Everything below is paste-ready. Fields appear in the order Devpost shows them (Manage team → Project overview → Project details → Additional info → Submit). Character limits are from the form. Copy was written to read like a person, not a model; edit freely but keep the numbers, they're all sourced from the impact model.
Grantweaver: turn your nonprofit's Slack into funding
(53 characters. Plain Grantweaver also works if you want the clean brand.)
A Slack agent for the whole grant lifecycle. It finds live federal funding, pulls cited impact evidence out of your own workspace, and drafts proposals into Canvases. Nothing ever leaves Slack.
(193 characters.)
- Use
assets/grantweaver-white-bg.png, cropped square then padded to 3:2 (Devpost recommends 3:2, JPG/PNG/GIF, 5 MB max). - Downscale first, the raw file is ~1.3 MB at print size. 1200×800 is plenty.
- Same image family as the video title card so the gallery looks coherent.
Paste this whole block:
## Inspiration
69% of nonprofits lost funding this year, and their leaders rank funder
reporting as their #2 stressor, right behind funding itself. Small orgs
can't afford a grant writer, so the ED writes grants at midnight.
Here's what kept bugging us: everything a funder wants to see already exists
in the org's Slack. The attendance numbers a program manager posted in May.
A parent's thank-you message from March. It all scrolls away, and six months
later somebody re-interviews the whole staff to reconstruct what the
workspace already knows.
Until this February, no third-party tool could safely touch that history.
Slack's Real-Time Search API changed that: query-based, permission-aware,
nothing copied out. So we built the tool that couldn't have existed five
months ago.
## What it does
Grantweaver runs the full grant lifecycle without leaving Slack.
**Discover.** Ask it for funding and it searches Grants.gov live, through an
MCP server we built, then scores each opportunity against your mission
profile. One click adds it to your pipeline.
**Gather evidence.** The agent searches your own workspace with the
Real-Time Search API and returns evidence cards, each one linked to the
original message. Staff can file new evidence just by reacting with 🧵.
**Draft.** It re-reads that evidence live, works it into a letter of intent
or a funder report, and streams the draft into a Canvas. Every claim carries
a citation back to its source. A human reviews everything before it goes
anywhere.
**Track.** The App Home holds the pipeline board, deadline nudges, a weekly
digest, and an Impact Meter that shows its own math.
## How we built it
Node and Bolt JS on the new `Assistant` surface: streaming replies, suggested
prompts, status updates, feedback buttons. The agent core is a
provider-agnostic LLM tool loop (any OpenAI-compatible endpoint; we picked
the model with an empirical tool-calling bake-off that ships in the repo).
The evidence engine is `assistant.search.context`, with a keyword fallback
when semantic mode is unavailable. MCP runs in both directions: the agent
consumes `grantsgov-mcp`, a Grants.gov server we wrote, and exposes
`grantweaver-mcp` so Claude, Cursor, or Agentforce can read the same
pipeline. Postgres stores permalinks and metadata, never message content.
Drafts land through the Canvas API. Hosted on Railway.
## Challenges we ran into
Getting the RTS `action_token` plumbing right on a bot token took longer
than the feature it powers. Keyword-mode search needed OR-expansion before
it felt semantic. Canvas markdown has opinions about formatting that we
learned one draft at a time. And keeping card order deterministic while the
agent streams was fiddlier than it sounds.
## Accomplishments that we're proud of
The citation chain survives end to end: you can click any claim in a draft
and land on the Slack message it came from. The privacy model is
architectural rather than a policy promise, since we store pointers and
re-read sources live, deletions and permission changes are respected
automatically. And the Impact Meter means our impact claims and the
product's own telemetry are the same numbers.
## What we learned
The Real-Time Search API rewards restraint. Fetching less and citing
everything beat fetching more every time we tested. We also learned that
nonprofit staff will not tag evidence in a form, but they will react to a
teammate's win with an emoji, so that became the capture mechanism.
## What's next
Nearest term: five pilot nonprofits, then a Slack Marketplace listing.
Private-channel evidence via user-token OAuth, plus foundation and state
funding sources; each new source is just another MCP server plugged into the
same funding plane. After that, coalition co-applications over Slack
Connect. The engine itself generalizes: schools, research labs, and city
agencies all turn institutional memory into cited, human-reviewed documents.
Grants are the wedge.Optional: embed the impact chart inside the story with
once the repo is public.
slack-bolt, slack-api, node.js, model-context-protocol, postgresql, grants.gov-api, railway
Plus the LLM provider tag once the bake-off winner is final (e.g. gemini
or nvidia-nim).
- GitHub repo URL (public before submission)
- Demo video YouTube URL (public, captions on, under 3:00)
Paste the sandbox workspace URL, e.g. https://yourorg.slack.com.
Before submitting, invite testing@devpost.com and slackhack@salesforce.com to the sandbox and confirm they can sign in. The form warns that missing access can disqualify the project. Do this a day early, not at the deadline.
We calculated it instead of claiming it; the full model lives in our repo.
Per organization: a small nonprofit with no grant writer spends roughly 284
staff-hours a year on grant work across discovery, evidence assembly,
drafting, and funder reports. Grantweaver returns about 236 of those hours,
six staff-weeks, or roughly $8,300 in staff capacity. Capacity is what caps
small-org applications, so those hours convert. Two more applications a year
at the sector's typical 20% success rate and a modest $25K average award is
about $10K in new funding per org, per year.
At scale: 1,000 orgs adopting (against 1.8M US nonprofits) means 236,000
hours a year returned to mission work. The equity point matters most to us.
About 88% of US nonprofits run on budgets under $1M and can afford neither a
grant writer nor $180/month discovery SaaS. Grantweaver gives a six-person
org the grant operation that well-funded orgs take for granted, on a Slack
plan that's already free for them.
Beyond the workspace: every marginal funded program lands on beneficiaries
who never touch Slack. Our grantsgov-mcp server is open source for any
civic-agent builder, and the zero-retention pointer pattern is documented as
a reference architecture for privacy-preserving Slack AI apps.
The claims are auditable inside the product: the App Home Impact Meter
computes the same counters live and discloses its heuristic in a tooltip.
Leave blank (we're entering the For Good track). If you decide to enter both
tracks, the App ID is on your app's Basic Information page at
api.slack.com/apps, the value starting with A.
Leave blank. Grantweaver is new for this hackathon.
- Render the architecture Mermaid diagram at mermaid.live, export PNG at 2x.
- Badge the three required technologies visually: Slack AI surface, RTS API, MCP (both servers). Judges should get the whole story from this one image.
- Sanity check that every box in the diagram matches a file that actually ships in the repo.
- Video: public YouTube, under 3:00, burned-in captions, no copyrighted music, no real-person data.
- Gallery: 5 or 6 images at 1600×1000, hero first.
- Repo public, README standing alone (no references to planning docs).
- Sandbox access confirmed for both judge accounts.
- Run the final story text through a fresh read-aloud pass. If a sentence sounds like a press release, cut it.