Skip to content
@basidiocarp

basidiocarp

Basidiocarp

Tools that make AI coding agents better at their job. Memory that persists across sessions, token compression that cuts costs 60-90%, code intelligence across 32 languages, and feedback capture that turns mistakes into lessons.

Everything runs locally. No cloud services, no API keys for the core stack. SQLite all the way down.

Install

curl -fsSL https://raw.githubusercontent.com/basidiocarp/.github/main/install.sh | sh
stipe init

Two commands. Stipe downloads binaries, finds your editor (Claude Code, Cursor, Windsurf, Continue, Claude Desktop), registers the MCP servers, and wires up hooks. stipe doctor confirms everything landed.

Projects

Project What it does
Mycelium CLI proxy. Rewrites command output before it reaches the agent. 70+ filters, 60-90% token savings. Docs
Hyphae Agent memory. Episodic recall, knowledge graphs, RAG with hybrid search, training data export. 39 MCP tools. Docs
Rhizome Code intelligence. Tree-sitter + LSP, symbol extraction, file editing, code graphs. 37 MCP tools, 32 languages. Docs
Cap Web dashboard. Browse memories, view token analytics, explore code with annotations. Docs
Spore Shared Rust library. Discovery, JSON-RPC, config, self-update, platform paths.
Stipe Ecosystem manager. Install, init, doctor, update.
Cortina Hook runner. Captures errors, corrections, code changes, session summaries in Rust.
Lamella Plugin system. 230 skills and 175 agents for Claude Code. Docs

Guides

Guide Covers
Integration How the projects connect, protocols, failure modes
AI Concepts RAG, DPO, fine-tuning, self-hosting, Bedrock comparison
LLM Training Data export, Axolotl, Together.ai, Ollama serving
Training Data Formats, volume estimates, SQL export

Architecture

graph TD
    Agent["AI Coding Agent"]
    Mycelium["Mycelium\nToken Compression"]
    Hyphae["Hyphae\nMemory + RAG"]
    Rhizome["Rhizome\nCode Intelligence"]
    Cap["Cap\nDashboard"]
    Cortina["Cortina\nHook Runner"]
    Stipe["Stipe\nEcosystem Manager"]
    Spore["Spore\nShared Infrastructure"]

    Agent -- "commands" --> Mycelium
    Mycelium -- "60-90% fewer tokens" --> Agent
    Agent -- "MCP (35 tools)" --> Hyphae
    Hyphae -- "memories, context, lessons" --> Agent
    Agent -- "MCP (37 tools)" --> Rhizome
    Rhizome -- "symbols, edits, analysis" --> Agent
    Cortina -- "hooks" --> Agent
    Cortina -- "errors, corrections, changes" --> Hyphae
    Rhizome -- "code graphs" --> Hyphae
    Mycelium -- "large outputs" --> Hyphae
    Cap -- "reads" --> Hyphae
    Cap -- "queries" --> Rhizome
    Stipe -. "installs + configures" .-> Mycelium
    Stipe -. "installs + configures" .-> Hyphae
    Stipe -. "installs + configures" .-> Rhizome
    Spore -. "shared lib" .-> Mycelium
    Spore -. "shared lib" .-> Rhizome
    Spore -. "shared lib" .-> Hyphae

    style Hyphae fill:#36b37e,stroke:#1f8a5a,color:#fff
    style Mycelium fill:#ff7452,stroke:#de350b,color:#fff
    style Rhizome fill:#6554c0,stroke:#403294,color:#fff
    style Cap fill:#4c9aff,stroke:#2571cc,color:#fff
    style Cortina fill:#00b8d9,stroke:#0095b3,color:#fff
    style Stipe fill:#97a0af,stroke:#6b778c,color:#fff
    style Spore fill:#ffab00,stroke:#ff8b00,color:#fff
    style Agent fill:#505f79,stroke:#344563,color:#fff
Loading

How It Works

Hyphae stores two kinds of data in one SQLite database. Memories are episodic: they decay over time based on importance and access frequency. Memoirs are permanent knowledge graphs with typed relations between concepts. Search blends FTS5 full-text (30% weight) with cosine vector similarity (70%) using fastembed locally or any OpenAI-compatible endpoint.

effective_decay = base_rate * importance_multiplier / (1 + access_count * 0.1)

Critical memories never decay. Frequently accessed ones slow down. The decay runs automatically on every recall.

Rhizome parses code with tree-sitter (18 languages, 10 with dedicated queries) and fills gaps with LSP servers (32 languages, 20+ auto-installed). The BackendSelector picks the right backend per tool call: tree-sitter for symbol extraction, LSP for references and renames. It also exports code structure graphs into Hyphae memoirs so the agent can query "what calls this function?" from memory.

Mycelium sits between the agent and the shell. git log -20 returns 5 lines instead of 200. cargo test with 500 passing tests returns only the 2 failures. Small outputs pass through untouched; medium ones get filtered; large ones get chunked into Hyphae for later retrieval. 70+ filters cover git, cargo, npm, docker, kubectl, and more.

Cortina runs as a Claude Code hook. After every tool use, it checks for errors, self-corrections, and code changes. On session end, it writes a summary to Hyphae and exports the code graph to Rhizome. Over time, extract_lessons surfaces patterns from these signals, and evaluate measures whether the agent is getting better. The accumulated data exports as SFT/DPO pairs for fine-tuning via Ollama. See the AI Concepts and Training guides.

flowchart LR
    Ingest["Ingest"] --> Chunk["Chunk"] --> Embed["Embed"] --> Store["Store"] --> Search["Search"] --> Inject["Inject"] --> Response["Response"]
    style Ingest fill:#36b37e,stroke:#1f8a5a,color:#fff
    style Search fill:#6554c0,stroke:#403294,color:#fff
Loading

Session Flow

sequenceDiagram
    participant A as Agent
    participant C as Cortina
    participant M as Mycelium
    participant H as Hyphae
    participant R as Rhizome

    A->>H: initialize
    H-->>A: auto-recalled context
    A->>M: git log -20
    M-->>A: 5 lines (90% saved)
    A->>R: get_symbols + edit
    C->>H: track edits, capture errors
    A->>H: extract_lessons
    H-->>A: patterns from past mistakes
    A->>H: session_end
    C->>R: export code graph
Loading

Popular repositories Loading

  1. mycelium mycelium Public

    Token-optimized CLI proxy for AI coding agents. Filters and compresses command output for 60-90% savings. Rust.

    Rust

  2. hyphae hyphae Public

    Persistent memory system for AI coding agents. Episodic memories + semantic knowledge graphs. MCP server + CLI. Rust.

    Rust

  3. cap cap Public

    Web dashboard for the Mycelium ecosystem. Memory browser + token savings analytics. React, Mantine, Hono.

    TypeScript

  4. rhizome rhizome Public

    Code intelligence MCP server. Symbol extraction and navigation via tree-sitter and LSP backends. Rust.

    Rust

  5. .github .github Public

    Organization profile and community health files

    Shell

  6. spore spore Public

    Shared IPC for the Basidiocarp ecosystem. Tool discovery, JSON-RPC 2.0, subprocess MCP communication. Rust.

    Rust

Repositories

Showing 8 of 8 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…