Cordyceps

Local-first agent harness. A daemon that spawns and drives PTY-based CLI coding agents — Claude Code, Codex, Gemini, Ollama — through a single JSON-RPC 2.0 control plane. Every primitive a human invokes from the shell (spawn, submit, interrupt, approve, kill) is reachable over the same API a manager LLM uses to drive a fleet of peers.

Why this exists

Most agent frameworks assume one model, one task, one process. Real agentic work is a fleet — a manager LLM driving peers, with humans escalated only when needed. The design insight that unlocked it: a manager LLM should be a peer client on the same control plane as a human operator, not a privileged special case. Once spawn / submit / interrupt / approve are reachable over a stable API, the manager gets everything the operator has — and a code review council, an MCP bridge, an autonomous repo grinder, and a deterministic test runner all fall out of the same plumbing.

Cordyceps is that plumbing. The PTY work is a backend detail; the control plane is the product.

Quick start

# Install
npm install -g @ellyseum/cordyceps

# Start the daemon (loopback-only, bearer-auth WebSocket)
cordy daemon start

# Spawn a Claude agent (uses your normal Claude auth — OAuth, memory, hooks)
cordy spawn claude --name smoke

# Send a prompt; print the response
cordy send smoke "respond with just: banana"

# Inspect; clean up
cordy state smoke
cordy kill smoke
cordy daemon stop

For one-shots without a persistent daemon: cordy --ephemeral spawn claude --name once.

Prerequisites

• Node ^20 || ^22 || ^24.
• node-pty is a native module. The npm registry ships prebuilt binaries for common Node ABIs; on a system without a matching prebuild, install needs Python 3 and a C++ toolchain (build-essential on Debian/Ubuntu, Xcode CLT on macOS).
• The CLI agents you want to drive (claude, codex, gemini, ollama) must already be installed on PATH. Cordyceps is a harness — it doesn't ship its own model access.

Install from source

git clone https://github.com/ellyseum/cordyceps
cd cordyceps
pnpm install
pnpm build
./bin/cordy daemon start

The manager-agent pattern

The agentic bit: a "manager" agent is just another client connected to the daemon's JSON-RPC API. Given a prompt, it spawns peer agents, drives them through work, responds to approvals, and tears them down. Gatekeeper, reviewer, orchestrator, human-in-the-loop escalator — these all fall out of one LLM holding the steering wheel of a fleet.

Every primitive is reachable over the same API surface a human uses from the shell, so a manager LLM operating cordy has the same capability as an operator driving it manually. The agent becomes a peer, not a special case.

cordy manager "summarize the git log since last week"

This spawns a Claude session wired with cordy's MCP bridge and a manager system prompt. The manager can spawn other agents (Codex for exec tasks, Gemini for search-heavy work, Ollama for cheap reasoning), coordinate their output, and report back. Press Ctrl+C to interrupt.

Code review council

cordy council review src/core/bus.ts
cordy council review src/core/bus.ts --panel claude,codex,gemini --chair codex --json
cordy council review src/core/bus.ts --no-chair --json   # synthesize yourself
cordy council diff                              # uncommitted changes vs HEAD
cordy council diff --staged                     # only staged changes
cordy council diff main..feature --scope src/   # branch comparison, scoped

N reviewers from different driver families run in parallel, each in a silo (no reviewer sees another's output), then a chair agent synthesizes their findings into a prioritized markdown verdict. The whole point is heterogeneous training lineages — intra-family ensembles correlate their blind spots; inter-family councils don't.

--no-chair (0.5.2+) skips the chair-spawn step and returns per-reviewer findings only. Useful when an LLM caller (e.g. a Claude Code session driving cordy) wants to synthesize the panel's output itself using full conversation context, instead of paying for a separate chair model.

review operates on a whole file; diff reviews the changes shown by git diff (with --no-ext-diff --no-textconv --end-of-options for hostile-repo safety). Tool-capable drivers (Claude PTY, Codex exec, Gemini exec) get a path-only prompt and read the file themselves; tool-less drivers fall back to inline mode with auto-chunking at ~30KB.

