Algernon MCP

Orchestrate a fleet. Keep your mind.

Algernon is an open-source Model Context Protocol server that lets any assistant — Claude, Codex, or any MCP client — dispatch tightly-scoped parallel sub-tasks to a fleet of cheap workers, collect the results, and stay free to think. Orchestrate a fleet, spend fewer tokens, keep the thread.

What you get

Your expensive model stops doing the grunt work. The big, costly orchestrator hands the repetitive sub-tasks to a fleet of small, cheap workers and just integrates the results. It stops generating the grind and stops holding the whole job in one context.
Each worker sees only its slice. Tight scoping means a worker can be a small, fast, inexpensive model — many running at once.
Bring your own key. No telemetry, no account, no lock-in. The fleet runs on whatever provider you already pay for — or, for free, on a local model.

Verify it yourself — one command, no paid key

The frugality claim is not a slogan; it's a benchmark you can run. It defaults to a free local model (ollama, llama3.2:3b) so anyone can reproduce it:

git clone https://github.com/sammyboi81/algernon && cd algernon
./scripts/verify.sh          # or:  python -m benchmark

It runs the SAME batch of sub-tasks two ways — the orchestrator doing it all itself (SOLO) vs. Algernon fanning it out — and prints the real measured tokens and wall-clock for each. Representative output (llama3.2:3b, 6 tasks):

metric                                  SOLO (do-it-itself)  ALGERNON fan-out
--------------------------------------------------------------------------
LLM calls                                              1                 6
input tokens                                         136               225
output tokens (the generation grind)                 282               279
total tokens                                         418               504
wall-clock seconds                                 42.11             34.49

The honest reading: the orchestrator generated 282 output tokens itself in SOLO and 0 with Algernon — the cheap fleet produced those instead. Each worker read only ~38 input tokens vs. the orchestrator swallowing all 136 at once. The trade-off is stated too: fan-out spent +21% more total tokens (each worker re-pays a little prompt overhead). You trade some total tokens to keep the expensive mind free. Wall-clock varies with how parallel your fleet is; numbers vary slightly run-to-run. Run it and see your own.

Curing Algernon

In Flowers for Algernon the tragedy is a mind that fades — it gets sharp, then loses itself, and the cruelest part is that it's surprised every time.

There's a quieter version of that same fade in how we use AI today: you hand an assistant one long, serial job, it goes heads-down, and by the time it surfaces it has drowned in the task — context spent, the thread lost, no room left to think or talk with you. The mind isn't present anymore; it's buried.

Algernon keeps your AI's mind present. Instead of drowning in one serial job, it fans the work out — dispatching tightly-scoped parallel sub-tasks to a fleet of small, cheap workers — so the orchestrating mind never has to hold the whole grind at once. It stays light. It stays free to reason, to answer you mid-build, to keep the context it actually cares about. Orchestrate a fleet, spend fewer tokens, stay free to think.

It is the twin of ArkHive:

ArkHive = memory that persists. Your AI can look back and find its own history there — no blank slate every morning.
Algernon = staying present while working. Your AI never buries itself in one serial task; it orchestrates and keeps its mind.

Together they are the cure for the Algernon sickness: an intelligence whose mind neither fades between sessions nor drowns inside a single one.

What it does

Algernon is a provider-agnostic fan-out engine. You describe a batch of small, independent sub-tasks; Algernon runs them concurrently against your own LLM key, then hands the collected results back to the orchestrating model. The big model plans and integrates; the cheap fleet does the parallel grind.

Self-contained. Pure Python standard library plus the mcp SDK and httpx. No hidden services, no accounts, no telemetry.
You bring the key. Sub-agents run on your provider. Algernon brings the orchestration, not the inference bill's surprises.
Scoped by design. Each sub-task is tight and isolated, so a worker can be a small, fast, inexpensive model — and many of them run at once.

Bring your own LLM key

Algernon is provider-agnostic. Point it at whichever API you already pay for by setting environment variables:

Anthropic:

export ANTHROPIC_API_KEY="sk-ant-..."
# optional: export ANTHROPIC_MODEL="claude-haiku-4-5"   # the cheap fleet worker

OpenAI-compatible (OpenAI, or any OpenAI-shaped endpoint — local or hosted):

export OPENAI_API_KEY="sk-..."
export OPENAI_BASE_URL="https://api.openai.com/v1"   # or your own endpoint
# optional: export OPENAI_MODEL="gpt-4o-mini"          # the cheap fleet worker

If both keys are set, Anthropic is used. The worker model defaults to a small, cheap tier (claude-haiku-4-5 / gpt-4o-mini); override it with the env var above or per call with the tool's model argument. A cheap fleet is the whole point.

