Tessera

Context & Memory Operating System for AI coding agents. Tessera captures knowledge from code, Git, docs, PRs, issues, chats, and IDE activity, then compiles task-aware context packages — not naive RAG dumps — using memory, a project knowledge graph, and hybrid retrieval. One platform runs Local, Self-Hosted, Managed Cloud, and Enterprise behind identical API + MCP interfaces.

A tessera is a single tile in a mosaic. Tessera assembles thousands of scattered fragments of project knowledge into one coherent picture, on demand, per task.

Why Tessera

AI coding agents fail less from weak models and more from bad context: stale snippets, missing decisions, no awareness that changing one function breaks three others. Tessera treats context as a compiled artifact with provenance, ranking, deduplication, and compression — and tracks effect-links so an agent learns that editing A requires touching B and C.

The two differentiators we build around:

1. Context Compiler — plan → retrieve → expand → rank → dedup → compress → assemble, every package explainable with provenance.
2. Effect-links — a first-class graph of "change here implies change there," surfaced to agents so they stop fixing one place and breaking others.

Quickstart

From source. Nothing is on npm yet — the release pipeline exists but has never been dispatched (see Releasing). Until then, run it from a clone.

git clone https://github.com/TesseraOS/Tessera.git && cd Tessera && pnpm install && pnpm build

Check the toolchain, config, storage and embeddings are healthy:

pnpm tessera doctor

Scaffold a Local deployment — SQLite, local embeddings, no external services and no API keys:

pnpm tessera init

Index a repository:

pnpm tessera source add --kind git --root /path/to/your/repo

Point an agent at it. mcp-config emits ready-to-paste configuration for Claude Code, Cursor, Cline, Codex and Continue:

pnpm tessera mcp-config --client claude-code

Or serve the REST API:

pnpm tessera serve

The full command set is init, serve, mcp, source, token, doctor, mcp-config and skills — run pnpm tessera --help for details. (pnpm tessera is a root alias for the built CLI; once the package is published this is just tessera.)

What's here today

The engine is built and verified: ingestion connectors (filesystem/Git/GitHub), versioned memory, the knowledge graph with effect-links, hybrid retrieval (five signals plus rank fusion), the Context Compiler, REST /v1 and MCP surfaces (stdio and remote HTTP), a generated TypeScript SDK, auth/RBAC with tenant and project isolation, an audit trail, usage metering and a billing port, Postgres/pgvector and S3 adapters for self-hosting, feature flags, a plugin host, and observability.

Three web surfaces ship alongside it: the dashboard (apps/web), the marketing site (apps/marketing) and the documentation site (apps/docs, with generated API and env references that are drift-gated against the code).

What remains is R4 launch work — deployment artifacts, notifications, i18n readiness, and the operator/legal finalization — tracked as open features in .harness/state/feature_list.json.

Everything is built one feature at a time under the in-repo agent harness: plan, implement, verify against gates, trace effects, record. See AGENTS.md.

• docs/PRD.md — product requirements.
• docs/architecture/ARCHITECTURE.md — system design.
• docs/adr/ — Architecture Decision Records (the "why").
• docs/glossary.md, docs/roadmap.md.

Layout (monorepo)

tessera/
├── apps/
│   ├── api/        # @tessera/api    — Fastify REST /v1 (modular monolith surface)
│   ├── mcp/        # @tessera/mcp    — MCP server (same services as REST; stdio + HTTP)
│   ├── server/     # @tessera/server — runnable bins (tessera-api / tessera-mcp / tessera-token)
│   ├── cli/        # @tessera/cli    — `tessera` one-command onboarding
│   ├── web/        # @tessera/web    — Next.js dashboard (app subdomain, ADR-0035)
│   ├── marketing/  # @tessera/marketing — public marketing site (apex domain; Terra Mosaic)
│   └── docs/       # @tessera/docs   — documentation site (Fumadocs, generated references)
├── packages/       # @tessera/*      — core, storage, ai, ingestion, memory, knowledge-graph,
│                   #                   retrieval, context-compiler, config, observability,
│                   #                   plugin-host, billing, sdk, skills, brand, mascot
├── tests/          # cross-cutting suites — bench, e2e-full, web-perf
├── docs/           # PRD, architecture, ADRs, design system
├── .harness/       # tool-agnostic agent harness (system of record)
└── .claude/        # Claude Code adapter for the harness

Development requirements

Verification gates (run in order; green is the only acceptable state):

node scripts/verify-state.mjs && pnpm -w typecheck && pnpm -w lint && pnpm -w format:check && pnpm -w test && pnpm -w build

Deployment modes (one architecture, configuration-selected)

Deployment is configuration, not a fork — see the ports & adapters design in docs/architecture/ARCHITECTURE.md.

Releasing

Releases are dispatched manually and do not publish by default: the workflow builds, packs and generates an SBOM on every run, and reaches npm publish only when explicitly asked to. The procedure and the supply-chain controls are in .harness/governance/release-checklist.md; the reasoning is in ADR-0062.

Contributing

Read AGENTS.md first — it is the operating manual for humans and agents alike, and it is binding. In short: one feature at a time, plan before code, verification is the proof rather than an assertion, and every deviation from a documented default gets an ADR.

Licensed under the Apache License 2.0. Third-party attributions are in NOTICE.md. Tessera is a project name and brand — see ADR-0008; the licence grants no trademark rights (Apache-2.0 §6).