Use from Claude Code

The recommended way to drive cordy from a Claude Code session is the claude-cordyceps plugin. It teaches Claude how to map natural-language asks ("consult peers on src/foo.ts", "spawn council", "brainstorm with multiple LLMs", "ask Codex about this design") to the right cordy invocation, and ships a SessionStart hook that detects which drivers are reachable on the current machine.

After installing cordy globally as above:

/plugin marketplace add ellyseum/claude-cordyceps
/plugin install claude-cordyceps

Then trigger via natural language or the /cordy <ask> slash command. See the plugin's README for the full setup and trigger phrase list.

MCP bridge — expose cordy to another agent

Experimental. The bridge currently speaks line-delimited JSON-RPC, which is what Claude Code's MCP loader accepts. Stricter MCP clients may reject the framing; full MCP stdio framing is tracked for a future release.

For lower-level integration (no plugin), add cordy mcp-stdio to a Claude Code MCP config — the agent gains delegated authority to spawn peers and drive them via tool calls (cordy_agents_spawn, cordy_agents_submit, cordy_drivers_list, …):

{
  "mcpServers": {
    "cordy": { "command": "cordy", "args": ["mcp-stdio"] }
  }
}

This is the substrate the claude-cordyceps plugin builds on, and is the path to take if you want raw tool calls instead of the plugin's opinionated skill body.

Drivers

Driver	Family	Modes	Notes
claude-code	Anthropic	pty, exec	Interactive TUI (pty) or one-shot via claude --print (exec). PTY supports approve / reject / mode-switch; exec is cleaner for one-shot prompts and council reviewers.
codex	OpenAI	exec	codex exec --json --skip-git-repo-check
gemini	Google	exec	gemini -p … --output-format stream-json; needs GEMINI_API_KEY in the daemon env
ollama	local	server-http	Streams NDJSON from /api/generate; needs a local Ollama daemon (ollama serve). Free for local models.

Drivers declare their modes; the AgentManager runtime registry maps each mode to a controller plugin without core changes. Adding a new CLI family is a driver (plus a runtime plugin if it speaks an unsupported protocol); nothing in the bus, transport, or existing drivers needs to change.

See docs/DRIVERS.md for the parser/control protocol and how to add one.

Architecture

┌──────────────────────────────────────────────────────────────┐
│                      Daemon Process                          │
│  ┌────────────┐    ┌──────────────┐    ┌──────────────┐      │
│  │  Transport │◄──►│ Service Bus  │◄──►│   Plugins    │      │
│  │  WS /rpc   │    │ (events+kv)  │    │ (audit, …)   │      │
│  └─────┬──────┘    └──────┬───────┘    └──────────────┘      │
│        │                  │                   ▲              │
│        │                  ▼                   │              │
│        │         ┌────────────────┐           │              │
│        │         │ Agent Manager  ├───────────┘              │
│        │         │ + RuntimeReg   │                          │
│        │         └────────┬───────┘                          │
│        │                  │                                  │
│        ▼      ┌───────────┼───────────┐                      │
│  WS clients   ▼           ▼           ▼                      │
│       ┌───────────┐┌────────────┐┌────────────┐              │
│       │  Pty      ││   Exec     ││ ServerHttp │              │
│       │Controller ││ Controller ││ Controller │              │
│       └─────┬─────┘└─────┬──────┘└─────┬──────┘              │
│             ▼            ▼             ▼                     │
│       ┌─────────┐  ┌─────────┐   ┌─────────┐                 │
│       │  PTY    │  │  exec   │   │  HTTP   │                 │
│       │(agent)  │  │(agent)  │   │(server) │                 │
│       └─────────┘  └─────────┘   └─────────┘                 │
└──────────────────────────────────────────────────────────────┘

