Forge

A local-first, per-project workspace for founders. Each project gets a chat-first home with a persona-aware assistant that can read your data and write notes, tasks, metrics, and saved resources directly. Sales, investor outreach, and hiring all live in the same pipeline - separated by Lead.kind, not separate apps.

Underneath the chat sits a versioned wiki, AI-grounded retrieval, a tool-use loop, and an evolving set of agents (lead discovery, reply triage, call coach, meeting prep, pipeline health, ICP refiner, news ingestion).

Stack: Next.js 16 (App Router) · TypeScript · PostgreSQL + Prisma · Anthropic / OpenAI · Whisper · Vitest

What's in here

Per-project assistant (the headline)

Chat-first project home at /projects/[id] - replaces the old dashboard.
Five personas, switchable per conversation: Lead expert · Investor · Hiring · Writer · Brainstormer. Each has its own system-prompt addendum; tools are shared.
Native Anthropic tool-use loop. Tools the assistant can call directly: create_note, create_task, record_metric, save_resource. Tool calls show as breadcrumbs in the chat and persist on the message row.
Retrieval composes wiki docs + saved resources + project state into every turn.

Project primitives (each one schema + REST + assistant tool)

Note { kind: NOTE | DECISION | DOC_DRAFT } with zero-or-one link to a lead, call, or insight.
Task (project-scoped, optional leadId).
Metric (append-only time series, project-scoped, named).
Resource (saved URLs with auto-fetched OG metadata, user notes, free-form tags) - feeds Assistant context via keyword retrieval alongside the wiki.

Pipeline (the original spine, still here)

Lead.kind = SALES | INVESTOR | HIRE - same stage machinery for all three.
Pipeline stages and orthogonal conversation depth, scored 0–100 by a pure function.
Touchpoint logging across email / call / DM.
CSV bulk import.

Calls + AI notes

Audio (m4a / mp4) → Whisper → Claude structured extraction (summary, key points, objections, signals, commitments, sentiment).
Manual notes path for un-recorded calls, same downstream extraction.
Rolling company-level and project-level summaries.

Second-brain wiki (per-project, gated by wikiEnabled)

External source ingest: URLs (Mozilla Readability + Turndown), PDFs (pdf-parse, 10MB cap), notes, articles.
Page kinds: PROJECT_INDEX, COMPANY, PERSON, CALL, TOPIC.
[[path/to/page.md]] backlinks parsed and queryable.
Topics: fixed (objections, competitors, icp-patterns, pricing-feedback) plus LLM-discovered.
Consistency lint + imputation pipeline (see Architecture).

Agents (src/lib/agents/)

In-flow: lead-discovery, reply-triage, call-coach, meeting-prep.
Scheduled (node-cron, wired in instrumentation.ts): pipeline-health (daily), icp-refiner (weekly), news-ingest (daily).
All structured-output agents go through runtime.ts (Zod-validated, retry-once-on-parse-failure, per-agent token budget).
All runs logged to AgentRun for cost observability.

Project feed

7-day timeline on the project home: new articles (from news-ingest), completed tasks, recent calls, new insights.
Articles expose a one-click "Save to Resources" to promote into the curated library.

Highlights

Things in here that aren't standard CRUD work, with file pointers:

Anti-hallucination guardrail on LLM extraction - proposed citations are rejected unless the URL appears verbatim in the search results we showed the model. (src/lib/wiki/lint/imputation.ts - parseProposal)
Native Anthropic tool-use loop - multi-turn tool execution with provider fallthrough (non-Anthropic providers get plain chat). Tool calls persisted and surfaced inline in chat. (src/lib/llm-tool-loop.ts, src/lib/assistant-tools.ts)
Persona-aware system prompts - five personas, single registry, single prompt builder. Persona stored per conversation. (src/lib/personas.ts, src/lib/lead-expert.ts)
Resource retrieval composes with wiki retrieval - same keyword-scoring shape, weighted with userNote highest. Wired into Assistant, outreach draft, and meeting prep. (src/lib/resources/retrieve.ts)
Content-hashed, versioned wiki document store - every page is immutable, deduped by hash, with supersededById chaining and full history. (src/lib/wiki/store.ts)
Idempotent fact-apply - accepting a proposal upserts a bullet under a structured ## Facts section; re-running is a no-op, same field with a new value replaces in place. (src/lib/wiki/lint/imputation.ts - upsertFactsSection)
Pluggable search + LLM providers via DI - WebSearchProvider and LLMFn interfaces; tests inject StubSearchProvider and a fake LLM. No network in the test path. (src/lib/wiki/lint/search.ts)
Scope-driven incremental compile - orchestrator supports call, lead, company, topic, all scopes. (src/lib/wiki/compile.ts)
Pure-function lead scoring - 0-100 score, weighted factors, deterministic. (src/lib/scoring.ts)
Best-effort OG metadata fetcher - native fetch + regex, 5s timeout, 1MB cap, never throws. (src/lib/og-fetch.ts)

