Prime Agent Hits 19K Stars With One Tool and No API Key Requirement
3 min readRunning Prime Agent on a Local Model — the models.json provider config for Ollama and vLLM, which compat flags matter for which backend, why auto-refine should be off before you trust it on a small model, and the telemetry and sandbox defaults worth changing. No Prime Intellect account required.
PrimeIntellect-ai/prime-agent has picked up 19,254 stars and 2,095 forks since its 8 May 2026 launch, on the back of Prime Intellect's $130M Series A. It is MIT-licensed TypeScript, and — unusually for a company-backed agent — it needs no account with the company that publishes it.
Two design choices make it worth a look if you run models locally.
It has one tool. Not read_file, write_file, bash and twelve others — one, called ipython. Every action the agent takes is Python executed in a persistent kernel. Even spawning a subagent is a function call inside that kernel rather than a tool schema entry. The pitch is that a smaller model has one interface to learn instead of a large tool menu to route between.
Local models are first-class. ~/.prime/agent/models.json takes arbitrary providers with a baseUrl, so Ollama, vLLM and LM Studio drop straight in, with per-provider compat flags for endpoints that deviate from the OpenAI schema. The docs name Ollama, vLLM and LM Studio explicitly.
Two things are worth reading carefully before you commit to it.
The "self-improving" framing does not mean what it usually means. There is no training, no RL, no gradients — the mechanism is an LLM call that emits JSON edits to a harness_state.json file holding prompt additions, memories, skills and subagent definitions. It runs automatically every 25 assistant turns by default. That is a genuinely useful feature and also an ordinary one, and on local models it is where the rough edges are: issue #1143 reports the refiner failing to return valid JSON against Ollama.
And the README is blunt that the agent's permission modes "are not a security sandbox." Actual sandboxing is an opt-in example extension you wire up yourself. An agent whose only tool is arbitrary Python execution deserves that warning taken at face value.
Also worth knowing: it is a rebrand of pi by earendil-works — the package is still published as @earendil-works/pi-coding-agent, and the local-model support is inherited from upstream rather than built for this release.
→ Running Prime Agent on a Local Model covers the models.json config for Ollama and vLLM, which compat flags matter for which backends, why you should turn auto-refine off on small models first, and the telemetry opt-out.
Source: PrimeIntellect-ai/prime-agent · Relevance: 8/10