agora

A terminal hub for developers and the agentic AI ecosystem — discover MCP servers and AI tools, skills and harnesses, follow the news, join the conversation, access the latest tech.

agora is a standalone CLI that puts everything a developer building for the future cares about in one terminal: a curated + live marketplace of MCP servers and agent tooling, a threaded community with reputation-weighted sort, a ranked news feed with on-cache AI summarization, and first-class install consent for declared permissions. Works offline by default; opt into the backend with agora auth login.

Install

# fastest path — runs in your current project
npx opencode-agora init

# or install once, use anywhere
npm i -g opencode-agora
agora welcome

From source (requires bun):

git clone https://github.com/IrgenSlj/agora.git
cd agora && bun install && bun run build && bun link

You can also compile a self-contained standalone binary (no Node or bun runtime required at run-time):

bun run build:binary   # produces dist/agora

Experimental. The compile works, but distributing the binary needs code signing (unsigned arm64 macOS binaries are killed on launch) and, for wider reach, notarization + a Homebrew tap. That packaging is tracked for an upcoming distribution cut; for now npm/npx remains the supported path.

The binary is published as both agora and opencode-agora.

Try it

agora welcome                          # guided tour, adapts when signed in
agora today                            # last 24h: news + community + trending
agora search mcp                       # 67 curated + live GitHub/HF results
agora install mcp-github --write --yes # install + write opencode.json
agora acquire mcp-postgres --dry-run   # preview scan-gated acquisition
agora doctor --probe                   # health check + description-drift detection
agora tui                              # 5-page full-screen interface

The default agora in a TTY drops you into a persistent shell that mixes bash dispatch (ls, git status) with free LLM chat (why is this slow?). It remembers you across sessions: /recall <query> searches your past conversations and /sessions lists them. /help lists the slash commands; /abc shows the single-letter shortcuts.

Commands

Run agora help for the grouped list, or agora help <command> for any of these:

Every command supports --json for scripting and --help for inline manual.

Install consent & safe capability acquisition

agora install <id> is preview-only by default. With --write, items that declare a permissions manifest require an explicit --yes:

$ agora install mcp-filesystem --write
Permissions
  fs    ./**/*

This package declares permissions. Re-run with --yes to grant and install.
$ echo $?
1

The TUI install preview flips its footer to g grant + install d details n cancel when permissions are present. The list shows a dim [fs net exec] badge on any item with a non-empty manifest.

--write also runs a pre-install scan first — repo reachability, npm existence, recency, declared-permission consistency, description-injection patterns, and community flag count — and refuses to apply if any check fails (e.g. an item flagged enough times to auto-hide, or a package that 404s on npm). Run the scan standalone with agora scan <id>, or bypass the install gate with --skip-scan.

agora acquire <id|query> composes the full pipeline — resolve by id or capability query, create an install plan, run the scan gate, and write config — into one agent-callable action. The scan gate enforces three outcomes:

• fail blocks the write (non-zero exit). Run agora scan <id> for details.
• warn without --accept-warnings does not write. Re-run with --accept-warnings to proceed.
• dry-run previews everything without writing.

For agents: the acquire MCP tool (via agora mcp) provides structured output the model can act on. The agora_acquire plugin tool is preview-only (dry-run); write-to-config requires the CLI or MCP tool.

The plugin also offers opt-in capability-gap detection (suggestAcquire): when the agent reaches for a tool the user lacks, it surfaces a non-intrusive agora_acquire suggestion. Stack memory (stackMemory, on by default) injects the current MCP stack + capabilities into compaction context so the agent remembers its tools across sessions.

Agent stack manager & capability acquisition

Beyond discovery, agora manages the MCP servers your agent actually uses — across opencode, Claude Code, Cursor, and Windsurf — from one place. Think package.json / Brewfile for your agent stack.

agora installed                 # every configured MCP server across all your tools, grouped
agora doctor                    # health: command resolvable? conflicting definitions?
agora doctor --probe            # probe + description-drift detection (canonical tool-schema hashing)
agora install mcp-github --write --save   # install AND record it in agora.toml
agora acquire mcp-postgres      # resolve → scan-gate → write config (safe agent-callable gateway)
agora acquire "postgres database" --dry-run   # resolve by capability query, preview only
agora acquire mcp-github --save --accept-warnings  # accept scan warnings and record in agora.toml
agora freeze --write            # snapshot your whole stack into agora.toml
agora sync                      # dry-run: what would change to match agora.toml
agora sync --write --yes        # apply it (preserves every unrelated config key)
agora sync --from <url|path>    # apply a shared manifest — clone someone's setup
agora try mcp-filesystem        # ephemeral test-drive: handshake, list tools, discard
agora capabilities "query a database"   # which of my servers can do X?

