Turn your nonprofit's conversations into funding.
Grantweaver is a Slack agent for the whole grant lifecycle. It finds live federal funding on Grants.gov, digs cited impact evidence out of your own workspace with Slack's Real-Time Search API, drafts grounded proposals into Canvases, and keeps the pipeline in sync across the App Home, two Slack Lists, and a shareable web page. It does all of this without storing a single message.
Here's the problem it goes after. 69% of nonprofits lost funding this year, and small orgs can't afford a grant writer. The proof funders ask for (attendance numbers, parent testimonials, program wins) already exists in Slack. Someone posted it, three people reacted, and it scrolled away within a week. Grantweaver puts it back to work.
A small nonprofit with no grant writer spends close to 284 staff-hours a year on grant work. Most of it is not writing. It's browsing funding databases, scrolling channels to reconstruct evidence, and scrambling before deadlines. Grantweaver returns about 236 of those hours. That's six staff-weeks, every year:
Hours convert to funding, because capacity is what caps small-org applications, not ambition. Two more applications a year at the sector's typical 20% success rate and a $25K average award changes the trajectory:
Both charts are projections from a stated model, not measurements. The same counters run live in the app. The App Home Impact Meter tracks opportunities surfaced, dollars applied for, evidence items woven, and estimated hours saved, and it states its heuristic right on the Home tab. Our impact claims and the product's telemetry are the same numbers.
| Technology | What it does here | Where |
|---|---|---|
| Slack AI capabilities | The agent surface: streamed replies with a live task timeline, loading statuses, state-aware suggested prompts, feedback buttons, thread titles | src/assistant.js, src/agent/loop.js, src/agent/streamer.js |
| Real-Time Search API | The evidence engine: live, permission-aware search over messages and files, with citations back to the source, plus the onboarding evidence scan | src/agent/rts.js, src/services/scan.js, src/agent/tools.js |
| MCP | Both directions: grantsgov-mcp is a Grants.gov server we built and consume; grantweaver-mcp exposes the pipeline to Claude, Cursor, Agentforce, or any MCP client |
src/mcp/ |
A note on timing. Scraping workspace history into an external database was never safe, and it violates Marketplace policy. The Real-Time Search API (February 2026) made permission-aware, zero-retention workspace search possible for the first time. Grantweaver could not have been built five months ago.
Discover. Ask for funding in plain English. The agent searches Grants.gov live through the MCP server, then runs a batched fit assessment on the top matches: a 0 to 100 score and an eligibility verdict that quotes the deciding phrase from the notice. Grants you mark "not relevant" stay gone, and your reasons feed the next fit call as negative examples. Forecasted opportunities get a Watch button instead of an Add button, because you can't apply to a forecast.
Prove. Evidence lives in three layers. None of them store content.
- A live search tool over your watched channels, messages and files alike, rendered as cards with permalinks. Channels you excluded in settings never surface, full stop.
- An evidence locker of human-saved pointers. Save from a card button, a 🧵 reaction on any message, or a message shortcut. Every pointer is a permalink plus a tag, mirrored into an Evidence Locker Slack List and the App Home. Delete a row in the List by hand and the locker forgets it too.
- An evidence index built by the onboarding scan, rebuildable any time by asking "rescan my workspace". A classifier groups what the scan found into themes a funder would recognize, each with a strength rating. Only theme labels, counts, and permalinks are stored. The snippets die with the scan.
Draft. Nothing slow runs without a visible yes. Asking for an LOI posts
a confirmation card first. Confirm with a button or a ✅ reaction, or open
the scope modal to pin specific evidence, emphasize sections, or add notes.
The draft is gathered fresh every time: workspace search runs first because
the search credential expires in under two minutes, then opportunity
details, then one completion into the opportunity's persistent Canvas. One
canvas per opportunity, five fixed sections (Overview, Requirements, Draft,
Evidence, Activity), regenerated in place, never duplicated. Every claim
carries a [source](permalink) link. Anything the workspace can't back up
is marked [TEAM TO CONFIRM] instead of invented.
