← Back to portfolio
In development

Anansi

A local agent harness built to make small models genuinely useful — on the laptop you already own, with no subscription and no API bill.

The name

In Akan folklore, Anansi is the spider who won all the world's stories from the sky god — not by being the strongest, but by being the cleverest. That's the bet here: a small model, wrapped in the right harness, can do far more than its size suggests.

Why it exists

Useful AI agents today mostly mean a subscription and someone else's datacenter. Meanwhile most laptops can already run capable 1–8B models locally — they just aren't reliable enough to do real work on their own.

Anansi is an agent harness designed around those limits: strict tool calling, small context windows, and models that forget. The goal is AI that costs nothing to run and works on hardware people already own — and benchmarks that show honestly how far that gets you, from CPU to GPU to NPU.

How it works

prompttool callsvalidateddelegateLocal modelOllama · OpenAI APIAgent loopplan → act → checkToolsfiles · shell · searchMemorypersistent notesContext budgettrim · summarizeSubagentsfresh context each

Roadmap

  1. 01

    Ollama backend

    First target: run any locally pulled model with zero setup beyond Ollama.

    Planned
  2. 02

    OpenAI-compatible API

    One standard interface that also covers LM Studio, llama.cpp, and vLLM.

    Planned
  3. 03

    Constrained tool calling

    Grammar-guided output and strict schemas so 1–8B models call tools reliably.

    Planned
  4. 04

    Core tools

    Read and edit files, run shell commands, and search a codebase.

    Planned
  5. 05

    Subagents

    Hand subtasks to fresh, small contexts and keep only their summaries — small models forget fast.

    Planned
  6. 06

    Persistent memory

    A notes file the agent maintains across steps and sessions.

    Planned
  7. 07

    Context budgeting

    Truncation and summarization tuned for small context windows.

    Planned
  8. 08

    Self-check & retry loop

    Catch malformed or failed tool calls and recover instead of derailing.

    Planned
  9. 09

    Hardware benchmarks

    Tokens/sec and task success across CPU, GPU, and NPU on consumer laptops.

    Planned
  10. 10

    Permission prompts

    Ask before running anything destructive.

    Planned

Follow along

Repo coming soon.