# Prime Agent: features, protocols, quickstart ## TL;DR Prime Agent is a Rust coding and research agent from Prime Intellect that runs the model inside a persistent Python REPL, where subagents are spawned as function calls. A daemon keeps sessions running after the terminal detaches. GitHub reports no license, but the LICENSE file is MIT text. It suits long-running, research-style tasks. ## Key facts | Field | Value | | --- | --- | | Type | Harness | | Languages / SDKs | Rust, Python | | License | unknown | | Pricing model | Open source, free | | Orchestration pattern | Supervisor | | GitHub stars | 21,399 (as of 2026-09-30) | | GitHub forks | 2,350 | | Last push | 2026-09-30 | | Latest release | v0.9.8 | | Repository | [PrimeIntellect-ai/prime-agent](https://github.com/PrimeIntellect-ai/prime-agent) | | Website | [primeintellect.ai](https://primeintellect.ai) | | Documentation | [github.com](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md) | | Last verified | 2026-09-30 | ## Key features - A persistent Python REPL is the model's built-in tool; file operations, shell commands, tool use, subagents and context management all go through code, following the Recursive Language Model design the README describes. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - `rlm.spawn(...)` starts real child agents for parallel or background work and returns their results programmatically; the runtime also exposes calls to list, collect and delete subagents. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/crates/pa-core/README.md)) - Agent-to-agent messaging through a local daemon: an agent can message its parent, siblings and direct children, with sender identity set by the daemon, and a read-only observe skill lists the family. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/skills/agent-message/SKILL.md)) - Daemon-backed sessions keep the agent, REPL state, schedules and subagents running when the terminal detaches, and `prime-agent attach` reattaches. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - `/refine` reviews the trajectory and can save small updates to supplemental prompts, memories, skill descriptions and subagent specifications, without rewriting the base system prompt, with snapshots for rollback. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - Long-task controls: `/goal` persistent objectives, heartbeats and schedules, and a bounded `/autonomous` mode with turn, token and time budgets and user-defined quality gates. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - An MCP skill gives the REPL a generic `mcp` module to search a service catalog, list connections, discover tool schemas and call tools on connected servers. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/skills/mcp/SKILL.md)) - An ACP stdio transport in the daemon (initialize, session/new, session/prompt, session/cancel and session updates) for Agent Client Protocol clients. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/crates/pa-daemon/README.md)) ## Architecture and orchestration pattern Pattern: Supervisor. The repository is a Rust workspace with crates for the agent loop, provider streaming, the session engine, a session-supervising daemon and a terminal UI, plus a Python runtime and a set of Python skills. The AGENTS.md file says the Rust implementation ports an earlier TypeScript product and is checked against it. The model does its work by writing Python into a persistent kernel; skills are importable Python packages. Subagents are children of the running session. The runtime's `rlm` module spawns them, lists them, collects results and deletes them, and the daemon carries messages between an agent's parent, siblings and children. A woken idle child continues in the same session and context, and a child stays available until its parent session closes. `prime-agent agents` browses running, idle and saved sessions. State is on disk and in the daemon. Sessions persist and resume, goals and scheduled jobs are file-backed, and the harness keeps supplemental prompts, memories, skill descriptions and subagent specifications as refinable state next to an immutable base prompt. ### Human in the loop Users can attach to a running session and keep interacting with it, browse and resume sessions, and pause, complete or clear a `/goal`. Autonomous mode runs only within configured turn, token and time budgets, and the README says a passed quality gate checks only what the gate verifies. Messages between agents use steering delivery, so a busy target sees them during its run. The README warns that model-generated Python and commands run with the user's permissions and that its worker and kernel processes are not a security sandbox. No tool-approval prompt is described in the README or crate docs read. ## Protocols | Protocol | Support | Evidence | Note | | --- | --- | --- | --- | | MCP | Yes (checked 2026-09-30) | [link](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/skills/mcp/SKILL.md) | Client side: the built-in mcp skill lets the REPL search a connector catalog, list connections and call tools on connected MCP servers; the pa-core README says the MCP protocol runs kernel-side over stdio or HTTP with an OAuth flow on the host. No MCP server mode found. | | A2A | Unknown (checked 2026-09-30) | — | No A2A / Agent2Agent mention in the README or crate docs read; code search hits for a2a were a battery README, a prompt layer, messaging code and a parity script, none read as protocol support. Agents talk through a daemon-local messaging skill instead. | | AG-UI | Unknown (checked 2026-09-30) | — | No AG-UI mention in the README or crate docs read; repo code search for ag-ui returned nothing; not checked against a listing in the AG-UI README. | ## Best for - Long-running research or evaluation work that needs a session to survive terminal disconnects, with goals, schedules and bounded autonomy. ([shortlist](https://multiagentguide.top/best/research-agents.md)) - Tasks where the agent spawns child agents from code and coordinates them by messages, using the parent, sibling and child roster. ([shortlist](https://multiagentguide.top/best/parallel-coding-agents.md)) - Experimenting with a harness that refines its own supplemental prompts, memories and skills while keeping a rollback history. - Developers who want a Rust terminal agent with an ACP transport for editors. ([shortlist](https://multiagentguide.top/best/coding-agents.md)) ## Not for - Untrusted repositories or instructions without an external sandbox; the README says the worker and kernel processes are not a security sandbox. - Users who need a standalone documentation site; the README's Documentation link points to its own install section. - Contributors who want to open pull requests freely; PRs are reviewed from maintainers