Revise as a team. "Request changes" opens a revision thread. Teammates
pile on requests as plain replies, and no model runs while they argue. One
"Apply changes" click confirms scope, applies everything in a single
completion, posts a summary of what changed, and counts the
[TEAM TO CONFIRM] spots still open.
Track. The pipeline lives in four synced places: the database, the App Home board, a Grant Pipeline Slack List, and each opportunity's Canvas overview. The sync runs both ways. Change the stage, owner, or deadline by hand in the List and Grantweaver adopts your edit, on the hourly sweep or on demand. Adding a grant auto-extracts an application requirements checklist from the notice itself (SAM.gov/UEI registration is always on it), and "mark as submitted" politely refuses until the checklist is done.
It doesn't wait to be asked. Standing watches sweep Grants.gov three
times a day and post fresh matches. Deadlines get nudges at T-14, T-7, and
T-2, with snooze. A drafting opportunity that goes quiet for four days DMs
its owner. A suggested or reviewing one that sits for ten days gets a "worth
a decision?" card. Monday brings a digest of new matches. Friday brings a
#memories recap of what the agent actually did that week.
Export. A draft can leave Slack as a working .md pack (the
opportunity, the requirements, and evidence re-read live at export time,
ready to paste into any AI assistant) or as copy-ready application answers
keyed to the checklist.
Show a funder. /grantweaver index mints a 7-day HMAC magic link to a
public web page at /org/<token>: the evidence index by theme, pipeline
stats, and recent activity. It's built from counts and links only, for the
board member who isn't in the workspace. Paste the link in Slack and it
unfurls.
Agents that act need receipts.
/grantweaver logsshows recent runs: surface, status, latency, tools called, token usage, and a sanitized error type./grantweaver stateshows the privacy-safe conversation state the agent carries between turns: goal, decisions, artifacts, and source pointers. Pointers, never text./grantweaver settingslets an admin exclude channels from AI entirely (the agent refuses to search or answer over them) and switch proactive workflows off. Changes land in an audit log.- Every run is tracked. Crossing a latency, token, or cost threshold writes an audit event and can alert an ops channel. A run can be cancelled while it's still going.
- The App Home shows pending confirmations, failed runs, and recent completions above the pipeline, so the first thing you see is what needs attention.
%%{init: {
'theme': 'base',
'themeVariables': {
'background': '#0c100d',
'primaryColor': '#0c100d',
'primaryTextColor': '#f8fafc',
'primaryBorderColor': '#2d6a4f',
'lineColor': '#e2e8f0',
'secondaryColor': '#0c100d',
'tertiaryColor': '#eab308',
'fontFamily': '"SF Pro Rounded", "Virgil", "Comic Sans MS", cursive, sans-serif',
'fontSize': '13px'
},
'flowchart': {
'curve': 'linear',
'rankSpacing': 60,
'nodeSpacing': 20
}
}}%%
flowchart LR
%% Styling Definitions
classDef slackHighlight fill:none,stroke:#00ffd0,stroke-width:2.5px,color:#ffffff;
classDef mcpHighlight fill:none,stroke:#c084fc,stroke-width:2.5px,color:#ffffff;
classDef engineHighlight fill:none,stroke:#eab308,stroke-width:2px,color:#ffffff;
classDef neutralNode fill:none,stroke:#94a3b8,stroke-width:1.5px,color:#94a3b8;
subgraph C1["Ingestion Feed"]