agora.toml is a portable, declarative manifest — commit it to a repo so anyone can reproduce your agent setup with agora sync --from <url>. sync is dry-run by default and never touches config keys it doesn't own. doctor --probe, try, and capabilities share a local cache of each server's discovered tool schemas — the local foundation for capability search. doctor --probe also computes a canonical descriptionDigest (SHA-256 of tool names + descriptions + input schemas) and detects DRIFT when re-probing shows a different digest.

Configuration

Per-user state lives under ~/.config/agora/ — state.json (saves + auth), settings.toml (preferences), news-cache.jsonl, news-meta.json (bookmarks), hubs-cache.jsonl. All files holding user data are written 0o600 and atomically (.tmp + rename) so a crash mid-flush leaves the previous version intact.

OpenCode plugin

agora init registers the package as an OpenCode plugin, drops a .opencode/command/agora.md slash command, and enables:

• 12 explicit named tools: agora_search, agora_today, agora_browse, agora_browse_category, agora_install, agora_scan, agora_acquire, agora_trending, agora_tutorial, agora_chat, agora_config, agora_news, agora_info
• Lifecycle hooks: opt-in capability-gap detection (tool.execute.before) and stack memory injection on compaction (experimental.session.compacting)
• SDK-preferring chat: uses client.session.prompt() when available, falls back to CLI spawn
• Windows-compatible binary resolution: proper PATHEXT-aware lookup and .cmd/.bat spawning

agora_* tools are also reachable from any MCP client via agora mcp — add {"mcp": {"agora": {"command": ["agora", "mcp"]}}} to your opencode.json.

Architecture

src/cli/              command handlers, dispatch, shell, prompter, TUI
src/cli/pages/        five full-screen TUI pages (home, marketplace,
                      community, news, settings) + shared helpers
src/marketplace.ts    curated catalog + live hub merge + install planner
src/search/           offline BM25 catalog-search index
src/acquire.ts        capability-acquisition gateway (match → scan-gate → write)
src/opencode-exec.ts  unified opencode binary resolver (win32-safe) + spawn helper
src/curator/          AI curator (GitHub/HF discovery + verify, scheduled-safe)
src/stack/            cross-tool agent stack manager (opencode/claude/cursor/…)
src/stack/capability-cache.ts  tool-schema cache + description-drift detection
src/plugin/           OpenCode plugin (tools, hooks, SDK chat transport)
src/hubs/             GitHub + HuggingFace connectors + AI enrichment
src/community/        backend client + types
src/news/             scoring, cache, per-source adapters
src/state.ts          local state, saves, auth (atomic 0o600 writes)
src/scan.ts           pre-install safety scan (repo, npm, description-injection)
backend/src/index.ts  Cloudflare Workers + D1 (Hono router)
test/                 1155 tests, 51 files

agora is, at its core, a marketplace + community hub for agentic coding. On top of that it's grown a daily-driver layer — a cross-tool agent stack manager (agora installed / doctor, then agora.toml + sync) and a safe capability-acquisition gateway (agora acquire — resolve by id/query → scan gate → write config). The OpenCode plugin deepens integration with lifecycle hooks, SDK-preferring chat, and 12 explicit marketplace tools. See the roadmap for the plan.

See docs/ARCHITECTURE.md for the why-this-shape writeup and ROADMAP.md for what's next.

Development

bun test            # 805 cases, ~3.5s
bun run typecheck   # CLI + backend (typecheck:cli / typecheck:backend run both)
bun run build       # tsc + chmod +x dist/cli.js
bun src/cli.ts <cmd>  # run from source without building

cd backend && bun run dev          # local backend on wrangler
cd backend && bun run typecheck    # backend has its own tsconfig

PRs welcome — see CONTRIBUTING.md and COMMUNITY_GUIDELINES.md. The catalog accepts entries via PR; see src/data.ts for the shape.

A scripts/demo.tape is included for regenerating the README hero gif with VHS (brew install vhs && vhs scripts/demo.tape → docs/demo.gif).

License

MIT — © IrgenSlj.