Read tools
Run automatically
Search, read, count, calculate, compare, profile a table, run a backtest. They look at things and change nothing.
An agent whose steps you can see.
Give it a multi-step task. It searches, reads, calculates and compares, and each step appears as it happens. It can create new documents for you, but it only proposes changes to the ones you already have.
Check the budget, work out an 18% contingency, and compare the two reports.
- Searching files for “budget”5 passages from 2 files12 ms
- Calculating 18% × 48,000= 8,640 · computed by Orbit’s calculator, not the model1 ms
- Comparing the two reports2 section changes, 1 changed figure4 ms
- Proposing an edit to Launch checklist › ScopeNot applied: waiting for youproposal only
Contingency at 18% is 8,640. The two reports disagree on the budget ($48,000 vs $52,000). I proposed one edit; nothing has changed yet.
A scripted preview of how the real interface behaves. In demo mode, plans are rule-based; with a live model the model chooses the steps.
What it may do is decided by the tool, not by the model.
Create tools
New, local and capped
Start a new document, a local draft or a brief suggestion. They never touch what already exists, and a run can only create a few.
Propose tools
Never change anything
Propose an edit and you see a diff before anything happens, or stop and ask you a question. You apply or dismiss.
Twenty-four tools. All local.
Every tool runs on your computer against your own files and documents. None of them reaches the internet, runs code or sends anything anywhere.
search_filesSearch filesFinds the passages in your files most relevant to a query, with reference ids it can cite.read_fileRead fileReads part of a file by line range and returns citable passages.list_filesList filesLists the files for this chat and its project, with type, size and passage count.list_documentsList documentsLists the workspace documents in this chat with version and last change.read_documentRead documentReads a workspace document, current or a past version, with its heading outline.find_in_textFind in textFinds the lines containing a word or phrase and returns line numbers.text_statsText statisticsWord, sentence and paragraph counts and a rough reading time.table_profileProfile tableRow counts, column types, sums, means, medians and data issues. Computed, not estimated.table_aggregateAggregate tableGroups by a column and sums, averages or counts another, returning the formula used.calculateCalculatorExact arithmetic from a hand-written parser. It is not a programming language and cannot run code.current_datetimeDate and timeThe date and time on this computer, and days until a date you give.compareCompareChanged sections and figures, removed sections, open questions and contradictions between two files.diff_versionsDiff versionsSummarises what changed between two versions of a document.project_briefProject briefReads the goals, decisions, open questions and next steps you have approved.list_draftsList draftsLists your local tasks and notes. Drafts are never sent anywhere.search_chatsSearch earlier chatsSearches saved chats. Temporary chats are never searchable.position_sizePosition sizeQuantity, money at risk and reward-to-risk from numbers you give. Arithmetic, not a recommendation.backtest_pricesBacktest a ruleTests a simple rule on price data you supply, with costs, and reports the last 30% out-of-sample.paper_tradesPaper tradesReads your simulated ledger. Nothing in it is a real position.create_documentCreate documentStarts a new document, table, code file or plan. It never touches an existing one.create_draftCreate draftMakes an inert local task, follow-up or note. Only you can approve it.suggest_brief_itemSuggest brief updateSuggests a decision or next step. Nothing joins the brief until you accept it.propose_revisionPropose editProposes a change to an existing document. You see a diff, then apply or dismiss it.ask_userAsk youStops the run and asks a question when it cannot continue without your input.
Limits on every run.
Each run is capped by the number of steps and by input tokens. You choose Orbit Swift or Orbit Depth.
Steps per run6
Input tokens per run30,000
Hit either limit and the run stops and tells you, instead of carrying on. Orbit Depth allows longer runs and uses your monthly allowance faster.
A grounding check on the answer
After the run, Orbit checks the figures in the final answer. Any number that no tool result or message of yours supports is flagged rather than presented as fact.
No hidden steps
The step list streams as it happens, with timings and what each tool found. Collapse it when you only want the answer, expand it when you want to check.
It cannot run code
There is no isolated sandbox, so generated code is never executed. The calculator is a fixed arithmetic parser, not a way in.
A closer look today. The working app comes next.
Explore the local preview through its screenshots and examples. Public accounts, uploads and live AI are not available yet.