Start your first research project

CoAutoResearch is a research system for working with an autonomous agent over multiple iterations. Its recursive self-improvement loop uses recorded results, review feedback and reusable lessons to refine subsequent research decisions. Bring a question, a proposal, or work already in progress. You can discuss findings as the research runs and decide which suggestions to send back into the research session.

Supported coding agents: ✓ Codex CLI · ✓ Claude Code

Set up with your coding agent

Give your coding agent this instruction:

Set up and launch CoAutoResearch from https://github.com/YihongT/CoAutoResearch. Follow docs/agent-setup.md, reuse my existing coding-agent login, configure paper generation, verify the setup, and open the dashboard.

The agent checks prerequisites, reuses your Codex or Claude Code login, starts the dashboard and prepares the paper tools. Complete any interactive provider login yourself. Dashboard readiness and paper-tool readiness are separate: research can start while a missing PDF dependency is being resolved.

The dashboard needs Node.js 20+, Python 3.10+, Git and one supported coding-agent CLI. Paper generation additionally needs Python 3.12+, LaTeX, Poppler and the four writing skills installed by the product’s setup command. Projects may need their own scientific packages. There is no dashboard build step. See agent setup for the executable checklist and paper generation for the PDF environment.

For a source checkout, the manual entry point is:

git clone https://github.com/YihongT/CoAutoResearch.git
cd CoAutoResearch
node bin/auto-research.js doctor
node bin/auto-research.js ui

Keep the server terminal open. Open the URL it prints. If the port is occupied, use --port 8766; use --projects-dir /path/to/projects to choose where research is stored. Before choosing the npm distribution, compare its published version with the source version. See the CLI guide for installation and updates.

Define your research

Choose Create project, name it, and select a backend. Creating a project makes a research workspace; it does not start an investigation.

In the research brief, explain:

the question you want to answer and why it matters;
relevant material or existing work;
limits on time, compute, data access and scope;
what would count as useful evidence, including a negative result.

You can request a general research report without choosing a conference or journal. Its review follows the audience, article type and evidence requirements in your brief. If you specify a venue, its official guidance and representative papers become additional review inputs; a general report is not a claim of submission readiness.

Use + to attach materials. Uploads copy files into the project. Linked folders remain at their original location and must stay accessible to the server. A path on another computer is not automatically available to a remote server. Review attachment names and paths before sending.

Send your brief. The agent prepares the research direction and may ask for missing information. Review the resulting brief, discuss changes, and choose Start autoresearch when you want autonomous work to begin. Check that the brief preserves your limits and any actions awaiting separate approval. For example, “ask me before using the final test set” must remain a pending decision, not become permission to test as soon as a plan is ready.

Choose the agent settings

Settings → Agent separates installed CLI, authentication and discovered models. Use Refresh login and models after updating or logging in. A CLI being installed does not prove the account can run a model. Resolve discovery errors before starting; choose from the models your provider actually exposes.

The composer controls the model and supported reasoning effort for your next message. Autoresearch has its own launch controls. A running turn keeps the settings captured when it started. Paper generation has separate controls and shows its selected model before you launch it.

Work together as research progresses

A Trial is one focused research iteration. It can test a method, analyze an unexpected result, acquire evidence or resolve a research blocker.

Use the autoresearch panel to follow the current action and live activity. Trials keeps earlier iterations, Reviews shows their checks, and Resources shows available material. A result recorded in the project has passed the applicable workflow checks; it has not been externally published or proved scientifically correct.

Choose New chat to discuss findings while the research session works. These discussions do not directly change research files. Add to research draft transfers a reply or plan to the research input. Review or edit the draft, then send it to make it an instruction for the main research process. When the main agent is busy, the input explains whether the message will wait for the current turn. Adding a draft alone does not apply a suggestion.

Pause, continue or start over

Action	What happens
Pause after current turn	Requests a pause after the active agent turn. The current trial may still be unfinished. Pause requested means the request is pending; Paused means automatic continuation has stopped.
Resume autoresearch	Continues from the current research state. Optional guidance applies to this resumed loop; leave it blank to continue with existing instructions.
Stop current run	Interrupts the active execution. The current turn or trial may remain incomplete. Read its status before resuming.
Restart autoresearch	Returns to the brief and materials saved at the first start. Later research work and auxiliary chats are archived, and a new process begins at Trial 1. Review the confirmation dialog carefully.

When resuming an unfinished trial, leaving the new instruction blank keeps any previous resume guidance; entering new guidance replaces it. The resume dialog explains this rule before you continue. Resuming an unfinished v2 trial also preserves its scheduled human review checkpoint. Changing the checkpoint interval explicitly starts a new window; a pause by itself does not extend that window.

If a recorded trial cannot start the next trial because of a setup error, the panel shows Paused and the last startup error. Resolve that error (for example, use Update project template), then Resume. The completed trial stays recorded, and this technical recovery preserves the existing checkpoint unless you change its interval.

An unambiguous execution-directory delimiter error can be repaired within the same unfrozen Trial. The service checks the assignment, quarantined bytes and destination conflicts before restoring files, then asks the agent to reconcile references and cumulative computation before fresh validation and review. This placement recovery is limited to two attempts per Trial. It does not approve rejected results, repeat experiments, or reset the research budget. A retained eligible error offers Repair and resume; ambiguous writes, changed protected files, conflicts or exhausted recovery attempts show Needs attention with details. Earlier recorded results remain available. A template update may retain this pending correction, but does not start the agent; use Resume when ready.

Editing an earlier message is different from sending a follow-up. Read the confirmation for that message: resending can replace later conversation or rerun work. Use a new message when you want to preserve the conversation and add a correction. Continue in agent CLI copies a provider command; it is separate from the dashboard’s Resume autoresearch action.

Closing a browser tab does not pause the server. Keep the server process alive for work to continue. A stopped or restarted server can interrupt execution; return to the dashboard and inspect the recovery action instead of assuming that work finished. See multiple projects and remote access when returning from another machine.

Read the manuscript and generate a paper

Manuscript brings the research narrative, supporting evidence and remaining gaps together. After reviewed results are recorded and project agents are idle, choose Generate paper, enter a venue and year or a general research-report target, and review the model settings.

The internal agent writes from a frozen evidence snapshot, builds figures, checks citations and compiles the PDF. Follow progress in Manuscript; you can cancel. After a successful generation, Paper appears beside Manuscript for preview, source downloads and review notes. When replacing a draft, the previous PDF stays available if the new generation fails or is cancelled.

Read the claims, figures, citations, limitations and venue requirements before sharing a paper. A green check means a draft was generated, not that it is ready for submission. See paper generation and best practices for review guidance.