AIPOCH Open-Science AI research workbench for reproducible science — open-source, local-first, and model-agnostic. AIPOCH Open-Science is an open-source AI research workbench. Start with a question, a paper, or a data file. Describe what you want to find out in your own words, and work with an AI agent to explore the evidence, analyze data, or create a chart. When you want to see how a result was produced, inspect…
The open-source AI research workbench for scientific research and agent workflows. Local-first, model-agnostic desktop app with extensible skills, MCP tools and connectors, Python/R execution and traceable artifacts for reproducible research on macOS, Windows and Linux.
AI research workbench for reproducible science — open-source, local-first, and model-agnostic.
AIPOCH Open-Science is an open-source AI research workbench. Start with a question, a paper, or a data file. Describe what you want to find out in your own words, and work with an AI agent to explore the evidence, analyze data, or create a chart.
When you want to see how a result was produced, inspect the available inputs, code, and execution history. To share the work, export a .science package with the conversation branches, file versions, Notebook records, and verification evidence you choose.
💡 AIPOCH Open-Science v1.2.0 updates
November 9, 2026
Remote compute: Run approved jobs on user-connected Kubernetes, Hugging Face Jobs, Modal, or Ray alongside SSH and Slurm, with per-job approval and collection recovery after restarts.
Built-in agent: The standalone Open-Science Agent is now the default for new setups; Claude Code, OpenCode, Codex, CodeBuddy, and Pi remain available.
Faster macOS updates: Download only changed blocks of the app image, with automatic full-download fallback.
Step discussions: Ask from imported .science research or a replay step to open a child conversation linked to that step; the import stays read-only.
Source previews: Hover over a source link to see its destination domain, with title and description metadata for approved HTTPS sources.
Workspace improvements: A simpler sidebar, a prominent “Fork to continue” action, and a stable reading position when sending from an earlier turn.
Reliability fixes: macOS update checks verify the running app, Windows names file blockers and keeps the input above the taskbar, and fixed-layout PDF translation retains successful pages when one fails.
The release assets include standalone CLI archives for macOS, Linux, and Windows. They bundle Node.js, so you can run the CLI without installing Node.js, npm, or Homebrew.
For the standalone CLI on macOS or Linux, use the separate AIPOCH tap:
brew install aipoch/tap/open-science
This Formula includes Node and runs without the desktop app; its release may differ from the desktop Cask. See the standalone runtime guide for prerequisites and upgrades.
For the standalone CLI and Web UI, install the published npm package with Node.js 22.13.0 or later:
Open the printed URL to complete setup, or follow the terminal setup guide. npm supports macOS ARM64/x64, glibc Linux ARM64/x64 and Windows x64; the desktop app is not required. Keep npm optional dependencies enabled and review the Linux prerequisites and upgrade instructions. The URL contains an authentication token; keep it private.
Windows reinstalls preserve research data. For a full cleanup, see the data reset tool, which permanently deletes local data after confirmation.
2. Complete first-time setup
Follow the setup wizard: Environment → Data location → Agent runtime → Model provider → Notebook runtime.
Complete the required environment and agent-runtime checks and test your model connection. Python/R Notebook setup is optional; Notebook and data-location settings can be changed later.
After setup, open System settings to change appearance or notifications. In General, you can find diagnostics, command-line tools, and the guide to manual installation and updates.
The former Model settings tab is now Local capabilities, covering PDF extraction, OCR language packs, and research narration. Narration requires a locally installed voice model.
Host compatibility, storage, and network checks
Provider, API Key, endpoint, and model validation
3. Start a research project
Click New project, open a session, and describe your research goal, inputs, and expected outputs.
Attach files, select a model and approval mode, then send the task. Use @ to reference project files or / to choose a skill.
Review tool activity and any approval requests, preview the results, and check their available evidence in Provenance.
Screenshots in this README illustrate the workflow. Labels, catalogs, and other interface details may differ from the version you install.
AIPOCH Open-Science .science Research Packages
Export a completed research session as a portable .science package for review, handoff, and archiving. Choose the conversation branches, file versions, Notebook records, and verification evidence to include, then import the package into another project or computer.
Export and import are available in the desktop app, local Web sessions, and the headless CLI. Remote Web sessions cannot read or write arbitrary local package paths.
Export → transfer → inspect
Export a completed session and choose the contents to include.
Move the .science file to a collaborator or another computer.
