SIA

The Self Improving Agent by nith studio.

Used across our projects to turn agent experience into reusable, reviewed skills. SIA gives each kind of work an agent, each agent a versioned skill, and each run a report. The next run starts with the lessons worth keeping.

The improvement happens in readable Markdown. Your model does the work; you review what it learned. No model training, hidden memory service, or magic "it gets smarter" button. Unfortunately, those still mostly come with a sales deck.

task → brief + skill → agent → deliverable + report
              ↑                    ↓
        next run ← reviewed lesson + skill version

Install

Python 3.10 or newer. No runtime dependencies or API keys.

python -m pip install "git+https://github.com/BuildAppolis/sia.git"

Or clone this repository and run python -m pip install .. The distribution is named nith-sia; the command is sia. python -m sia works too. The initial release is installed from Git or source; it is not published on PyPI.

Start in a project

cd your-project
sia init
sia new-agent writer --role "Writes guides from verified sources." --tools Read,Write,Edit,Bash,Glob,Grep
sia new-skill write-guide --purpose "Write a short, verified guide for a named reader." --owner writer

Open .claude/skills/sia-write-guide/SKILL.md and replace the placeholder procedure with real steps. Then:

sia on
sia brief writer --task "Write a setup guide from README.md; save it in this run's output directory." --out brief.md

Give the brief to the named agent in your host. SIA does not launch agents. After reviewing the deliverable and any proposed lesson, save the agent's report to .claude/sia/out/r0001/report.md and record the run:

sia report writer --run r0001 --file .claude/sia/out/r0001/report.md
sia annotate r0001 --note "Checked the guide's commands against the project."
sia check

Use the actual run ID printed by brief. If you changed the skill, run sia bump-skill write-guide --note "what improved and why". An unchanged skill does not need a version bump. No new lesson is a valid outcome.

Bring your host

Host	Setup	What SIA provides
Claude Code	sia init --claude-hooks	Project agents, skills, /sia, and context hooks
Codex	Follow the adapter guide	Explicit status checks, shared skills and reports, tool translation
Other agents	Read the generated protocol and pass briefs manually	The same CLI and learning records

Claude setup merges settings without removing unrelated hooks or permissions. It refuses to replace an existing, different /sia command. The CLI must be available on the host's PATH. SIA starts off in new projects; sia off keeps all learning on disk.

What it tracks

• Agents, their roles, owned skills, and declared tool grants.
• Versioned skills with procedures, checks, and a bounded lessons section.
• Run briefs and the skill versions they used.
• Reports, reviewer annotations, and orchestration lessons.
• Tool requests, escalation decisions, grants, and revocations.

Everything lives in the project's .claude/ directory, retained for backward compatibility even when using another host. Each project has its own learning store. Copy reviewed skills deliberately; there is no automatic cross-project sync or private-history upload.

See the command reference, host setup, architecture and limitations, and migration guide.

Trust and limits

SIA is a local workflow and audit tool, not a security sandbox. Its grants record intended access; your host enforces actual permissions. --by user is an actor label, not proof of consent. A process that can edit the files can change the records. Read SECURITY.md before using it with sensitive projects or untrusted agents.

The package makes no network requests, contains no model client, and sends no telemetry. Agents you run through your host may have their own network access.

Development

python -m pip install -e .
python -m unittest discover -s tests -v
python -m pip install build
python -m build

Tests use temporary projects. They cover the run lifecycle, grant escalation, revocation, legacy upgrade, root discovery, hook setup, and command locking. CI runs on Windows and Linux with Python 3.10 and 3.13, then builds wheel/sdist.

Contributions: CONTRIBUTING.md. Licensed under MIT.