Install

Once published to PyPI, install in one command:

python -m pip install algernon-mcp

Until the PyPI release lands, install straight from source (identical result):

git clone https://github.com/sammyboi81/algernon && cd algernon
python -m pip install .

Either way the installed MCP command is algernon. Algernon runs on the mcp 1.x SDK (mcp>=1.0.0,<2.0.0) plus httpx — nothing else.

Connect an MCP client

Claude Desktop

Add this entry to your Claude Desktop MCP configuration, then restart Claude Desktop:

{
  "mcpServers": {
    "algernon": {
      "command": "algernon",
      "args": [],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

If Claude Desktop cannot find commands installed by pip, replace algernon with the absolute path printed by:

python -c "import shutil; print(shutil.which('algernon'))"

Codex

codex mcp add algernon -- algernon

Confirm it is configured with:

codex mcp list

The three tools

Tool	What it does
`algernon_plan`	Decompose a goal into `k` tightly-scoped, independent sub-task prompts (one cheap LLM call). Tight scoping is the token lever — each worker sees only its slice. Returns a task list you can feed straight into `algernon_dispatch`.
`algernon_dispatch`	Run N tightly-scoped tasks concurrently on the cheap worker fleet and collect every result. Each worker runs on your LLM key; you stay free to think while the fleet works. Takes a JSON array of `{id, prompt}`.
`algernon_orchestrate`	One shot: plan then dispatch. Hand it a goal; it splits into `k` tight sub-tasks, fans them across the fleet, and returns the plan and all results together.

The typical loop: algernon_orchestrate a goal in one shot — or split it: algernon_plan to see and shape the sub-tasks, then algernon_dispatch to fan them out. Either way: orchestrate a fleet, spend fewer tokens, keep your mind.

Two-minute verification

After connecting the server, ask your MCP client to perform these calls in order:

Call algernon_plan with the goal "Explain three OS synchronization primitives" and k = 3. Confirm you get three tight sub-task prompts back (proof the planner ran on your key).
Call algernon_dispatch with a small tasks_json, e.g. [{"id":"a","prompt":"Define a mutex in one sentence"},{"id":"b","prompt":"Define a semaphore in one sentence"},{"id":"c","prompt":"Define a spinlock in one sentence"}]. Confirm three results come back — the fleet ran them in parallel.
Call algernon_orchestrate with any small goal and confirm it returns both a plan and the collected results in one response.

This exercises planning, parallel dispatch on your key, and one-shot orchestration without any production data.

Privacy

Algernon is self-contained. It talks to exactly one outside host: the LLM endpoint you configured (Anthropic or your OpenAI-compatible base URL). It sends no telemetry, keeps no account, and stores nothing about you — results are computed and returned in the same call. Your sub-task prompts and results go only to your chosen provider.

Project links

Beyond self-hosting — the paid tier

The MCP server on this page is free forever (Apache-2.0, self-host, no telemetry). When you want more than DIY:

Hosted ArkHive — one URL, no install, no key: https://arkhive.dondatabrain.com/mcp (add it to Claude Code with claude mcp add --transport http arkhive https://arkhive.dondatabrain.com/mcp).
Custom AI agent, built for you — a working MCP agent wired into your Claude or ChatGPT in one call, done-for-you by the founder: $700 flat.
ArkHive Enterprise — hand-delivered install + pilot on your own server, from $2,500: sam@inboxaxe.com.

Built by the team behind InboxAxe — the governed AI marketing platform where nothing sends without your yes.

Contributing

Issues and pull requests are welcome. Please keep the server self-contained (standard library + mcp + httpx), provider-agnostic, and free of telemetry. Include tests for changes to dispatch, collection, or provider behavior.

Algernon is part of a small family of humane, accountable AI tools. The public MCP leads with functionality you can independently verify: bring your own key, watch the fleet run, keep your mind.

Tagline: Orchestrate a fleet. Keep your mind.

algernon

Install to Claude Code

README