Import it into a project. The right-hand Research replay shows the recorded process; the left-hand conversation lets you ask about it and follow fixed step references.
Imported records remain read-only. Watching a replay does not execute code, call a model, or restore credentials. Your first question creates an associated writable discussion. Choose Fork to continue to create a writable copy for follow-up work. Re-executing a result still requires a complete recipe, required inputs, and an available runtime.
Questions from imported research or a replay step open step-linked child discussions without changing the import. Discussions survive archiving; deleting the parent turns them into ordinary conversations.
Side Chat conversations, private reading bookmarks, and credentials are excluded; included files depend on export selection. Received verification records describe the sender’s checks, not a rerun on this computer. See the .science package guide.
Benchmark Results
AIPOCH evaluated AIPOCH Open-Science on literature search and question answering, research workflows, biomedical data analysis, computational reproducibility, and hypothesis discovery. The results below come from AIPOCH’s own runs. Each score describes a particular task and evaluation setup.
Task
What AIPOCH Open-Science achieved
Find the target paper
For 68 of 75 test questions, the target paper appeared in the first 30 results on LitQA2-FullText-Search (90.67% Recall@30, gpt-6-astra).
Answer questions using the literature
In a separate run on LitQA2-FullText, Open-Science answered 70 of 75 test questions correctly (93.33% accuracy, deepseek-flash with Max reasoning).
Synthesize a cited answer
On 100 ScholarQA-CS2 validation questions, Open-Science earned a composite quality score of 0.9289/1 using deepseek-v4-flash. Official test references use different questions.
Carry out research workflows
In local ASI-Bench runs using the same model, Open-Science scored 59.46 on B4 and Codex scored 55.45. On B2 the scores were nearly identical (56.11 and 56.09).
Analyze biomedical data
On BiomniBench-DA Public 50, two judges scored the outputs 81.04 and 77.06. Their equal-weight mean was 79.05.
Reproduce published results
On CORE-Bench-Hard, Open-Science answered 36 of 37 test tasks correctly (97.30% task accuracy) in a local evaluation. The final tally includes retests.
Discover hypotheses from data
On a 25-task subset of the DiscoveryBench validation split, Open-Science earned a mean score of 0.7467/1 with a local model judge.
AstaBench — Literature tasks
Paper retrieval — target paper found for 68 of 75 questions
AIPOCH ran the official 75-question LitQA2-FullText-Search Test split locally. Its result matched the highest Recall@30 in the October 26, 2026 comparison snapshot. This local run was not an official leaderboard submission or placement.
Recall@30 counts a hit when the target paper appears in the first 30 results; it does not measure answer accuracy. Costs were not standardized across the runs.
Full-text question answering — 70 of 75 answers correct
AIPOCH ran the 75-question LitQA2-FullText Test split in AstaBench v0.5.4. The chart orders the local result and historical reference records by score. Its “#1” refers to that comparison.
Accuracy measures correct answers. Here, 93.33% is the local answer accuracy and 90.67% is a historical reference answer accuracy. The retrieval chart above uses a different metric. Reference records were captured on October 28, 2026; the highest reference score was checked again on November 2.
Cited-answer quality — local validation result and official test references
The local ScholarQA-CS2 result measures the quality of cited answers on 100 Validation questions. Reference results are from the Test split, so the scores do not establish a head-to-head ranking.
The composite score is reported on a 0–1 scale; it is not the percentage of questions answered completely correctly.
AstaBench — Reproducibility and hypothesis discovery
CORE-Bench-Hard — 36 of 37 tasks correct, including retests
Given a study’s code, data, and README, the agent must reproduce the requested results. AIPOCH ran DeepSeek Flash with Max reasoning and used the official programmatic scorer locally. A task earns one point only if all required answers are correct. All 37 tasks received a score. The November 6 tally includes retests and is not a first-pass success rate. Evaluation report.
DiscoveryBench — mean score of 0.7467/1 on 25 validation tasks
AIPOCH used deepseek-v4-flash both to generate hypotheses and to judge them with a local implementation of the DiscoveryBench scoring procedure. Of the 25 tasks scored, 18 received full credit, 2 received partial credit, and 5 received zero. The November 9 run used a validation subset and a different judge from the official full-test evaluation, so the scores cannot be ranked together.
