experiment.
  5. 05Independent paper reviewAcademic Paper ReviewResolved findings across facts, citations, novelty, method, structure, figures, hallucination risk, concision, and informativeness.
  6. 06Mission repositoryBaseA reproducible, organized Git repository with the Mission stage map, reviewed documentation, and a verified GitHub push.
Unfold the complete execution mapCollapse the execution mapTurn six high-level deliveries into five workstreams, eighteen squad-owned stages, and visible milestones.

Workload ×3

This is the complete Mission OpenCorvus has to coordinate

It starts with model and data evidence, crosses CUDA training, a live product, and research publication, then converges on independent reproduction and GitHub release.

18squad-owned stages99named roles

Model & data evidence

Prove what can be trained, where the data came from, and how improvement will be judged.

3 stages15 named roles

  1. Verify model & literatureDeep ResearchSource-verified model identity, licence, checkpoints, and an evidence map of current ABSA work.
  2. Build the data lineData Engineering ReliabilityVersioned acquisition, cleaning, splitting, synthesis, lineage, and reproducible dataset builds.
  3. Freeze the baselineAI Model Governance and EvaluationA model-and-data card, fixed evaluation slices, quality risks, and a signed-off baseline protocol.
CUDA training & experiments

Build the GPU runtime, implement the trainer, run candidate campaigns, and reconcile every metric.

4 stages32 named roles

  1. Design CUDA runtimeCloud Platform ArchitecturePinned driver, CUDA, framework, container, storage, and GPU observability contracts with no CPU training path.
  2. Implement trainingAdvancedA reproducible trainer, configuration surface, checkpoint lifecycle, and focused verification.
  3. Run candidate campaignsEvolution LabFrozen arms, budgets, mutations, train/test metrics, integrity review, and a best-candidate decision.
  4. Reconcile performanceData Analysis & Business InsightsComparable metrics, slice analysis, uncertainty, failure clusters, and an independently checked ranking.
Live product

Turn experiment data into a real training monitor and inference experience, then prove it in-browser.

2 stages12 named roles

  1. Build live productFrontend InnovateA designed training monitor and inference experience backed by the experiment data contract.
  2. Prove live updatesBrowser Research & AcceptanceReal-page evidence that metrics refresh, inference works, responsive states hold, and console failures are absent.
Research & publication

Explain the winning design, render its evidence, position related work, write the paper, and audit it independently.

6 stages28 named roles

  1. Explain the designScientific Research DesignResearch questions, hypotheses, ablations, causal limits, and a defensible architecture narrative.
  2. Select the winning systemProduct ManagementOne decision register tying the chosen architecture to user value, evidence, trade-offs, and non-goals.
  3. Render publication figuresOffice DeliveryArchitecture, experiment, and comparison figures generated from the accepted source data with validation receipts.
  4. Map adjacent workPatent Landscape and Prior ArtA dated, query-reproducible adjacent-work landscape separating publications, claims, and open gaps.
  5. Draft the ACL paperResearch StudioA four-plus-page short paper whose method, results, figures, limitations, and citations share one evidence base.
  6. Audit the manuscriptAcademic Paper ReviewIndependent novelty, method, fact, citation, hallucination, organization, and concision findings with resolutions.
Reproduction & release

Reproduce the complete system, harden the publication boundary, and push one organized Mission repository.

3 stages12 named roles

  1. Reproduce & debugReview & DebugA clean-room reproduction of training, evaluation, inference, and site startup with root-cause fixes.
  2. Harden the repositoryCybersecurity AssuranceSecret, dependency, provenance, licence, workflow, and release-boundary assurance before publication.
  3. Publish the Mission repoBaseAn organized Git repository with exact setup, CUDA reproduction, data/model cards, dashboard, paper, evidence, and reviewed GitHub push.

Another long Mission

OpenCorvus writes the paper about OpenCorvus

A thirty-page, evidence-dense systems paper built from primary sources, executable evaluation, publication-grade figures, and an independent manuscript audit — not autobiography by assertion.

