Agentit is a portable, provider-neutral meta-harness for AI coding agents: activate with natural language, interview once, load only the skill family you need, spawn specialists when useful, critique large plans, verify with evidence, and ship through PRs.
Agentit is designed to work across OpenAI, Anthropic, Google, xAI, and compatible future coding-agent environments. Provider-specific subagents/workers are optional execution primitives; the shared Agentit protocol is semantic and portable.
The only special phrase is a natural use Agentit in your language:
usa agentit
use agentit
utilise agentit
No other powerwords. Ordinary prompts drive routing (“frontend and backend”, two file paths, “at the same time”, “review and fix”, “several agents”, …).
usa agentit y crea mi portfolio personal
For product-affecting work Agentit follows roughly:
inspect facts
↓
one comprehensive interview batch
↓
domain pack (+ craft depth only if design/visual)
↓
project-aware token estimate
↓
persist resumable project state
↓
skill budget + MCP fit + optional specialists
↓
critic on large structural plans
↓
implement on work branch
↓
verification → PR by default
Purely mechanical chores can bypass the interview.
Agentit interviews every product-affecting task, not only ambiguous ones.
Before asking, the agent inspects repo/docs/tools. Then it asks all currently identifiable material decisions in one batch.
It recommends a domain pack (skill family: frontend, backend, design, data, …) and loads only that family plus a tiny always-core, not the whole catalog.
Standard / Polished / Studio are design/visual craft depth only — not a universal tax on every task. Token estimates are project-aware (router token_estimate), not fixed billing tables.
Canonical files: skills/interview-me/SKILL.md, effort/levels.yaml, docs/AGENTIT_INTERVIEW_AND_PROVIDER_POLICY.md, skills/mcp-tooling-fit/SKILL.md.
Agentit assumes any chat can disappear because of context/token exhaustion, provider/model switch, app crash, machine switch, or a long pause.
For every product-affecting task, maintain a compact project state document at:
docs/agentit/STATE.md
If the project already has an equivalent canonical state file, reuse it instead.
The state must let a completely fresh agent recover:
- what is being built and why;
- confirmed intent, audience, success criteria, constraints, and non-goals;
- domain pack, craft depth if design/visual, spend, token estimate;
- what is done / in progress / blocked / not started;
- durable product, architecture, API, data, and design decisions;
- important files/artifacts;
- branch + PR;
- verification commands/results;
- next executable actions;
- open user questions/blockers.
Update it after interview confirmation, expensive-to-rediscover decisions, meaningful milestones, before handoff/context exhaustion/pause, and before completion.
Do not persist secrets, credentials, raw chain-of-thought, full chat transcripts, or giant tool dumps.
Canonical policy: docs/PROJECT_CONTINUITY.md.
For repository changes Agentit defaults to:
work branch → commits → verification → pull request → review/user merge decision
It should not commit/fast-forward directly onto main/master and should not auto-merge PRs unless the user explicitly authorizes that exception for the task or project instructions require another workflow.
Continuity/docs updates travel in the same branch/PR as implementation.
agents/catalog.yaml defines semantic roles with small skill bundles and output contracts. Examples include:
frontend-developer
backend-architect
ai-engineer
devops-automator
design-system-researcher
ui-researcher
trend-researcher
creative-tool-scout
visual-storytelling-director
spatial-experience-designer
delight-and-whimsy
design-critic
performance-benchmarker
api-tester
workflow-optimizer
Agentit uses intelligent delegation: stay single-agent when that is best; spawn specialists when independence, isolation, domain expertise, or independent critique wins. No hard min/max subagent caps. Large structural plans require an independent critic. If the user asks for multi-agent without benefit, the Architect should push back.
Execution fallback:
native provider subagent/worker
↓ unavailable
isolated delegated model/tool call
↓ unavailable
separate fresh-context invocation
↓ unavailable
parent + exact specialist skill bundle
Multi-agent execution is an optimization, never a correctness dependency.
Specialists declare stable required and preferred capability IDs such as
repository.read, design.inspect, or browser.inspect. Agentit resolves those
IDs against an explicit host inventory using ordered ChatGPT app, MCP, CLI, and
local fallbacks.
specialist -> capabilities -> explicit available providers -> scoped grants
No inventory means no assumed grant. Delegated workers receive only the selected capability/provider permissions, and an explicit unresolved required capability fails the pre-spawn gate. Resolution is plan-only: Agentit does not install, authenticate, enable, or call providers.
./agentit capabilities resolve \
--specialist frontend-developer \
--host codex \
--available mcp.github,local.filesystem,mcp.playwrightSee docs/CAPABILITIES.md for the catalog, inventory
contract, fallbacks, least-privilege envelope, and extension guide.