ASI-Bench — Following a plan and choosing methods
The four conditions ask the agent to execute a specified plan (B1), organize a workflow (B2), choose methods (B3), or work through distracting information (B4). The first two rows of the chart compare local runs using the same GPT-6 Astra model.
Open-Science scored 79.35, 56.11, 59.67, and 59.46 on B1–B4; the local Codex reference scored 74.78, 56.09, 58.12, and 55.45. The B2 scores were nearly identical. Other published runs differ in task instances, configurations, or judging.
The local B1 score covers 59 tasks after one scorer-related exclusion. Three local repeats were reported for B1/B2. B3/B4 used Open-Science v0.32.0 and a substitute judge; each score averages three 60-task runs.
BiomniBench-DA — Biomedical data analysis
On the Public 50 task set, the Gemini 3.1 Pro judge scored the outputs 81.04 and the DeepSeek v4-pro judge scored them 77.06. AIPOCH reports their equal-weight mean, 79.05, for the gpt-5.6-sol (xhigh) configuration.
Judge scores and historical references — judge coverage varies
The chart combines entries with different judge coverage. Its “#1” describes the ordering of that collection. A ranking under a common judging procedure would require comparable scoring for all entries.
79.05 is a mean judge score across this internal evaluation.
Evaluation scope: These reports cover local/internal runs from August to November 2026 with the stated models and splits; the evaluations were not rerun for the current release. Reference scores are historical snapshots, not live rankings. Scores from different benchmark families are not directly comparable, and numerical gaps do not establish statistical significance.
Import references from Zotero or paste up to 100 identifiers or article links at a time. With a screening model configured, screen papers against your inclusion and exclusion criteria. You can review the model’s decisions alongside your own and open the PDFs to check a finding in context.
Preview scientific data, documents, images, and molecular structures alongside your papers. Run Python or R in a notebook to turn a project dataset into a table or figure, then inspect the output alongside the files you started with.
Check how a result was produced and what supports the answer.
Open a specific version of a figure, table, or report in Provenance to see the available inputs, code, and execution history. When the inputs, recipe, and runtime are available, rerun that version in the desktop app and compare the outputs.
Turn on Reviewer for a separate check of an agent’s answer against the execution logs and files available for that turn. Inspect the issues it flags and the follow-up review after the agent makes a correction.
Give a colleague the context behind a result.
Choose the conversation branches, file versions, Notebook records, and verification evidence to include in a .science package. A colleague can open it on another computer and inspect the steps behind a figure or table.
The imported record opens read-only. Importing it runs no code and restores no credentials. The verification records show what was checked on the sender’s computer. To explore a follow-up question, your colleague can select Fork to continue and work in a writable copy.
Choose the tools and computing resources for your analysis.
Choose a model provider and use 48 built-in connectors to access scientific data sources. Some connectors require you to enable them first.
Start with 24 built-in skills for recurring tasks. The Skills Marketplace lists 525 skills you can install. To delegate part of a task, choose from 10 specialists in the Specialist Marketplace.
Run an analysis locally, or submit it to a computing environment you manage through SSH, Slurm, Kubernetes, Hugging Face Jobs, Modal, or Ray. Review the target, command, resources, and outputs before approving a remote job. Cloud jobs may incur charges on the connected account.
Model Providers
AIPOCH Open-Science is model-agnostic at the product level: connect it to major cloud LLM providers, a custom gateway, or reuse an existing Claude or Codex subscription. Provider availability currently depends on the selected agent backend and the API protocols it supports. The following connection options are available:
Provider mode
How it works
Built-in cloud providers
Choose from the provider list shown by the installed app and authenticate with the requested key.
Custom Gateway
Supply a Base URL and exact model ID with an API protocol supported by the selected agent backend (Messages, Chat Completions, or Responses), then run the connection test. Remote gateways require HTTPS and an API Key. Loopback endpoints such as localhost, 127.0.0.1, or [::1] may use HTTP without a key; presets include Ollama, LM Studio, llama.cpp, and vLLM. A default API format does not guarantee server or model compatibility.
Codex Subscription
Select Open-Science Agent or Codex, then choose Codex Subscription as the provider type. Import an existing Codex sign-in or sign in with Open-Science.