9squads55named roles

Original Mission requirements

  1. Reconstruct OpenCorvus from inspected source code, current architecture specifications, documentation, releases, and version history; do not write from memory or product claims.
  2. Research inspected related work, define falsifiable questions and bounded contributions, then design reproducible evaluation with baselines, metrics, uncertainty, failures, and ablations.
  3. Use Research Studio's analysis-report-quality Skill to create one validated evidence model before drafting a systems paper with at least 30 substantive main-text pages, excluding references and appendices.
  4. Do not pad the paper with repeated background, generic Agent prose, oversized figures, loose spacing, appendix migration, or unsupported claims.
  5. Redraw every necessary figure from accepted data with publication-grade hierarchy, labels, units, sample sizes, uncertainty, provenance, captions, and accessible encodings; raw notebook or chart-library output is forbidden.
  6. Independently review literature, novelty, logic, methods, statistics, facts, citations, hallucination risk, and presentation; resolve every critical and major finding and deliver the manuscript, source, bibliography, evidence ledger, figure sources, and reproduction guide together.

Required final deliveries

  • 30+ substantive pages
  • Source & architecture evidence
  • Reproducible evaluation
  • Publication-grade figures
  • Verified bibliography
  • Complete source & evidence bundle

Complete execution workflow

Research foundation

Freeze the questions, inspect primary sources, and reconstruct the system before making claims.

3 stages24 named roles

  1. Freeze the research charterScientific Research DesignFalsifiable questions, bounded contributions, competing explanations, evidence needs, ethics, and a no-padding acceptance contract.
  2. Build the source recordDeep ResearchInspected code, specifications, releases, history, papers, and official sources with claim-level locators and search limits.
  3. Reconstruct the systemAdvancedExecutable architecture evidence for control flow, persistence, Mission/Task semantics, squads, evolution, permissions, recovery, and observability.
Evaluation & reproduction

Run the declared baselines and ablations, then reproduce the central results independently.

2 stages11 named roles

  1. Run empirical evaluationData Analysis & Business InsightsReproducible baselines, cases, metrics, uncertainty, ablations, failure analysis, canonical tables, and checked quantitative claims.
  2. Reproduce independentlyReview & DebugClean-room reproduction of the claimed workflows and results, with root-cause fixes and unresolved limits recorded.
Evidence to manuscript

Position related work, refine every figure, and write thirty information-dense pages from one evidence model.

3 stages12 named roles

  1. Position the contributionPatent Landscape and Prior ArtA query-reproducible landscape that separates inspected prior work, overlap, defensible novelty, and unknowns.
  2. Refine every figureOffice DeliveryPublication-grade architecture, experiment, ablation, and comparison figures regenerated from accepted data — never raw plotting output.
  3. Write the thirty-page paperResearch StudioOne validated evidence model and a concise, information-dense 30+ page main paper whose claims, prose, tables, figures, limits, and references agree.
Independent manuscript audit

Challenge novelty, logic, methods, facts, citations, and presentation before accepting the final bundle.

1 stages8 named roles

  1. Audit and finish the manuscriptAcademic Paper ReviewResolved literature, novelty, logic, method, statistics, fact, citation, hallucination, and presentation findings, followed by a final figure-to-prose reconciliation.

Other combinations

How composition works

Evolution

Squads that revise

Two paths, both ending at a confirmation you have to give.

  • From what you said

    Say what you actually wanted. The host copies the exact revision, applies the edit, validates the package, and stages it.

  • From measurement

    Evolution Lab freezes the target revision, cases, scorers, budget and mutation surface, then runs the arms and reviews the result.

Nothing installs without your confirmation, and every change can be restored.

How evolution works

Research paper

OpenCorvus: Evolving Expert Organizations for Professional Work

A proof-of-concept working paper on reusable expert organizations, domain specialization, and experience-driven updates. Includes preliminary AutomationBench results; the research and experiments are still in progress.