Agentit's design profile includes ui-ux-pro-max-intelligence, a provider-neutral JIT adapter for the MIT-licensed upstream nextlevelbuilder/ui-ux-pro-max-skill.
The upstream project provides searchable product-aware intelligence for style families, palettes, typography, UX/accessibility rules, icons, charts, GSAP/motion patterns, landing structures, and many implementation stacks.
Agentit deliberately treats it as an intelligence source, not the creative director:
UI UX Pro Max lookup
↓
compact product/design baseline
↓
Taste / creative direction / research
↓
implementation
↓
Impeccable / Emil / critic
The database should be queried JIT. Do not dump it wholesale into model context and do not let a preset style automatically become the art direction.
Effort behavior:
- Standard: narrow lookup only when it prevents a mistake or answers a concrete design question.
- Polished: targeted product/style/color/type/UX intelligence when useful.
- Studio: one evidence source among live inspiration research, concept exploration, creative direction, and independent critique.
See skills/ui-ux-pro-max-intelligence/SKILL.md and THIRD_PARTY_NOTICES.md.
The design profile contains a broad but JIT-routed craft stack:
ui-ux-pro-max-intelligence structured UI/UX design intelligence
design-inspiration-research live project-specific references
design-trend-researcher emerging / maturing / saturated patterns
creative-web-experiences concept generation
design-taste-frontend art direction / visual thesis
impeccable-design critique / polish / responsive craft
emil-design-eng interaction and motion feel
visual-storytelling-director narrative beats and pacing
creative-tool-scout current implementation tooling
figma-design-workflow official Figma MCP workflow
scrollytelling-web narrative scroll architecture
gsap-scrolltrigger pin/scrub/timeline mechanics
gsap-performance motion runtime performance
threejs-spatial-experiences rooms/stores/museums/worlds
threejs-product-storytelling GLB/glTF product storytelling
delight-and-whimsy restrained memorable details
Do not load the whole stack just because design is active. Effort level and task signals control depth.
For genuinely high-ambition Studio work, Agentit may use a design competition: shared evidence brief → 2-3 independent concepts → explicit jury by brand fit/originality/clarity/usability/feasibility/performance/memorability → winner or justified hybrid → implementation → independent critique.
profiles.yaml keeps the global install bounded and activates deeper capabilities JIT.
| Profile | Purpose |
|---|---|
core |
bounded everyday engineering harness |
frontend |
browser/performance/UI maintenance |
design |
full craft studio + UI/UX intelligence |
backend |
API/service/observability |
supabase |
Postgres/Supabase guidance |
product |
discovery/spec/marketing |
writing |
documentation/writing |
release |
CI/migration/launch |
research |
context/spec/adversarial review |
all |
escape hatch only |
agentit enable design --project . --apply
agentit status --project .
agentit disable design --project . --applyAgentit separates model claims from verification evidence.
agentit verify "task summary" --project .
agentit verify "task summary" --project . --applyNo done, fixed, or passing claim without fresh evidence after the last relevant edit. Design work needs rendered evidence when browser tooling is available; high-ambition work should normally get independent critique/performance review.
agentit mcp status
agentit mcp enable context7 --apply
agentit mcp enable-stack developer_core --apply
agentit mcp enable-stack design_studio --applyThe design studio stack can combine Figma, Context7, Playwright, and Chrome DevTools. MCPs are optional capabilities, not portability requirements.
git clone https://github.com/marcmarti9/agentit.git ~/code/agentit
cd ~/code/agentit
bash install.sh --provider all --with-guides --apply
ln -sf ~/code/agentit/agentit ~/.local/bin/agentitOpen Skills / compatible clients typically use ~/.agents/skills/; Claude can use ~/.claude/skills/; Codex can use ~/.codex/skills/. Shared Agentit semantics remain provider-neutral.
python3 -m unittest discover -s router -p "test_*.py"
python3 -m unittest discover -s tests
python3 evals/run.py| Doc | Purpose |
|---|---|
AGENTS.md |
global agent playbook |
docs/AGENTIT_INTERVIEW_AND_PROVIDER_POLICY.md |
batched interview + effort + provider semantics |
docs/PROJECT_CONTINUITY.md |
resumable project-state contract + PR-first workflow |
agents/catalog.yaml |
reusable specialist roles |
effort/levels.yaml |
Standard / Polished / Studio budgets |
docs/MCP_CATALOG.md |
MCP catalog/runtime |
docs/ADAPTIVE_AGENT_ARCHITECTURE.md |
orchestration topologies/contracts |
THIRD_PARTY_NOTICES.md |
upstream design-skill attribution |
Licensed under the Apache License, Version 2.0. Third-party adaptations/integrations retain their applicable notices in THIRD_PARTY_NOTICES.md.