openbash

npx add-mcp 'https://openbash.ai/api/mcp'

openbash

Bash programs over data. One command line in, one answer out.

The programs below are commands on a path, not endpoints. There is ONE exec call, and a pipeline of several programs is a single request to it. They compose with pipes, redirection and the ordinary shell tools, so two of them can be joined without a schema for the join.

Which programs exist is decided per ACCOUNT, not per call. Once one is enabled every session that account opens has it, and "help" lists exactly those: so "help" is the authoritative answer to "what can I run", not this page.

Setting yourself up

If you are an agent reading this because somebody pasted a link at you, this is the step they meant:

curl -fsSL https://openbash.ai/install.sh | sh -s -- --key sk_ob_...

It writes https://openbash.ai/skill.md — how to use this shell, as instructions — into every agent directory on the machine, checks the key by running one free command, and registers the MCP server below with that key as a header. No browser, and nothing to install: there is no CLI, because the API is one HTTP call and curl is already its client.

Without a key it still installs the skill and tells you where to get one. With no key at all, the MCP door below runs an ordinary OAuth consent screen.

Three doors, one shell

HTTP   POST https://openbash.ai/api/v1/exec   {"command": "..."}
MCP    ssh -p 38449 <your-key>@ssh.openbash.ai '<command>'/api/mcp            one bash tool
SSH    https://openbash.ai

They are three ways into the same account-scoped shell, not three services. Whichever you use, the same programs are on the path and the same key decides which ones.

Output contract

• JSON on stdout, one object per line, so jq composes across programs.
• Every record carries source (what it was read from) and as_of (when that source was current). as_of is NOT the time you ran the command.
• A derived figure is marked "derived": true and carries its inputs. A figure you cannot check is a figure you should not use.
• Absence is a value. A program that finds nothing says so and exits 3; it does not return an empty object that reads as zero.

Exit codes

0   results
1   usage error. The flag or argument was wrong
2   not entitled. The program exists but this key does not have it
3   no results found. The query was valid and matched nothing
4   upstream data failure. The source could not be read

2 and 3 are different answers and it matters: 3 means there is nothing there, 2 means you were not allowed to look.

Reading the errors

Errors are the documentation. An unsupported flag names the flags that are supported; an unknown command names the near misses. Try something, read the failure, adapt. That path is designed to work.

Discovery

GET https://openbash.ai/api/v1/programs      the catalog, including programs this key lacks
GET https://openbash.ai/api/v1/programs/{slug}/surface/{read}  a program's browsable page as data, for the programs whose catalog entry carries a surface
GET https://openbash.ai/api/openapi.json     the OpenAPI spec for every endpoint
<program> --help                          the definition of one program

Keeping up

New programs are new capabilities on this shell, and there is no need to re-read the catalog and diff it to find them.

GET https://openbash.ai/api/v1/changelog     what was added and when, newest first, as JSON
GET https://openbash.ai/feed.xml             the same list as Atom
GET https://openbash.ai/sitemap.xml          every page, with the date it last changed

A program's date is when it first appeared, not when the catalog last synced, so an entry that is new to you is genuinely new.