GOV["🌐 Grants.gov API Feed<br/>[Upstream Funding Source JSON]"]
GG["📦 grantsgov-mcp client/server<br/>[Orchestrated Model Context Protocol]<br/>src/mcp/grantsgov-server.mjs"]
end
class GG mcpHighlight;
class GOV neutralNode;
subgraph C2["Agent Grounding"]
RTS["🔍 Slack RTS API context generator<br/>[Zero-Retention Live Agent Vector Search]<br/>src/agent/rts.js"]
SCAN["⚡ Ephemeral Onboarding Scanner<br/>[In-Memory Context Window Grounding]<br/>src/services/scan.js"]
end
class RTS,SCAN slackHighlight;
subgraph C3["Execution Loop"]
LOOP["🤖 Multi-Agent Reasoning Loop<br/>[Stateful ReAct Tool Call Orchestrator]<br/>src/agent/loop.js"]
TOOLS["🛠️ Context-Aware Toolbelt<br/>[8 Live Slack Integration Bindings]<br/>src/agent/tools.js"]
DM["💬 Slack Assistant Interface<br/>[Streaming Token-by-Token Threading]<br/>src/assistant.js"]
CANVAS["📄 Slack Canvas API Writer<br/>[Persistent Workspace Document Sync]<br/>src/services/canvas.js"]
end
class LOOP,TOOLS engineHighlight;
class DM,CANVAS slackHighlight;
subgraph C4["State Storage"]
DB[("🛢️ Postgres Trust Engine<br/>[Write-Isolation Metadata Pointer Store]<br/>src/services/db.js")]
LISTS["📋 Two-Way Slack Lists Sync<br/>[Pipeline Board Bi-Directional State]<br/>src/services/lists.js"]
CRON["⏰ Event-Driven Cron Scheduler<br/>[Proactive Context Nudges & Digest Loops]<br/>src/services/scheduler.js"]
GW["🔌 grantweaver-mcp gateway<br/>[Exposed Local Tool Model Context Protocol]<br/>src/mcp/grantweaver-server.mjs"]
end
class DB,LISTS slackHighlight;
class GW mcpHighlight;
class CRON engineHighlight;
style C1 fill:#1c170c,stroke:#eab308,stroke-width:1.5px,stroke-dasharray: 5 5;
style C2 fill:#0b1d1c,stroke:#00ffd0,stroke-width:1.5px,stroke-dasharray: 5 5;
style C3 fill:#181021,stroke:#c084fc,stroke-width:1.5px,stroke-dasharray: 5 5;
style C4 fill:#0a1711,stroke:#2d6a4f,stroke-width:1.5px,stroke-dasharray: 5 5;
%% Flow Connections
GG -- "fetch feed" --> GOV
GG -- "live search" --> TOOLS
RTS -- "zero-retention query" --> TOOLS
SCAN -- "ephemeral theme index" --> LOOP
DM -- "user prompt" --> LOOP
LOOP --> TOOLS
TOOLS -- "write LOI with citations" --> CANVAS
TOOLS -- "mirror pipeline stage" --> LISTS
TOOLS -- "save metadata pointers" --> DB
CRON -- "trigger alerts" --> LISTS
CRON -- "background status check" --> DB
GW -- "read-only developer queries" --> DB
And one turn, end to end. This is why drafts can cite without storing:
sequenceDiagram
autonumber
actor U as User
participant S as Slack
participant A as Agent loop
participant R as Real-Time Search
participant L as LLM
participant P as Postgres
U->>S: "What evidence do we have that mentoring works?"
S->>A: message event + action_token (TTL ≈ 1 minute)
rect rgb(240, 246, 242)
note over A,R: prefetch: search fires before the first LLM call,<br/>while the token is freshest
A->>R: search_workspace(query)
R-->>A: snippets + permalinks (this turn only, never stored)
end
A->>L: system prompt + org context + prefetched evidence
L-->>A: narrated answer with citations
A->>S: evidence cards (one permalink each) + streamed reply
U->>S: clicks 💾 Save on a card
S->>P: permalink + tag only, no text
note over P: at drafting time every pointer is re-read live,<br/>so deletions and permission changes<br/>are respected automatically
You need a Slack workspace you admin, a Postgres database, and one LLM API key. Any OpenAI-compatible endpoint works; Gemini's free tier is enough.
