Coding agent (CLI) · Prime Intellect
Prime Agent
Rust coding and research agent built around a persistent Python REPL where subagents are function calls; sessions run in a background daemon, can be reattached, and refine their own harness state.
Prime Agent review
Prime Agent is a Rust coding and research agent from Prime Intellect, built around a persistent Python REPL. The tool is free and MIT licensed. We read the vendor's repository in October 2026. We did not run an independent test.
Who it is for
Researchers and power users who want long-running, programmable agents with subagents and a persistent Python environment. It is not for anyone who needs a sandbox out of the box.
What stands out
Subagents are function calls in the REPL (`rlm.spawn`), for parallel or background work. Agents can message one another. Sessions run in a background daemon, so they survive terminal disconnects and can be reattached. The README also describes schedules, persistent goals and a bounded autonomous mode.
The CLI has a `-p` print mode and json, rpc and acp output modes, so it can be scripted. We took this from the CLI source, not prose docs. It includes an MCP client. At `/login` you choose a subscription or your own API key. Install is a one-line shell script.
Price and licence
The tool is free, as of 28 October 2026. The licence is MIT, with copyright to Prime Intellect Ltd. An appended note says it is a port of a TypeScript product by Mario Zechner. GitHub shows "Other" only because of that extra credit paragraph. Model cost comes from the provider you pick at login.
Where it falls short
- The README says it is not a security sandbox. Model-generated Python and commands run with your user permissions.
- Documentation is thin. The docs folder holds a Slack MCP guide and release notes, so supported models and local-model use are not documented.
- It is young: the repository was created in May 2026, and it is a Rust rewrite of an earlier TypeScript version. Behaviour may still shift.
- The REPL-centric workflow is aimed at research and long-running evaluation work.
Alternatives
See terminal coding agents. pi is a smaller MIT-licensed agent, and OpenCode is a general open-source option.
Facts: platforms, price and licence
- platforms
- macOSLinuxWindowsOne-line installer for Unix shells plus a PowerShell installer for Windows x64 (README). macOS and Linux are named in the changelog (kqueue/pidfd daemon code); the README does not list platforms explicitly. Rust rewrite of an earlier TypeScript version.
- price
- FreeTool is free and MIT licensed (seen 2026-10-28). Model cost comes from the provider you pick at /login: a subscription or your own API key.
- source
- Open sourceLicence: MIT
- github
- PrimeIntellect-ai/prime-agent 22.4k stars, seen 2026-10-28
Prime Agent features
- ✓ yes MCP
- ✓ yes Multi-agent
- ✓ yes Headless / CI
- – no Sandbox
- ? not verified Local models
- ✓ yes Your own API key
- ? not verified Git worktrees
- ? not verified SSH
✓ documented · – absent or not in its feature list · ? not verified
Prime Agent strengths
- Built-in subagents (rlm.spawn) for parallel or background work, and agents can message one another
- Daemon-backed sessions survive terminal disconnects, with heartbeats, schedules, persistent goals and a bounded autonomous mode
- Print mode and json/rpc/acp output modes in the CLI make it scriptable
- MIT licensed, Rust binary, frequent releases from a known lab
Prime Agent weak points
- README states it is not a security sandbox: model-generated Python and commands run with your user permissions
- Documentation is thin; the docs folder holds only a Slack MCP guide and release notes, so model and local-model support are not documented
- Young project (repo created May 2026) and a Rust port of an earlier TypeScript version, so behaviour may still shift
- Oriented to research and long-running evaluation work, with an unusual REPL-centric workflow
Who Prime Agent is for
- best for
- Researchers and power users who want long-running, programmable agents with subagents and a persistent Python environment.
- look elsewhere if
- Anyone who needs a sandboxed agent out of the box or a well-documented list of supported models.
Lists that include Prime Agent
Sources and what we could not verify
Checked 2026-10-28. Prices and features change; the vendor's page is the authority.
- github.com/PrimeIntellect-ai/prime-agent — README: description, install, subagents, daemon, sandbox warning, MCP, license statement, stars
- github.com/PrimeIntellect-ai/prime-agent/blob/main/LICENSE — MIT text and Mario Zechner port note
- github.com/PrimeIntellect-ai/prime-agent/blob/main/CHANGELOG.md — v0.10.1 dated 2026-10-21; macOS and Linux daemon details
- github.com/PrimeIntellect-ai/prime-agent/blob/main/crates/pa-cli/src/command_registry.rs — CLI flags: -p/--print, --mode text|json|rpc|acp, --provider, --api-key, --offline
Not verified
- local_models: only a test fixture and models.json custom-provider config mention ollama/baseUrl; no docs statement, left null
- git_worktrees: README only advises using a clean worktree; no worktree feature found
- no standalone product site found; app.primeintellect.ai/prime-agent hosts the installer only
- headless/scripting evidence comes from CLI source and help text, not prose docs
- platform list is inferred from installers and changelog, not an explicit support table