Search, read, test, conclude.Every step leaves evidence.
An open-source research workbench you drive in plain language. It runs the whole loop, from literature to ideas, experiments, reviewed conclusions and a report. Everything stays on your machine, and every conclusion can be traced back to the run that produced it.
Chat assistants answer questions. Research needs a record.
A general assistant can summarise papers and write code. It forgets the project between sessions, grades its own claims, and treats a pipetting step like any other tool call. Research Agent is built around the three places that breaks.
A chat log is not a research record
Six months later you need to know which papers, which run and which observation a conclusion rests on. A conversation history cannot answer that.
A model should not grade its own claims
When a model says a citation is real, an idea is novel or a result supports a conclusion, something other than the model has to check.
Physical actions need a named person
An agent that can write a protocol, approve it and run it has no safety gate at all. Wet-lab steps have to stop and wait for a human.
One project, one loop, one evidence graph
Your research question is the project. Each stage writes its outputs into the same graph of records and links, so the final report is assembled from evidence rather than from memory.
A · Literature and writing
Cross-source search with de-duplication, open-access PDF download, structured reading cards from full text, review drafts with every citation checked, BibTeX and CSL export.
B · Experiments
Dry: simulation platforms such as OpenMM, scanpy, pydeseq2 and cobrapy, with a design, dry-run, collect, analyse, conclude state machine that resumes after a crash. Wet: plain-language protocol to Opentrons code, safety gate, approval, official simulator.
C · Records
Typed records and links, artifact versions with lineage, a timeline, and a research report that carries its evidence chain. Projects export to JSONL with checksums and re-import elsewhere.
D · Novelty
Claims are extracted and searched against the literature. The rating is then checked by deterministic code, so a thin search cannot produce a confident "novel".
E · Conclusion review
Citation integrity, data consistency, capability labelling and statistical plausibility checks. Only conclusions a named person approves reach the report.
F · Compute
Plan, approve, run, collect. Any job that costs money, uses the network or needs a key waits for a person. Local execution works today; remote back-ends have their contract in place and are not live yet.
Three bets that shape the code
The project is the root
Every output hangs off a project: library, ideas, experiments, observations, conclusions and artifacts. State lives on disk, not in a process, so a crashed run resumes and a project survives a new machine.
Code constrains the model
Wherever a model's judgement drives a later action, a pure function checks it against computable facts. A citation must exist in the library. A conclusion must cite observations that really occurred. The model can be a witness, never the judge.
No automatic approval of physical work
Passing the safety gate is necessary, not sufficient. A protocol then waits in awaiting_approval until a named person approves that exact version, identified by its hash.
- Wet lab stops at the simulator. Protocols compile, pass the gate, get approved and run in the Opentrons simulator. No real robot is connected until the safety gate covers more phrasing.
- The safety gate's reagent list is finite. A reagent it does not know is reported as unchecked, never as compatible. Concentration and biosafety rules only see the simplest single-sentence phrasing so far.
- Retrieval ranking is still weak. Highly cited off-topic papers can outrank a field's foundational work. Add key papers by DOI.
- Reviews only cover the library. Citation checks prevent invented references; they do not guarantee coverage.
- Remote compute is not live. The job contract and recordings exist; the gateway does not yet.
From a terminal, a browser or another agent
The command line and the HTTP API are the source of truth. The web workbench mirrors them, and external agents can use the whole toolkit over MCP.
# one research thread, end to end spark-research project new folding \ --description "Can we drop the MSA?" spark-research lit search "protein structure prediction" --add spark-research lit read --all spark-research lit review --topic folding spark-research idea check <idea-id> spark-research exp run <exp-id> spark-research lab compile "Add 50 µL to a 96-well plate…" spark-research lab approve <exp-id> --actor "A. Chen" spark-research report export --out report.md
// connect any MCP-capable agent { "mcpServers": { "spark-research": { "command": "spark-research", "args": ["mcp"] } } } # 30 tools exposed: search, library, ideas, # novelty, experiments, records, reports … # 8 withheld on purpose: approvals, simulator # runs, conclusion review, paid compute. # An agent can propose; only a person approves.
Web workbench
spark-research server opens projects, sessions, the evidence graph, long-running tasks with live progress, and cost attribution per command and model.
Bring your own model
Model-agnostic, with any OpenRouter-compatible key. A budget cap stops a project at a set spend and keeps finished work. Keys live in a 0600 file read only by the daemon, never in prompts.
Extend it
Six extension points: skills, connectors, simulation platforms, wet-lab back-ends, safety rules and prompts. Three have scaffolds whose generated tests run on the spot. capabilities --json describes what is installed.
Open source, Apache-2.0, v0.10
Install a single binary or from source. Local-first. Issues and pull requests are welcome, and so are real research projects that push it past its current limits.