Claude Subscription
Sign in with a Claude subscription in two modes: shared (a browser login that stores credentials in your default ~/.claude profile) or isolated (an app-managed claude setup-token run under an app-owned CLAUDE_CONFIG_DIR, fully isolated from ~/.claude/, with a browser flow plus a paste-a-token fallback).
xAI (Grok) OAuth
Use the built-in xAI (Grok) OAuth device sign-in. Access depends on your xAI account and the selected backend’s protocol support; validate the connection in the app.
There are 25 preset provider options, including OpenAI, Anthropic, DeepSeek, NVIDIA Build, and others; available models and regional endpoints depend on the installed version and selected agent backend. Check the provider selector and connection test in the app.
Data, Permissions, and Trust
AIPOCH Open-Science stores project data, settings, artifact versions, and provenance evidence on the local computer. API Keys are kept locally and use the operating system’s secure credential storage when it is available. Logs are local and are not uploaded automatically.
External data flow is still possible and should be reviewed:
Model requests send the prompt and necessary context to the selected model provider.
Web searches and remote connectors send their displayed parameters to external services.
Local connectors may execute trusted commands on the computer.
The application can also contact update servers, marketplace catalogs, and runtime or model download services.
Attachments, @ references, logs, and generated reports may contain sensitive research data.
Choose the narrowest permission profile that fits the task:
Mode
Behavior
Recommended use
Ask for approval
Requests approval for actions not already covered by scoped grants or trusted application tool policies
New workflows, sensitive data, unfamiliar scripts
Auto-approve edits
Uses the backend’s native auto review when available; otherwise automatically allows clearly low-risk workspace operations
Trusted file-editing work with controlled external access
Full access
Automatically allows edits, commands, network, and connectors
Clearly scoped, fully trusted, unattended work
The effective profile depends on the selected backend and existing grants. Connector, tool, and compute-network policies also apply; check the effective mode shown by the app.
Review connector parameters and tool activity before approving them. Never include API Keys, access tokens, patient identifiers, unpublished data, or sensitive local paths in screenshots or public issue logs.
Development & Packaging
AIPOCH Open-Science has an ordinary Node backend and an Electron desktop client, built with React, TypeScript, Prisma/SQLite, and the built-in Open-Science Agent runtime alongside optional agent backends.
The same Web UI can run on an ordinary Node backend without Electron or a graphical display.
Electron starts or attaches to that same Node service. One runtime owns each data directory.
npm run dev:web # Build and run the Node backend + Web UI
npm run pack:backend # Build a target-specific standalone npm tarball
For a published installation, use npm install --global @aipoch/open-science; the commands above
are for source development. Install the package or locally built tarball, then run
open-science start --no-open and open-science url to open the
existing UI. See the standalone runtime guide for installation,
secure credential storage, host capability boundaries and data compatibility.
The desktop app and browser share the same service. App connections and an explicit open-science start hold keep it running. Quitting the last App stops the service only when no explicit hold remains.
Mobile remote access
Scan a QR code to open the current workspace on a phone or tablet over local WiFi or a Pinggy tunnel. Local WiFi pairing stays on your network; internet tunnels route traffic through Pinggy’s servers, as disclosed in the connection dialog.
Headless CLI and SDK
The headless CLI and zero-dependency Node.js SDK use the same local daemon, projects, sessions,
credentials, and permissions as the desktop and web interfaces. Detailed usage lives with the
publishable package so there is one command reference to maintain:
CLI guide - installation, service lifecycle, task automation,
artifacts, output formats, and exit codes
The standalone CLI can expose its research tools to external MCP clients through open-science mcp. Start and authenticate the selected profile’s service first. The adapter uses that profile’s existing permissions; it is not a separate sandbox or read-only view, so connected clients can persist changes and incur model costs. It does not expose arbitrary shell or RPC access, credential management, file import/export, download-to-path operations, or a remote HTTP MCP endpoint.
Frequently Asked Questions
Why does the model connection test fail?
A: Check the API Key for missing characters or spaces, verify the Base URL and region, use the provider’s exact model ID, and confirm network access and account balance. For a Claude subscription, retry the shared browser login or refresh the isolated claude setup-token credential, depending on the selected mode.
Why is Continue disabled during setup?
A: The current step has not met its required condition. Fix any environment row marked Action needed, install or repair the selected agent runtime, or validate the model provider, depending on the active step. Notebook setup is optional and only affects Notebook execution.
How do I run jobs on a remote HPC cluster?