and vouched contributors. ## Quickstart ```sh curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh ``` Install not yet verified by this site. ```bash # Install the latest build (rolling beta from the rust branch; stable on release) curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh # Start in the project you want it to work on (use a disposable clone or worktree) cd /path/to/project prime-agent # on first launch, run /login to choose a subscription or API-key provider # Session management prime-agent agents prime-agent attach prime-agent --resume prime-agent status prime-agent doctor prime-agent shutdown ``` ### Common pitfalls - The agent executes model-generated Python and project commands with your user permissions and is not sandboxed; use a disposable clone, clean worktree or external sandbox. - The installer serves a rolling beta from every push to the `rust` branch, while the stable channel ships on release, so the installed build can change between runs. - Run `/login` on first launch to select a provider. - Daemon-backed sessions keep running after the terminal closes; `prime-agent shutdown` stops every agent, worker and background service. - A child subagent stays available only until its parent session closes. - GitHub's license detector reports no license; the LICENSE file is MIT text followed by an attribution paragraph. Official quickstart: https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md#install ## Pros - Built-in subagents with a roster, result collection and daemon-mediated messages among parent, siblings and children; the daemon sets the sender identity. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/skills/agent-message/SKILL.md)) - `/refine` changes only supplemental harness state, never the immutable base system prompt, and records snapshots for rollback, according to the README. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - The installer and everything it downloads (version pointers, tarballs, checksums) come from one domain, per the README. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - Autonomous mode is bounded by turn, token and time budgets and can run quality gates. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - Frequent tagged releases, with stable v0.9.x versions published in late September 2026. ([source](https://github.com/PrimeIntellect-ai/prime-agent/releases)) ## Cons - Not a security sandbox: the README says model-generated Python and commands run with the user's permissions and the worker and kernel processes only improve lifecycle isolation. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - Documentation is the README and per-crate notes; the README's Documentation link points to its own install section rather than a docs site. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - The default installer tracks a rolling beta built from every push to the rust branch, with the stable channel shipping on release. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - Pull requests are reviewed from maintainers and vouched contributors, and questions and bugs start as GitHub Discussions. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md)) - GitHub cannot classify the license because the LICENSE file adds an attribution paragraph after the MIT text, crediting an earlier TypeScript product by Mario Zechner. ([source](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/LICENSE)) ## Alternatives - [Oh My Pi (omp)](https://multiagentguide.top/tools/oh-my-pi.md) - [Pi](https://multiagentguide.top/tools/pi.md) - [OpenCode](https://multiagentguide.top/tools/opencode.md) - [Claude Code](https://multiagentguide.top/tools/claude-code.md) - [Codex CLI](https://multiagentguide.top/tools/codex.md) ## FAQ ### Does Prime Agent support MCP? Yes, as a client. A built-in mcp skill lets the REPL browse a connector catalog, list connections and call tools on connected MCP servers, and the engine handles OAuth. No MCP server mode was found. ### What license is it under? GitHub reports no license because the LICENSE file appends an attribution paragraph to the MIT text. The file states the MIT license with copyright to Prime Intellect Ltd, and the README says it is fully open source under MIT. ### How does it coordinate several agents? The agent writes Python in a persistent REPL and calls rlm.spawn to start child agents, then lists and collects them. A local daemon delivers messages between an agent's parent, siblings and children, and the TUI shows subagents. ### Is it a sandbox? No. The README says worker and kernel processes improve lifecycle isolation and recovery but are not a security sandbox, and recommends a disposable clone, worktree or external sandbox. ### Is it free? The repository is MIT licensed and no paid plan is documented. You sign in with a subscription or API-key provider through /login, and provider usage is billed by that provider. ## Sources - [Prime Agent README](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md) - [pa-core crate README (session engine)](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/crates/pa-core/README.md) - [agent-message skill](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/skills/agent-message/SKILL.md) - [mcp skill](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/skills/mcp/SKILL.md) - [pa-daemon crate README (ACP transport)](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/crates/pa-daemon/README.md) - [Prime Agent releases](https://github.com/PrimeIntellect-ai/prime-agent/releases) - [Prime Agent LICENSE file](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/LICENSE) - [Prime Agent README, install section](https://github.com/PrimeIntellect-ai/prime-agent/blob/main/README.md#install) - [Prime Intellect website (linked from README)](https://primeintellect.ai) - [PrimeIntellect-ai/prime-agent GitHub repository](https://github.com/PrimeIntellect-ai/prime-agent) ## Unknown fields license: GitHub reports NOASSERTION (unknown). The LICENSE file (25 lines) is the standard MIT text with copyright to Prime Intellect Ltd, followed by an attribution paragraph saying the software is a port of the Prime Agent TypeScript product, originally copyright 2025 Mario Zechner, used under the MIT License; the README also says MIT. pricing_model is oss-free on that basis. protocols.a2a: no mention found in README or crate docs; code search hits were not read as A2A support. protocols.agui: no mention found. harnesses: the README shows no launching of, or plug-in to, Claude Code, Codex, OpenCode, OpenClaw or Hermes. hitl: no tool-approval mechanism is described in the docs read. Corrections or removal requests: support@multiagentguide.top --- Data as of 2026-09-30. Not affiliated with listed projects. HTML version: https://multiagentguide.top/tools/prime-agent