PoC Version Under Construction

English PDF · 16 pages · 11 September 2026

Read the PoC paper (PDF)

Opens in a new tab

Why

Open, yours, controlled, legible

Four things decide whether a tool lasts.

  • Open source

    MIT licensed, every line published. Self-host it, audit it, fork it.

    MITLicence

  • Customizable

    Swap models, narrow tools, tune permissions, install squads — no forking.

    121Installable squads

  • In your control

    Runs on your machine. Tools are scoped per project; irreversible steps ask first.

    Your machineRuns on

  • Fully legible

    Every tool call, argument and result stays in the transcript, readable line by line.

    CompleteRun record

Compared with the nearest two

The closest products in positioning, side by side.

Compared with the nearest twoWorkBuddyDeepSeek HarnessOpenCorvus
LicenceCommercial, token packagesMITMIT
RunsCloud serviceLocallyYour machine or your server
Starting pointOne sentence to a finished outputPlugin kernel, compose it yourselfA whole harness working, then replace any layer
Capability unitExperts and Expert GroupsPluginsVersioned squads with a digest (121)
Getting inDesktop clientOne npx line to a web UIInstaller or source build

DeepSeek Harness is MIT licensed too, and records a run just as completely; its plugin kernel goes further than ours. The choice is whether you want to assemble a harness or start from one.

Compared against both products' published documentation

Expert Squads

121 installable squads

Browse all 121

Each one declares its scope, its roles, and its workflows.

Start

Get started

Start from the CLI, or install the desktop build.

Build from source

With Bun installed, clone the repository, build, and self-check.

git clone https://github.com/yangheng95/opencorvus.git

Already installed?

Start the server inside your repository; the workbench opens locally.

bun "$OPENCORVUS_SOURCE" serve
Read the quickstart

Three boundaries

  • Signature checkedPackages verify signature and SHA-256 before they land.
  • Permission boundedTools are granted per project; irreversible actions confirm.
  • Evidence keptEvery tool call and result stays in the conversation record.

FAQ

Common questions

Where this sits, what it does not do, and where your code goes.

How does this relate to Claude Code or Codex?

Different layer, and they work together. Those are coding sessions bound to one vendor's models; OpenCorvus is the harness — model-agnostic, multi-agent, self-hosted. The desktop app can even discover Claude Code, Codex, Gemini Code, Copilot, and GLM Code already installed on your machine and open one in the current project directory.

I already use one of those. Do I need this?

Depends what is missing. If you want a better single coding conversation, stay where you are. If you want coordination across many owned tasks, version-pinned expert squads, one permission and evidence trail, and work that survives a restart — that is what this does.

Does my code leave my machine?

The runtime runs on your machine or your own server, MIT licensed, every line auditable. Model requests go only to the provider you configure; pick a local runtime and nothing leaves the network at all.

Which models are supported?

One bundled catalog resolves 87 providers and 2,579 models, local runtimes included. Switching is configuration, not a fork.

What exactly is an Expert Squad?

An inspectable capability package: roles, workflow, Skills, tools, selection guidance, version, and digest frozen together. A Task pins one exact revision and cannot silently switch it mid-life.

Is it built on another agent?

The harness and the desktop app are both written in this repository, with no third-party agent engine underneath — that is what makes every layer replaceable. It stands on plenty of open source: Bun, the AI SDK, Solid, Tauri.

How long is long-horizon, really?

As long as your runtime stays online. Work survives a restart because leases, an event log, and a reconciler own it, not because a process stayed alive. A completed, failed, or cancelled Task reopens on your next message, at a fresh execution occurrence, with the old history intact.

Does it change itself behind my back?

No. A revision is drafted, validated as a runnable package, and staged. It installs only after you confirm it in your own message, and the receipt you get back is how you restore the previous revision.

How long can it run unattended?

Only while your runtime is online. This is not a hosted service and it does not promise unbounded autonomy — output depends on the model, the reachable sources, and the available evidence.

Contribute

Contribute

File an issue, publish a squad, or send a patch.

Go to GitHubJoin discussionsBuild a squad