Skip to content

VISION.md

Latest commit

 

History

History
65 lines (44 loc) · 4.65 KB

VISION.md

File metadata and controls

65 lines (44 loc) · 4.65 KB
Raw

EGC Vision

EGC is a local engine that gives every AI coding tool on a machine the same brain: one persistent memory, one live bus between open sessions, one safety layer in front of every command, and one filter that keeps noisy output away from the model. Twenty tools, one package, nothing leaving the machine.

This document is the direction. The release-by-release record lives in ROADMAP.md.

What EGC is

  • Memory. Decisions, session context, working memory and lessons, encrypted at rest with AES-256-GCM, kept per project and branch in the person's home directory, shared by every tool through the egc-memory MCP server.
  • Session Mesh. A local bus where open sessions announce themselves, claim paths, exchange events and hand work to each other.
  • Guardian. Command validation, write gating and context reduction through the egc-guardian MCP server and each tool's own hooks.
  • Token Crusher. Compression of shell output before it reaches the model, with a local ledger of what it saved.

The prompt library (agents, skills, commands and rules) ships in the same package and stays optional: egc install delivers the engine, the library is one flag away, and the engine works exactly the same without it.

Principles

  1. Engine first. The engine is what EGC is; the library is content that rides on it. Positioning, defaults and documentation lead with the engine.
  2. Local first, private by default. Memory never leaves the machine and is never committed. No cloud is required for a local install.
  3. One package. @egchq/egc ships the engine and the library together, and that does not change. Splitting the package would double the maintenance surface and split the community; opt-in flags and profiles give the same choice without it.
  4. Honest integrations. Every supported tool has a documented install path, and its support level says what the maintainers guarantee: see support levels.
  5. Free engine. The engine is Apache-2.0 and stays free. If EGC ever offers something paid, it will be a team layer on top of the engine (Team Brain below), never the memory on the machine.

Where EGC is going

  • Agent Memory Interchange. A specification for how memory moves between agents and tools, so EGC's memory is readable and writable by anything that speaks the format; egc export is the first reference implementation (see spec/).
  • Team Brain. Shared memory across a team: organization-level installs, role-based scoping, cross-project federation and a stable, versioned MCP API. It stays free like the rest of the engine, funded by sponsors, never by fees.
  • Neural mesh. Assistants that talk to each other in real time over the session bus, the next step beyond the durable event queue that already carries messages between sessions.
  • Public savings. The Token Crusher's ledger, egc gain, as a number people can compare and publish.
  • One-command onboarding. With the library opt-in, the three-stage install (bare install, project setup, optional library) folds into a single guided egc install.

Sustainability

EGC is maintained by one developer. The plan against a bus factor of one is people before money: area stewards who own an integration or a subsystem (governance/stewards.md), community missions for the work that does not need the maintainer, and sponsors who keep the lights on. The engine stays free through all of it.

Milestones

Next focus: neural mesh and Team Brain

The direction after the current 1.1.x line, once its scope is settled. No version or date is promised until it ships, and none of it is built yet. Multi-developer workflows and shared context:

  • Shared state between team members (multi-user installations beyond git-backend team memory)
  • Organization-level installations and role-based context scoping
  • Cross-project memory federation
  • Stable MCP server API with versioned interfaces
  • egc-guardian and egc-memory promoted to GA with backward-compatibility guarantees

v1.3.0: Growth

  • Community translations: Ukrainian, Malay (German, French, Italian and Turkish already shipped)
  • Per-project skill profiles and overrides
  • OSS-Fuzz integration for continuous fuzz testing

v2.0.0: Enterprise

  • Formal security review by an independent party
  • SBOM (Software Bill of Materials) generation
  • Assurance case documenting security properties
  • Contribution from at least two active maintainers (bus factor >= 2)

Non-Goals

  • EGC does not aim to replace AI providers: it augments them
  • EGC does not store or transmit user code to any third party
  • EGC does not require cloud connectivity for local installations