AIRAS

AIRAS - an open-source project for research automation

AIRAS is open-source software for automated research. It gives a coding agent (Claude Code, Cursor, or any MCP client) everything it needs to take a research topic through literature survey, hypothesis, experiments, and a finished paper, and it makes the paper's claims verifiable: the paper is preregistered in git before any experiment runs, every reported number is realized from run outputs, and CI re-checks all of it before the PDF of record is produced.

AIRAS ships as one PyPI package (airas) that provides:

• an MCP server with the research tools the flow needs (paper search, hypothesis and experimental design, experiment results, figure rendering, LaTeX, Overleaf, and the record/verification tools),
• a Claude Code plugin that bundles the server with the auto-research workflow skills,
• a small CLI (airas verify-record, airas verify-paper) that the experiment repository's CI uses as the verification gate.

Currently, it focuses on the automation of machine learning research.

Quick Start

No clone, no Docker. Only uv is required; uvx fetches the package on first run.

1. Install

AIRAS is meant to be driven from Claude Code through its plugin. The plugin installs the MCP server together with the auto-research workflow skills and the hooks that record the agent's state:

/plugin marketplace add airas-org/airas

/plugin install airas@airas

The MCP server can also be used on its own, without the skills and hooks. In Claude Code:

claude mcp add airas -- uvx airas

In any other MCP client, add it to the client's MCP configuration (e.g. .mcp.json):

{

"mcpServers": { "airas": { "command": "uvx", "args": ["airas"] } } }

Upgrading. uvx keeps the version it fetched the first time, so an existing install does not move to a new release on its own. Run uv cache clean airas (or uvx airas@latest) once to pick up the latest version.

2. Configure credentials

Credentials live in ~/.airas/credentials.json and are re-read on every tool call, so you can create or edit the file at any time:

mkdir -p ~/.airas

cat > ~/.airas/credentials.json <<'EOF' { "GH_PERSONAL_ACCESS_TOKEN": "ghp_...", "SEYVAL_API_KEY": "..." } EOF chmod 600 ~/.airas/credentials.json

3. Start a research project

In Claude Code, invoke the orchestrator skill and give it a topic:

/airas:auto-research

It walks through the flow below, asking you to settle the operational choices (repository visibility, execution platform, compute target) once up front. Other MCP clients call the tools directly; each tool's description says what comes before and after it, and the MCP documentation walks the flow.

The auto-research flow

auto-research owns only the ordering and the rules that span steps. Each step is its own skill with a stated contract, so you can also invoke a single step on an existing repository.

What the flow guarantees

• Declare before you run. Nothing is dispatched before the preregistration commit exists, and every run descends from it. Fixes are committed on top, never rebased away.
• The record only grows. .research/record.json is an append-only tree of hypotheses, claims, designs, runs, and results. A reworded claim, a changed run condition, or a dropped result all fail the same check. A claim that misses its criterion is reported as a negative result, not rewritten.
• No experimental number is ever typed. Numbers reach the paper only through \airasval{...} references to a run's measured metric or declared parameter, rendered from the record. Anything else is marked \unverified{...}.
• The gate is enforced, not advisory. Branch protection makes the verification workflow a required check, so a red run cannot be pushed past.
• The repository is the state. Everything a later step needs is committed, so a fresh session can resume from the clone alone.

Execution platforms and LLMs

Experiments run through the same three MCP tools on either backend: dispatch_experiment starts the run, get_experiment_run_status follows it, and import_run_outputs copies its outputs from where the backend keeps them (Seyval's storage, or the workflow's artifact on GitHub Actions) into .research/results/ with a provenance manifest. The record gate cross-checks the committed bytes against that same store, and once the store has dropped the run, against the sha256 hashes the import recorded. Seyval (bring-your-own Slurm compute) and GitHub Actions are supported; a backend for machines you run yourself (RunPod, a lab cluster) is not yet, since it needs a store the agent cannot rewrite.

The flow needs no LLM key: the agent authors the hypothesis, design, analysis and paper itself, guided by the skills. Backend-LLM tools (analyze_experiment, and verify_paper_values with a model) and get_prompts, which hands a client the prompts they use, exist for use outside the flow; the former need a provider key (get_available_llms lists the models your keys allow). Supported providers: OpenAI, Anthropic, Google Gemini, OpenRouter, Amazon Bedrock, and Vercel AI Gateway.

Companion repositories

AIRAS relies on three sibling repositories under the airas-org organization. Each keeps one piece of the workflow outside the agent's reach.

MCP tools

The auto-research flow uses the following tools; the skills above are thin contracts over them. The server exposes more, but these are the ones a research project goes through.

See the MCP documentation for descriptions, credentials per tool, and configuration options.

CLI

uvx airas                 # MCP server on stdio (default)

uvx airas verify-record # check .research/record.json against run outputs, git history, and the platform uvx airas verify-paper # check paper values and citations against the record and run outputs — no PDF build uvx airas verify-record --implementation-verifier-model # review implementation against declared designs uvx airas verify-paper --citation-verifier-model # judge unjudged passage citations uvx airas publish-paper # build the paper's PDF for publishing (values already verified by the gate)

verify-record checks declarations and execution evidence; verify-paper independently checks the paper against them. They are the two required checks the experiment repository's CI runs on the protected branch; publish-paper builds the PDF in the separate, non-required publish workflow.

Development

cd backend

uv sync uv run airas

Lint and type checks are wired through pre-commit (pre-commit install once per clone). See CONTRIBUTING.md.

Roadmap

AIRAS is developed in stages. Reliability comes first: an automated research pipeline is only worth scaling once its outputs can be trusted and reproduced.

1. Reliability (in progress). The record now guarantees that every number in the paper traces back to a declared run and its outputs, that each claim's verdict is derived from the criterion it declared