Noosphere

Open-source growth intelligence: discover the right market, run outbound, publish inbound content, and turn conversations into calls.

Documentation française · English documentation · Architecture · Self-hosting · Required subscriptions

Noosphere brings the GTM loop into one multi-workspace application:

flowchart LR
  O[Product and offer] --> I[ICP research]
  I --> C[Outbound campaigns]
  O --> P[Inbound content]
  C --> M[LinkedIn, email and WhatsApp conversations]
  P --> M
  M --> R[Qualified calls]
  R --> L[Durable learning]
  L --> I
  L --> P

Loading

The normal experience stays intentionally simple:

1. launch an ICP study;
2. let campaigns and LinkedIn content run within deterministic policies;
3. answer from the unified inbox and collect calls.

The AI never owns provider state. PostgreSQL, durable jobs, leases, idempotency keys and outbox events remain authoritative. Closing a page or drawer stops browser polling only; it does not cancel the work.

Quick start

Requirements: Bun 1.3+, Docker with Compose, and uv for crawler development.

cp .env.example .env
bun install
bun run dev:setup
bun run dev

Open http://localhost:3000.

Self-hosting

The same responsive UI can be hosted privately behind Caddy and used from a phone or desktop browser. Supply any domain or subdomain that points to the VPS; the application hostname does not need to be the future marketing domain.

Noosphere supports three distribution paths without YAML edits:

1. pull official public images from ghcr.io/ignitionai;
2. tag a public or private fork and let its workflow publish images under the lowercase fork owner;
3. set DEPLOY_MODE=local-build to build all application images on the VPS.

GitHub creates a fork's first GHCR packages as private even when the fork is public. The release workflow fails closed and prints the three package landing pages. Sign in to GitHub with an account that administers the packages, open each landing page, click Package settings, then set it to Public once and re-run the workflow. GitHub deliberately returns 404 for a private package page when the browser is signed out or the account lacks access. Later tagged releases keep that visibility.

If GitHub displays Setting is disabled by organization administrators, an organization owner must first open Organization settings -> Packages and enable Package Creation -> Public. The release workflow prints the exact organization settings URL for organization-owned forks.

Create a secure environment interactively:

bash deploy/configure.sh
ENV_FILE=.env bash deploy/doctor.sh
ENV_FILE=.env bash deploy/release.sh

quickstart gives one light workspace HTTPS, daily local backups and a clear warning that VPS loss also loses those backups. production additionally requires encrypted off-VPS Restic storage, monitoring and restore drills. The configurator generates secrets locally, writes .env with mode 0600 and refuses to overwrite it by default.

See the complete self-hosting runbook for fork releases, private GHCR login, updates, exact-digest rollback and backup restoration.

A running UI is not yet an operational prospecting system: AI features need a configured Codex CLI, Kimi Code or OpenAI route, and outreach needs at least one connected LinkedIn, email or WhatsApp provider account.

Verification

bun run check
bun run test:integration

The repository also contains effect-free capacity, shadow, Setter-quality and operator-comprehension gates. Their current evidence and remaining production gates are recorded in the Prospect 360 validation report.

Measured on 23 August 2026: the real-data shadow gate passed on 1,000 IgnitionAI contexts; 100/100 synthetic Codex Setter dry-runs were generated with zero provider effects and resolvable memory receipts. An isolated 2-vCPU/8-GiB VPS remained functional but missed the concurrent-memory p95 target.

Production sizing

• Light usage, one workspace: Netcup RS 2000 G12, with 8 dedicated cores, 16 GiB RAM and 512 GB NVMe. This is the acceptable minimum for a canary or one lightly loaded workspace, provided heavy crawling, indexing and campaigns are not run concurrently.
• Recommended production: Netcup RS 4000 G12, with 12 dedicated cores, 32 GiB RAM and 1 TB NVMe. This is the recommended target for the full platform, including PostgreSQL/ParadeDB, MinIO, Chromium crawling, workers, Qwen embedding and BGE reranking.

TEI keeps Qwen and BGE resident in RAM to avoid cold starts, but this is not continuous compute. Embeddings are generated only for new or changed knowledge, hybrid-search queries, and full reindexing; unchanged content is skipped by hash. Message synchronization, prospect sourcing, post writing, sends and normal Setter execution do not currently use Qwen Embedding. See the localized deployment guides below for the complete workload explanation.

Use x86_64/AMD64 and NVMe storage. A GPU is not required. Shared-vCPU plans are suitable for short preproduction tests but are not the preferred production target. See the French deployment guide, English deployment guide and capacity report for assumptions, upgrade thresholds and operational configuration.

License

Noosphere is licensed under the GNU Affero General Public License v3.0 only. If you modify and operate it over a network, the AGPL requires offering the corresponding source to its users.