  HTTP /health        — liveness probe (no auth)
  WS   /rpc           — JSON-RPC 2.0 (Authorization: Bearer <token>)

Two keystone patterns:

• Service Bus (src/core/bus.ts) — pub/sub events + flat key-value state. on() returns an unsubscribe handle so plugins clean up cleanly. Plugins coordinate through agreed bus-key prefixes; nothing imports across plugins.
• Plugin architecture (src/plugins/api.ts) — every extension is a CordycepsPlugin declaring methods, lifecycle hooks, and bus subscriptions. Loaded in priority groups with topological sort. The audit plugin is the reference implementation.

For Mermaid diagrams of the spawn / send / response flow and the manager-agent sequence, see docs/ARCHITECTURE.md. JSON-RPC method reference: docs/PROTOCOL.md.

Notable design decisions

Manager as peer, not special case. The same JSON-RPC surface that powers the cordy CLI also powers manager LLMs talking through cordy mcp-stdio. A manager has zero capabilities beyond what an operator has — which means every shell workflow tested manually maps directly to an agent's behavior.

Glyph-based state machine over regex scraping. PTY output is messy: escape sequences, partial writes, spinner frames over the same cursor position. The Claude parser walks Unicode glyphs (● message, ⎿ result, ⏵/⏸ mode) as state-machine transitions and ships a fixture-replay harness as a regression net. Regex would have to handle every render-time race; a state machine fed by the actual TTY stream doesn't.

Bus-coordinated plugins. Plugins talk through bus.on() / bus.emit() rather than importing each other's types. That keeps the plugin set loosely coupled, lets the loader topo-sort by declared order, and makes hot-load / unload practical without a registry refactor.

What it isn't

• Not a human-facing TUI or web UI — pure agentic infrastructure.
• Not a provider SDK client (drives installed CLIs; no direct Anthropic / OpenAI / Google API calls). The trade-off is that you bring your own auth and pay for whatever the underlying CLI bills you for.
• Not a multi-machine orchestrator — the daemon is loopback-only by design.

Persistence and security

• ~/.cordyceps/ (mode 0700) — daemon state, instance files, optional audit logs (opt-in via --audit).
• ~/.cordyceps/instances/{pid}.json (mode 0600, atomic writes) — the client uses these to discover the running daemon and read its bearer token. The token is regenerated on every daemon start.
• ~/.cordyceps/env (mode 0600) — optional env file auto-loaded at daemon start. Shell env wins over file values. Useful for driver API keys like GEMINI_API_KEY. Override the path via CORDY_ENV_FILE; per-repo override is <repo>/.cordyceps/env.
• Loopback-only transport (127.0.0.1); upgrade rejects non-loopback Host / Origin headers as defense in depth.
• 192-bit bearer token (base64url of 24 random bytes), regenerated per daemon start, presented via Authorization: Bearer <token> on the upgrade request.
• The bearer token is shell-execution-equivalent — anyone who reads it can spawn arbitrary subprocess via cordy's spawn API. Treat it like an SSH key.

Tests

pnpm test          # vitest, full suite green
pnpm build         # tsc clean

The repo uses pnpm; npm test and npm run build work too if you don't have pnpm installed.

When a CLI release shifts glyph spacing, mode-line shape, or spinner frames, the fixture-replay harness in test/fixtures/ catches the drift on the next test run and the fix lives in the relevant driver's directory.

Status

Pre-1.0. Ships a working daemon, four built-in drivers, the council and manager plugins, an MCP stdio bridge (experimental framing), and the plugin / runtime registries. See STATUS.md for what's in progress and what's sketched.

Contributing

Issues and PRs welcome. Read docs/ARCHITECTURE.md first if you're touching anything beyond a typo; touching a driver, docs/DRIVERS.md; touching a plugin, docs/PLUGINS.md.

License

MIT — see LICENSE.

Built by Jocelyn Ellyse.