Architecture: the wiki AI pipeline

   ┌──────────────┐   ┌─────────────┐   ┌──────────────────┐
   │ Audio upload │──▶│   Whisper   │──▶│  Claude extract  │
   │ (m4a / mp4)  │   │ transcribe  │   │ structured notes │
   └──────────────┘   └─────────────┘   └────────┬─────────┘
                                                 │
                            ┌────────────────────▼─────────────────────┐
                            │  Wiki compile (scope=call)               │
                            │  → company / person / call / topics      │
                            │  → versioned, content-hashed, backlinked │
                            └────────────────────┬─────────────────────┘
                                                 │
                            ┌────────────────────▼─────────────────────┐
                            │  Consistency lint  +  Imputation         │
                            │  • find missing fields                   │
                            │  • web search per field                  │
                            │  • Claude extracts (JSON, URL allowlist) │
                            │  • persist as MISSING_DATA findings      │
                            └────────────────────┬─────────────────────┘
                                                 │
                            ┌────────────────────▼─────────────────────┐
                            │  UI: triage panel on company wiki page   │
                            │  Apply → upsert under ## Facts, new ver. │
                            └──────────────────────────────────────────┘

Architecture: the assistant turn

   User message on /projects/[id]
              │
              ▼
   ┌─────────────────────────────────────────────┐
   │  buildLeadExpertSystemPrompt(projectIds,    │
   │     query, persona)                         │
   │  → base prompt + persona addendum +         │
   │    project context + wiki retrieval +       │
   │    resource retrieval                       │
   └────────────────────┬────────────────────────┘
                        │
                        ▼
   ┌─────────────────────────────────────────────┐
   │  runAssistantWithTools (Anthropic native)   │
   │  loops: tool_use → execute → tool_result    │
   │  tools: create_note · create_task ·         │
   │         record_metric · save_resource       │
   └────────────────────┬────────────────────────┘
                        │
                        ▼
   AssistantMessage row { content, sources, toolCalls }
                        │
                        ▼
   Breadcrumbs + answer rendered in chat

Every step is independently invocable via REST endpoints under src/app/api/.

Setup

git clone https://github.com/ShuklaA11/Forge.git
cd Forge
npm install --legacy-peer-deps
docker compose up -d                 # PostgreSQL on :5433
echo 'DATABASE_URL="postgresql://postgres:postgres@localhost:5433/lead_management"' > .env
npx prisma migrate dev
npm run dev

Open the dev URL → Settings → add an Anthropic or OpenAI key. Tool-use (writes from chat) currently requires Anthropic.

Tests: npm run test (Vitest, no DB / network - provider stubs).

Scheduler note: Background agents (pipeline-health, icp-refiner, news-ingest) start with the Next dev/prod server via instrumentation.ts. Disable with AGENT_SCHEDULER_DISABLED=1.

Tech stack

Layer	Tool
Framework	Next.js 16 (App Router, Server Components, Server Actions, Turbopack)
Database	PostgreSQL 16 (Docker)
ORM	Prisma
UI	shadcn/ui + Tailwind
LLM	Anthropic Claude (default, required for tool-use) · OpenAI · Ollama
Transcription	OpenAI Whisper
Scheduling	`node-cron` (wired in `instrumentation.ts`)
Tests	Vitest
Search	Pluggable `WebSearchProvider` (DuckDuckGo default)

Where to look

Concern	File
Schema (all entities, all relations)	`prisma/schema.prisma`
Per-project assistant home	`src/app/projects/[projectId]/page.tsx`
Chat component (reusable across surfaces)	`src/components/assistant-chat.tsx`
Assistant API route	`src/app/api/assistant/route.ts`
Persona registry	`src/lib/personas.ts`
System prompt builder	`src/lib/lead-expert.ts`
Tool registry (create_note / create_task / record_metric / save_resource)	`src/lib/assistant-tools.ts`
Anthropic tool-use loop	`src/lib/llm-tool-loop.ts`
Resource retrieval	`src/lib/resources/retrieve.ts`
Wiki retrieval	`src/lib/wiki/retrieve.ts`
Wiki versioned store + hash dedup	`src/lib/wiki/store.ts`
Wiki compile orchestrator	`src/lib/wiki/compile.ts`
Imputation pipeline + apply	`src/lib/wiki/lint/imputation.ts`
Scheduler + cron jobs	`src/lib/agents/scheduler.ts`
Agent runtime (Zod + retry)	`src/lib/agents/runtime.ts`
News ingest agent	`src/lib/agents/news-ingest.ts`
Project feed	`src/components/project-feed.tsx`
Lead scoring (pure function)	`src/lib/scoring.ts`
LLM provider abstraction	`src/lib/llm.ts`
OG fetch helper	`src/lib/og-fetch.ts`

License

Private project.

Name		Name	Last commit message	Last commit date
Latest commit History 32 Commits
prisma		prisma
public		public
scripts		scripts
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
components.json		components.json
docker-compose.yml		docker-compose.yml
eslint.config.mjs		eslint.config.mjs
instrumentation.ts		instrumentation.ts
next.config.ts		next.config.ts
package-lock.json		package-lock.json
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json
vitest.config.ts		vitest.config.ts

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Forge

What's in here

Highlights

Architecture: the wiki AI pipeline

Architecture: the assistant turn

Setup

Tech stack

Where to look

License

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

Forge

What's in here

Highlights

Architecture: the wiki AI pipeline

Architecture: the assistant turn

Setup

Tech stack

Where to look

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages