Why Keyword Search Makes You Miss Grants
There are a great many government support programs, and their names often diverge from who they're actually for. Search 'AI startup,' for instance, and you mostly get R&D projects or startup packages — yet the same team could also apply to content-production grants, art-meets-tech residencies, or social-service startup support. Programs like those don't surface under the keyword 'artificial intelligence.'
So to survey them properly you need not a keyword search but an exhaustive one — scanning every open notice without skipping a single one. The catch is that this runs to hundreds of notices. Doing it by hand every time is exhausting, and when you're exhausted, you miss things. ir-search starts from the idea of handing that exhaustive scan to a tool.
What It Does and How
ir-search crawls notices from five main sources: K-Startup (founding), Bizinfo (SMB support across all ministries and local governments — the widest coverage), NIPA (AI/ICT), KOCCA (content), and SMTECH (SMB ministry R&D). Other bodies (NIA, IITP, IRIS, and more) are handled via a built-in source registry.
Rather than just listing what it collects, it first narrows candidates against your project's profile — startup stage (pre-founding / early / scale-up), region, and what you need (funding / space / R&D). Then it opens the full text of each candidate's detailed notice and checks the eligibility criteria (age, years in business, region, target) one by one to verify whether you can actually apply. The crawlers also run standalone as Python scripts, so a developer can wire them straight into a research pipeline.
It Sorts Them Into Three Groups: A, B, C
The verified programs are organized into three groups. Group A is 'qualify right now' — programs you meet as you currently stand, sorted by nearest deadline. With the deadline and D-day attached, you can tackle the urgent ones first.
Group B is 'once you meet a condition (roadmap)' — programs that aren't open to you now but unlock once you hit a trigger like incorporating or raising investment. It even shows the chain of 'do this, and this opens,' so you can work backward from grants to design your growth stages. Group C is 'eligible if you reframe' — programs you qualify for by restating your item in another field's language. Rephrase 'AI voice technology' as 'an audio content-production pipeline,' say, and you become a target for content grants. On top of this it adds absence checks like 'not currently recruiting,' and a deadline-ordered priority action list.
Install and Use — One Line Does It
Installation is one line. Run the install.sh below and it auto-detects your installed agents (Claude Code, Codex, agy, Gemini), installs to all of them, and even handles the Python dependency (curl_cffi). Claude Code also works via the plugin marketplace, and Cursor / Grok Build work with a single clone into the shared skills directory.
Using it is one line too. With your project folder open, say 'run an exhaustive grant survey for our item,' and the agent reads the project info from the folder, asks only about the missing fields, and begins. It's designed for repeat use: your profile is saved to a file in the project folder, and on re-survey it automatically diffs against the previous result to report only new notices, deadline changes, and closed opportunities — meaning it doesn't re-read 250 notices every time.
curl -fsSL https://raw.githubusercontent.com/djfksjd/ir-search/main/install.sh | bashrun an exhaustive grant survey for our item
# or in Claude Code: /ir-searchCan You Trust It — Source URLs and 'Unknown' Marks
The most dangerous thing in grant information is a plausible wrong answer. Misstate a deadline or an eligibility rule and the whole application becomes wasted effort. ir-search attaches a source URL to every notice and marks anything not stated in the notice as 'unknown' rather than guessing. It's designed so a human can trace the basis of each judgment back to the original text.
That said, remember that the tool's output reflects the notice text at the time of the survey. Deadlines, eligibility, and amounts change frequently, so it's safest to confirm with the receiving agency before applying. The tool takes the 'exhaustiveness' and 'speed' of the search; the human does the final check and decision — that division of labor is the right way to use this skill. The author (djfksjd) released it under the MIT license, so if you're interested, it's worth checking the repository directly.