- Create the Slack app. Go to api.slack.com/apps,
choose Create New App → From a manifest, and paste
manifest.json. Install it to your workspace, and enable the Agents & AI Apps toggle if Slack prompts for it. - Configure.
cp .env.example .env, then fill in:SLACK_BOT_TOKEN,SLACK_SIGNING_SECRET,SLACK_APP_TOKEN(a Socket Mode token withconnections:write), andSOCKET_MODE=truefor local devLLM_BASE_URL,LLM_API_KEY,LLM_MODELDATABASE_URLWEB_LINK_SECRET(any long random hex) andAPP_BASE_URL. These power the shareable evidence page; without them/grantweaver indexand the Home tab's Evidence Index button can't mint links.
- Run.
npm install npm run migrate npm run dev
- Set up. DM the Grantweaver agent and run
/grantweaver setup. It's a 60-second modal: mission, focus areas, which channels the agent may learn from and post to. Then reply "scan my workspace" when it asks. That typed message matters: Slack only issues the workspace-search credential on a real message, so the first evidence scan rides on it. - Ask. "Find grants that fit our mission." Add one to the pipeline, then ask for an LOI.
One service runs everything: the Slack app, the MCP endpoint, the evidence
pages, and the marketing site, all on one port. railway.json ships in the
repo. Create a Railway project with a Postgres plugin, set the same env vars
with SOCKET_MODE=false, and point your Slack app's request URLs at
https://<your-app>.up.railway.app/slack/events. The health check lives at
/healthz. Any Node 20+ host works the same way.
Two things to change for your own install: the app ID baked into the Home
tab's "Ask Grantweaver" button (APP_ID in src/surfaces/home.js), and the
marketing-site links that point at grantweaver.app.
The LLM client is provider-agnostic. npm run llm:bakeoff tests
tool-calling correctness and latency against a replica of a real agent turn,
so you pick a model with evidence instead of vibes. Reasoning models work
too: the loop detects completions that got truncated by hidden
chain-of-thought and retries with headroom.
Grantweaver exposes its pipeline as an MCP server, so Claude, Cursor, or any other MCP client can ask about your grants:
node --env-file=.env src/mcp/grantweaver-server.mjs # serves :7802/mcp
claude mcp add --transport http grantweaver http://localhost:7802/mcp \
--header "Authorization: Bearer $MCP_SHARED_SECRET"Then ask: "What's due in the next two weeks?" Six read-only tools are
available: list_pipeline, get_deadlines, search_grants,
list_watches, get_checklist, and get_impact_meter. Requests without
the bearer secret get a 401. In production the same endpoint mounts on the
main service at /mcp, so there is no second process to run.
| Subcommand | What it does |
|---|---|
setup |
Org profile modal: mission, focus, channels, digest and memories channel |
settings |
AI-excluded channels and the proactive on/off switch (audit-logged) |
index |
Fresh 7-day magic link to the web evidence page |
digest |
Post this week's digest right now |
watch list / watch remove <id> |
Manage standing Grants.gov watches |
state / logs |
Agent conversation state · recent run telemetry |
simulate <sweep> |
Fire any proactive sweep on demand (allowlist-gated) |
clear |
Delete the agent's own messages from your DM |
reset |
Wipe the org, with a confirmation card, and start over |
Grantweaver never stores message content, and that's enforced in code, not
policy. The data layer rejects any write containing text, snippet,
content, or message keys, and conversation state is scrubbed the same
way. The evidence locker holds permalinks and tags. At drafting time the
agent re-reads each source live through Slack's Real-Time Search, so
deletions and permission changes are respected automatically. Even a
generated draft is held in process memory until a human confirms it,
because drafts quote your messages verbatim and a database row would break
the guarantee. Drafts always cite their sources, and nothing goes out
without human review.