You normally use these tools through a conversation rather than calling them yourself. State the input, intended output, and limits.
Which tools the agent sees follows permissions, not the wording of your request. Research starts from a default set (bash, read, glob, grep, edit and write or apply_patch for GPT-family models, webfetch, research_search when a search provider is connected, literature, recall, todowrite, task, skill, question in interactive clients, python, compute_job, artifact). Everything else appears when a loaded skill declares it in allowed-tools, when the agent's configuration allows it by name, or when a specialist worker carries it: r, the biological query tools, the science_* connectors, experiments and study, generate_image, lsp, codesearch, and remote compute.
Read, search, and edit files
| Operation | Tool names you may see | Example request |
|---|---|---|
| Read a file | read | Read methods/protocol.md and summarize the assay steps. A PDF over the provider's page limit (100 pages) is read as text without being attached; an image the provider's content filter refuses is withheld from the transcript and the step runs again without it. |
| Find paths | glob | Find analysis scripts and data dictionaries in this project. |
| Search file contents | grep | Find every place the analysis applies a sample exclusion. |
| Edit a file | edit, apply_patch | Correct the unit conversion and show the resulting diff. |
| Write a file | write | Save the agreed methods note to results/methods.md. |
| Run a command | bash | Run the project's documented tests and report failures. A command with no timeout given stops after twenty minutes; longer work goes through compute_job. |
Give exact paths for important inputs, and preserve original data. Use Code and terminal for repository work and Files for locations and editing.
Calculate and run experiments
| Operation | Tool names | Usage guide |
|---|---|---|
| Python execution | python | Run calculations, inspect values, and save plots. |
| R execution | r | Analyze with R and save reproducible scripts. |
| Planned and detached compute jobs | compute_job | Plan, start, inspect, cancel, and recover outputs. |
| Supported provider account/status checks | provider_compute | Connect remote resources and understand scope. |
Prepare a compute plan for the existing analysis script. Verify input
paths and expected outputs, show the target and resource request, and
wait for approval before launching the job.
Search and retrieve information
| Operation | Tool names | Usage guide |
|---|---|---|
| Web, research, news, and developer search | research_search | Choose source, filters, and content depth. |
| Read a known web URL | webfetch | Read and verify source documents. |
| Find and read papers | literature | Search OpenAlex and arXiv together, then read the full text by DOI or arXiv id. |
| Recover something seen earlier in this session | recall | Searches every earlier message, tool result and saved tool output by regular expression, including turns compaction summarized away. |
| Search technical reference material | codesearch | Use source-grounded code and repository workflows. |
| Discover scientific databases | science_list_dbs | Choose a source by domain. |
| Search a scientific database | science_search | Use native queries and source identifiers. |
| Fetch a record or supported scientific file | science_fetch | Retrieve and validate the selected record. |
Specialized biological queries
The agent also has targeted query tools for common biological sources:
| Source | Tool | Ask for |
|---|---|---|
| UniProt | query_uniprot | Protein records and accession-based information. |
| Ensembl | query_ensembl | Supported gene, transcript, and genomic lookups. |
| KEGG | query_kegg | Supported pathway and biological record queries. |
| PubMed | query_pubmed | Literature records for a focused scientific query. |
| NCBI Gene | query_ncbi_gene | Gene records with the intended organism and identifiers. |
| STRING | query_string | Supported protein-association queries with organism context. |
| PDB | query_pdb | Structure records and identifiers. |
Use identifiers returned by the source, confirm the organism, and distinguish an association or prediction from an experimentally established mechanism. The database directory links each registered source.
Plan, delegate, and retain outputs
OpenScience can track work, ask questions, delegate a bounded task, load a skill, and save a file as a Result. These operations support the workflow you request; they do not define its scientific acceptance criteria for you.
| Capability | How to request it |
|---|---|
| Track steps and progress | Define a plan or goal; the plan panel tracks the steps. |
| Ask about missing information | Choose the appropriate independence setting and state important uncertainties. |
| Delegate independent work | Enable delegation and give separable questions with a shared output format. |
| Load a procedure | Select a skill and describe its task. |
| Generate or edit an image | Use the image tool with an output path. |
| Retain a file | Ask to save the verified output as a Result. |
For programmatic workflows, artifact supports save_file with a workspace
path, and read_file with an exact artifact_id and version_id. Reading a
version verifies its stored hash and stays within the current project. Text is
returned in bounded windows with a byte offset for continuation; binary Results
and files above the 8 MiB inline limit remain available through Files and the
artifact API. Large-file responses explicitly contain stored metadata only.
Saving a file captures its
bytes; it does not by itself verify the file's scientific claims.
Historical tool inputs retain their complete arguments when older outputs are
compacted. A copied legacy …[+N chars] argument preview is rejected before a file
mutation when its source can be identified. Literal examples already in a file
remain editable. Recover the original source before using a shortened preview as
input. When the model or permissions change the advertised tool set, the next
model request receives a compact added/removed notice; permissions still apply at
execution time.
| Keep sources and evidence | Ask for a source list and claim-level evidence in the research deliverables. |
Optional language-service or batch operations can appear in configurations that enable them. External MCP servers, plugins, and project tools add their own operations; inspect their documentation and availability before relying on them.
Read tool results carefully
A tool result can report success, a partial result, a prerequisite, or failure. Check the output file, source coverage, job state, or validation result appropriate to the operation. If a run was interrupted, inspect existing work before retrying it.