Skip to main content
Glama

OpenCode Project Knowledge

Universal, one-time-installable project understanding for OpenCode and Antigravity IDE/CLI. Install once, open any repository, and get efficient project context automatically — no per-project setup. The engine is the same on both platforms; thin adapters expose the same project_context as an OpenCode plugin and an Antigravity MCP server + hook.

What it does

  • Detects the project type (Python, JS/TS, Vue, React, Electron, C#/.NET, Java, C/C++, PHP, Go, Rust, Android, mixed, or unknown→generic).

  • Bootstraps a concise .project/ knowledge baseline safely when missing.

  • Injects a compact (~35-line) PROJECT CONTEXT at task start — never the whole knowledge base, never large source excerpts.

  • Refreshes incrementally: file change → affected modules → stale markers → targeted doc updates → baseline update.

  • Works with local filesystem + Git only. On OpenCode: no MCP required (plugin + instructions); on Antigravity: same engine via MCP project_context

    • PreInvocation hook + skills (open Agent Skills spec).

Related MCP server: @mempalace/mcp-dev

Install (Windows, one click)

  1. Get this repository onto your machine (Clone it, or Download ZIP from GitHub and extract it anywhere).

  2. Double-click Install.bat.

  3. Done. No admin rights needed — it installs for your Windows user only, and the window stays open so you can see the result.

To remove it later, double-click Uninstall.bat (your other OpenCode settings are left untouched).

Advanced (terminal — does exactly the same thing):

powershell -ExecutionPolicy Bypass -File install.ps1
powershell -ExecutionPolicy Bypass -File install.ps1 -DryRun   # preview only

One-liner (no clone needed):

irm https://raw.githubusercontent.com/usm007/opencode-project-knowledge/main/install.ps1 | iex

Linux/macOS:

./install.sh

Uninstall (terminal — same as Uninstall.bat):

powershell -ExecutionPolicy Bypass -File uninstall.ps1
./uninstall.sh

How automatic operation works

  1. Open any repo in OpenCode or Antigravity.

  2. Global AGENTS.md + project-knowledge-auto skill fire before substantial work (Antigravity also auto-injects via its PreInvocation hook and exposes project_context via MCP).

  3. .project/ present → context.js builds PROJECT CONTEXT from overview.md + knowledge.json + state/stale.json; only the touched subsystem's detail docs are read via the modules.md map.

  4. .project/ absent → bootstrap.mjs creates it safely (detects type, respects .gitignore, skips build output/secrets, never touches source, never commits, creates root AGENTS.md only if missing).

  5. Edits → refresh.js marks affected modules stale; agent updates only those docs; status.js reports health read-only.

Debug directly:

node project-knowledge/detect.js <repo> --json
node project-knowledge/context.js <repo> "one-line task"
node project-knowledge/status.js <repo>
node project-knowledge/refresh.js <repo>
node project-knowledge/bootstrap.mjs <repo>

Supported project types

Python · JavaScript · TypeScript · Vue · React (incl. Next/Angular hints) · Electron · C#/.NET · Java · C · C++ · PHP · Go · Rust · Android · Ruby · Dart · Swift · Docker-aware · mixed-language · unknown → generic structural analysis.

Global vs project-local

Global (installed once, ~/.config/opencode/ + ~/.gemini/ + ~/.agents/)

Project-local (per repo, .project/)

AGENTS.md section, skill, plugin (OpenCode) / MCP + hook + plugin (Antigravity), project-knowledge/*.js, templates, commands

overview.md, architecture.md, modules.md, data-flow.md, dependencies.md, conventions.md, decisions.md, known-issues.md, test-map.md, knowledge.json, state/stale.json

Limitations

  • Heuristic detection (markers + extension census) can misread exotic layouts — detect.js --json shows evidence; source always wins.

  • Bootstrap stubs contain (agent) placeholders until completed with real reads.

  • Git-absent repos lose change tracking (filesystem fallback only).

  • The plugin is optional/defensive; CLI + instructions are the guaranteed path. It ships as ESM per OpenCode's documented plugin format and installs its @opencode-ai/plugin dependency automatically at first OpenCode startup.

Security

Never records passwords, API keys, tokens, certificates, or env secret values. Respects .gitignore + info/exclude; skips node_modules/bin/obj/dist/build/ target/vendor, caches, binaries, and secret-like paths. See docs/.

Safety guarantees

  • --force bootstrap backs up overwritten docs to .project/.backup/<timestamp>/.

  • Your own keys in knowledge.json survive regeneration; pre-existing payload files are backed up (.bak) on first install and restored on uninstall.

  • Installers refuse silent downgrades, support --dry-run/-DryRun previews, and verify installed files against SHA256SUMS.

  • npm test (52 unit + E2E) runs on Windows + Linux via CI (--allow-downgrade and POSIX paths covered on Ubuntu).

Contributing

See docs/development.md. Requires Node ≥ 18, no deps. Run npm test before PRs. Bump VERSION on releases.

Layout

├── README.md  LICENSE  VERSION  AGENTS.md  package.json
├── install.ps1  uninstall.ps1  install.sh  uninstall.sh
├── project-knowledge/  skills/  plugins/  commands/  antigravity/
├── tools/  tests/  docs/  examples/

Related MCP Connectors

Related MCP Servers

  • F
    license
    Not graded
    quality
    D
    maintenance
    Intelligently analyzes codebases to enhance LLM prompts with relevant context, featuring adaptive context management and task detection to produce higher quality AI responses.
    2
    -
  • A
    license
    Not graded
    quality
    D
    maintenance
    Enables AI coding assistants to automatically detect and inject team-specific codebase patterns (e.g., error handling, imports, naming conventions) via the Model Context Protocol, ensuring suggestions follow project conventions.
    MIT
  • A
    license
    B
    quality
    A
    maintenance
    Enables AI coding agents to maintain persistent project context, including rules, decisions, environment intelligence, and Git history, using a local-first MCP server with automatic project detection and token-efficient retrieval.
    13
    MIT