Getting Started

Install one binary, run three commands, and watch an agent write state that outlives the terminal. primkit needs no server, no account, and no config file — each primitive auto-creates its SQLite database on first use. This page gets you to durable state in about a minute; the per-primitive guides and the agent reference cover everything after that.

Let your agent install it

The lowest-friction path: paste this into Claude Code, Cursor, Codex, or any agent with shell access. It reads the docs, installs the primitives, and verifies them — no manual steps on your end.

Read https://github.com/propifly/primkit/blob/main/README.md and install primkit for me:
the four primitives (taskprim, stateprim, knowledgeprim, queueprim) via go install, or the
pre-built binaries if Go isn't available. Then run a quick command to confirm each works.

Prefer to do it by hand? The manual steps are below.

Install

The fastest path is go install (requires Go 1.26+). Install only the primitives you need:

go install github.com/propifly/primkit/taskprim/cmd/taskprim@latest
go install github.com/propifly/primkit/stateprim/cmd/stateprim@latest
go install github.com/propifly/primkit/knowledgeprim/cmd/knowledgeprim@latest
go install github.com/propifly/primkit/queueprim/cmd/queueprim@latest

Prefer a pre-built binary? Download one for your platform from the latest release and move it onto your PATH. The README lists every install option, including curl one-liners and building from source.

Your first durable state

These three commands create a task, query what's open, and mark it done — the full lifecycle, all from the shell. The database is created automatically at ~/.taskprim/default.db the first time you run add.

# 1. Create a task (source = who created it, typically an agent name)
taskprim add "Deploy v2 to staging" --list ops --source agent

# 2. List what's open
taskprim list --list ops --state open

# 3. Mark it done (use the ID printed in step 2)
taskprim done t_abc123

Close the terminal, open a new one, and run step 2 again — the state is still there. That persistence across sessions is the whole point.

Every command speaks JSON for programmatic callers:

taskprim list --list ops --state open --format json

Wire it into an agent

An agent with shell access uses primkit like any Unix tool — no SDK to import, no framework to adopt. To expose the primitives over the Model Context Protocol (for IDE agents such as Cursor or Claude Desktop), run a primitive in mcp mode:

taskprim mcp --transport stdio

The README has ready-to-paste .mcp.json and claude_desktop_config.json snippets for all four primitives.

Where to go next

You now have one primitive running. The other three follow the same shape — a single binary, its own SQLite file, CLI by default:

• taskprim Guide — task lifecycle, lists and labels, dependency graphs, and the "what can I work on next?" frontier.
• stateprim Guide — namespaced key-value state, dedup checks, and append-only logs.
• knowledgeprim Guide — a knowledge graph with hybrid full-text + vector search.
• queueprim Guide — persistent work queues with priority, retries, and dead-letter handling.

When you want primitives to compose into agents that investigate open questions and accumulate knowledge across sessions — not just react once — read Building Curious Agents. For the exact command and JSON surface an agent calls, see the agent reference.