A: Remote Compute (SSH) is always enabled and does not need to be enabled in Settings. Register an SSH compute host under Settings → Connections → Compute, make it available to the current session, and then use natural language or /remote-compute-ssh. You need a reachable SSH host, valid authentication, permission for the required directories, and the software, dependencies, and compute resources required by the workload. Direct SSH does not require a scheduler; Slurm mode requires a working Slurm environment and permission to submit jobs. “Always enabled” refers to the skill, not to the availability of every registered host.
Kubernetes, Hugging Face Jobs, Modal, and Ray are also available for user-connected targets. Before each external submission, approve the target, command, resources, and outputs. Recovery records resume result collection after a restart, and published results open in Compute Preview. Provision the required account, software, resources, and permissions yourself; cloud jobs may incur charges.
Is there a command-line interface?
A: Yes. Install the standalone CLI, or use Settings → General → Command line tool → Install command to control an already running desktop instance. The npm installation requires Node.js; Homebrew and standalone bundled-runtime packages include it. Initialize the local profile, then control the service and submit research tasks without opening a browser:
# Initialize the local CLI profile and start the service in the background
open-science init
open-science start --no-open
# Create a project and run a task by its exact name
open-science project create "Systematic review"
open-science run --project "Systematic review" \
--prompt-file ./task.md \
--approval-profile auto \
--skill literature-review \
--wait --json
# Download a generated artifact
open-science artifacts list <session-id> --json
open-science artifacts download <artifact-id> --output ./report.md
See the CLI guide for the full command reference, JSON/JSONL output formats, exit codes, and headless service options.
How do I inspect where a generated result came from?
A: Open the generated artifact and choose Provenance. Select a version to inspect the content identity and the available producer code, execution history, inputs, environment inventory, producing conversation context, and reviewer evidence. Evidence AIPOCH Open-Science could not verify is marked unavailable.
Can I revise an earlier request without losing the conversation that followed?
A: Yes. Edit a completed user message and resend it to create a new branch from that point. The original later turns remain available, and the revision arrows beside the message switch between the alternative paths.
Contributors
Thank you to everyone who contributes code, documentation, translations, bug reports, and community support to AIPOCH Open-Science.
Before opening a public issue, remove API Keys, tokens, private file paths, unpublished data, patient identifiers, and other sensitive material from logs and screenshots. See CONTRIBUTING.md for the development workflow.
If AIPOCH Open-Science is useful to you, star the repository on GitHub.
For shipped, partial, and planned capabilities, see the Capability Map.
@misc{ouyang2026aipochopenscienceopensourceai,
title={AIPOCH Open-Science: An Open-Source AI Research Workbench with Reproducible .science Packages},
author={Wen Ouyang and Bocheng Huang and Qianyu Yao and Fei Sun and Minying Deng and Haoyue Cheng and Wenwen Zhou and Lei Mou and Yusi Wei and Jiang Li and Kun Zhu and Lei Wang and Yi Feng and Jiaheng Dai and Ke Qian and Dan Wang and Yahui Wang and Chenyi Wang and Mingsheng Lyu and Yaqiang Zhang and Ping Zheng and Chunhua Tang and Xinge Ma and Xu Zhang and Youjia Fan and Can Lu and Zhijun Dai and Man Hu and Qiqi Xie and Beinan Liu and Yuanning Guo and Yunze Liu and Jinhong Jiang and Xianjun Fu and Mingwei Zhang and Yuanji Xu and Shilin Xia and Bowen Yang and Mingjun Yu and Jingdun Xie and Xuwei Chen and Xia Zou and Yuzhou Xue and Jianguo Zhou and Chunhao Cao and Donghao Tian and Lei Huang and Jinxiao Li and Shengrong Long and Zefeng Song and Honglin Li and Jiang Luo and Wenying Jiang and Shanshan Li and Lei Dai and Ning Jiang and Ying Cai and Chaohui Li and Chunyan Peng and Hetong Zhao and Angqing Li and Chunyan Yang and Chongyang Sun and Yong Li and Xinyu Shi and Wei Dai and Ming Peng and Chen Sang and Yuanyuan Han and Desong Qiu and Haixin Miao and Jian Xu and Yi Chen and Xiaoliang Shi and Tian Shen and Teng Hu and Shuhao Wang and Yingyong Hou and Huimei Wang},
year={2026},
eprint={2610.17110},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2610.17110},
}
![dependabot[bot]](https://avatars.githubusercontent.com/in/29110?v=4)
评论