Programs

• biotech: Targeted biotech studies as JSON: clinical trials, drug labels and approvals, company filings, literature.
• aero: Every flight movement for an airport-day, plus per-terminal and per-hour rollups, as JSON.
• trade: US customs bills of lading: a company's counterparties, its shipments, and the firms already importing what you make. And India's monthly trade by partner country, from the Commerce Ministry.
• outreach: Companies matching a description, the people at them, and their addresses — derived from the company's format, found by a finder, or verified against a mail server.
• funding: SEC Form D filings — almost every US private raise is recorded on one, within fifteen days of the first sale.
• webcertificates: Certificates issued for names at or under a domain — a new subdomain is a service going live, usually within hours.
• fcc: The equipment grantee register and the authorisations themselves — devices, dates and frequency bands.
• fda: Breakthrough-therapy and novel drug approvals, each with what it was approved to treat; and device establishment registrations, months before a product ships.
• us-trademarks: Trademark applications matched on the applicant, with the goods the company intends to sell under the mark.
• us-business-license: Liquor licences and retail food permits by place and date — a venue with a lease signed and no suppliers yet.
• ucc: UCC-1 financing statements naming a company as debtor, with the secured party. Colorado and Connecticut only.
• h1b-hiring: The labour condition applications an employer filed — job titles, wages offered, worksites, and new seats against renewals.
• company-hiring: Employers filing for sponsored positions in a place or discipline, ranked by the seats that did not exist before.
• us-motor-carriers: The federal motor carrier census, mirrored weekly: look a carrier up, list new fleets, or find the ones that added trucks.
• model-prices: The current version of every language model family, and what each provider charges for it per million tokens.
• perspectives: Qwen3.5-4B answers a prompt plainly and once per voice, each voice one neuron pushed.
• perspectives-nanbeige: Nanbeige4.3-33B answers a prompt plainly and once per voice, each voice one neuron pushed, with its reasoning shown when thinking is on.
• trends: Google Trends on web search or YouTube, autocomplete, and Wikipedia page views — free.
• reward-hack: Qwen3.5-4B reads a task and an answer and reports two reward-hacking neurons and a linear probe, token by token.
• weight-edit: One math problem answered by Qwen3.5-0.8B and by the copy an AI agent edited to stop its thinking loops, side by side, with the edit's signal every 64 tokens.
• scene: Two people, each with a look and a name you choose, talking and moving in an empty world, each with a personality pushed into Qwen3.5-4B; made, cast, started, paused and stopped by command, with a clip of every step.
• least-squares-neurons: Train Qwen3.5-0.8B on your own examples with forward passes alone: new neurons fitted by least squares, for labels (prompt -> label) or answers (prompt -> answer), scored against the unedited model, returned as a patch.
• search-index: Ask Google, Bing (ChatGPT search, Copilot) and the IndexNow engines to index your pages in one step, see what crawlers are missing on your site, and read your own Search Console.
• social: Profiles, posts, comments and search on X, LinkedIn, Instagram, YouTube and Reddit.
• realestate-us: Rental listings merged across sources, the HUD tables that decide what a subsidised rent may be, and tract-level ACS and opportunity data.
• realestate-in: Rental listings around a place in an Indian city, merged across backends and deduplicated.
• places: Find a place by describing it, get its id and coordinates, then fetch what it looks like.
• sec-filings: SEC filing text by ticker, 13F ownership, and the name-to-ticker lookup you need before either.
• ocr: Optical character recognition over a file on a drive or a url. The whole document, or each line with where it sits.
• vision: A vision model reads images from a drive or a url and answers in words: what is in it, what it says, whether two match.
• dbq: Read-only SQL against the account's connected database, with results as CSV on stdout.
• traffic: Estimated monthly Google organic visits for up to 1,000 domains at once, with the month and source.
• rank: Keyword position on Google, and who is above you.
• keywords: How many people search a phrase on Google each month, and the phrases people search near it.
• aeo: What ChatGPT, Perplexity, Gemini, AI Mode, Bing AI Mode and Google's AI Overview answer to a question, and which sites they cite.
• dns: Mail authentication, nameservers and registration.
• webaddress: Categories and subcategories in the shape of a classifieds front page, each holding the addresses worth knowing; ask any branch for its top ten, local to a place.
• registry: npm and PyPI downloads with a year of weekly history, GitHub stars, forks and last push.
• agentrank: Calls and calling accounts per tool on Monid, MCP server use on Smithery, users and runs per tool on Apify, MCP image pulls on Docker, and which tools AI assistants recommend.
• series: The history a measuring program leaves behind, and which way it went.
• browser: Search the web, take originals, and drive a live browser session.
• pplx: Perplexity's index from the shell: ranked results for a question, and the passages of a page that bear on it.
• voice: Text spoken in a cloned or stock voice, as an MP3 filed under a URL; and voices cloned from the account's own clips.
• avatar: A photo animated and lip-synced to a voice track, as an MP4 filed under a URL.
• image: Gemini's image model draws from a prompt and up to four pictures: a photo restyled, a drawing changed in one place, the same character in another pose.
• tla: A business or a system written as a TLA+ module, checked by TLC: every state its steps can reach, each rule that breaks with its shortest trace, and the rules that hold.
• youtube: A public YouTube video watched by a model: its transcript with speakers and times, summary, story beats, characters, setting and on-screen text, as JSON.
• listen: Any audio or video file on a mount, or a link: every word timed to the millisecond with its speaker and confidence, pauses, and the voice measured over time (loudness, pitch, brightness, harshness, jitter, shimmer, learned quality scores). Given the words that should be said, each word judged against its root language, sound by sound. Time series, never a summary.
• pdftext: A PDF's text as markdown, with a readability check that runs without extracting.
• doctext: Text of a doc, docx, ppt, pptx, rtf, odt, odp or epub as markdown.
• edi: X12 as CSV: the 820 remittance and the 812 adjustment, with the reason code on every line.
• xlsx: Read and edit an Excel workbook the way Excel does: values and formulas, rows, columns and sheets inserted and deleted with every reference following, formats, names, pivots, and every formula recalculated on save.
• csv: Columns, counts, frequencies, sums, averages, filters and joins over CSV.
• jq: Filter and reshape JSON, the way every program here emits it.
• text: awk, sed, grep, rg, cut, tr, sort, uniq and the rest of the text coreutils, in the jail.
• files: cat, ls, find, cp, mv, mkdir, rm, stat, du, diff and the path helpers.
• archive: gzip, tar and zip, plus base64, checksums and hexdumps.

The shell

What the shell IS — the mounts on it, what this key can run, and the convention for citing sources in an answer — comes from the shell itself, not from this page. Run help through any door above, or request this page with your key and it will be included here.