Open source agentic framework
A cross-tool operating layer for AI agent work. It helps teams turn one-off chats into repeatable work: clear scope, the right specialist, remembered decisions, checked outputs, and human approval before high-risk actions.
It covers strategy, software delivery, finance, leadership, engineering, design, marketing, sales, ops, research, people, and customer success.
Works across Claude Code, Cursor, Codex, opencode, Antigravity. MIT open source, 71 skill protocols, 36 commands, 12 domains.
/goal /plan /grill /build /debug /ship Loads the shared rules before the agent starts.
SKILL.md Owner, authority, write set, definition of donecontract -> invoke -> verify -> stop Turns a vague task into bounded work.
.resonance/02_memory.md Memory and ledger records survive the chatlesson -> decision -> metric -> next session Keeps durable context in the project.
docs/EVALS.md Golden cases and keep-or-revert improvementwith/without -> scorecard -> remeasure Makes skill improvement measurable before it stays.
Resonance gives product and engineering teams a shared way to run AI-assisted work. Product people see the job, owner, risk, and outcome. Technical people can inspect the skill file, write boundaries, validators, and evidence trail.
Operating contracts
Before an agent acts, the skill can define the owner, allowed actions, required output, touched files, helper skills, and stop rules. Product teams get fewer surprises. Technical teams get inspectable boundaries.
Read the sourceProject knowledge
Resonance can record project identity, state, decisions, lessons, metrics, customers, and experiments in the repo. The next agent starts with that context instead of asking the same questions again.
Read the sourceSkill improvement
Resonance ships eval cases and tooling, not public performance results. If a skill change does not improve the measured work, the operator can revert it.
Read the sourceCross-tool compilation
Forge compiles the same skill sources into host-specific skills, prompts, commands, and context bridges. The working method travels with the repo.
Read the sourceUse case
The easiest way to understand Resonance is to follow a normal request. A conductor frames the job, calls only the specialists needed, keeps the work inside clear limits, and stops before publish, spend, send, or ship actions need approval.
Common use case
We need a first campaign for a technical audience that hates hype. Find the angle, name the risks, prepare the assets, set up tracking, and stop before anything is published or spent.
run-campaign Creates the campaign spine: audience, offer, channels, budget, proof standard, approvals, and outcome check.
/grill Tests weak assumptions, claims, audience proof, and risk before assets are made.
copywriter Drafts the plain-language argument, hooks, subject lines, CTAs, and anti-slop pass.
channel skills Routes to content distribution, lifecycle, paid acquisition, analytics, or studio only when the brief needs them.
legal Stops high-risk claims, privacy, consent, or regulated outreach until the owner reviews them.
launch proposal Returns assets, checks, open risks, and the exact publish or spend action for human consent.
What comes back
Deliver a software change
deliver-change/grill/plan/build/test/audit Clear scope, implementation, checks, review, and release proposal without auto-shipping.
Run a revenue motion
run-revenue-motionaccount intelligencelead opsrevopscustomer success Buyer next actions, sources, CRM change proposal, forecast method, and approval gates.
Run a finance cycle
run-operating-cyclefinanceanalyticslegalledger Sourced numbers, assumptions, options, approval record, and follow-up date.
Run a leadership cycle
run-operating-cyclefounder-oshiringproductivityretro Decision brief, cadence, delegation map, communication plan, and review date.
Protocol in action
An operating contract is a small rulebook for the task. It says who owns the work, what inputs are needed, what output must come back, which files or systems may be touched, which specialists can be called, and when the agent must stop. The goal skill is one example. See the goal skill source.
owner -> authority -> inputs -> outputs The skill names the job and the expected return before work starts.
The contract declares which project areas it may touch, so coordinated work avoids silent collisions.
Thin conductors call specialist skills for copy, design, finance, sales, security, or engineering work.
High-authority actions stop for human approval. Publishing, spending, sending, and shipping are not automatic.
Completion attaches checks, decisions, or outcome dates instead of relying on model confidence.
Define the job, owner, authority, inputs, and outputs before work starts.
Use the conductor only when coordination is needed. Use the specialist directly for a narrow job.
Respect write sets, privacy, legal, money, people, and shipping boundaries.
Ground completion in tests, validators, audits, source notes, approvals, or dated outcome checks.
Write durable decisions, lessons, metrics, customers, or experiments into the project memory or ledger.
Framework anatomy
A prompt can be useful once. A skill can be reused by a team. It defines who owns the work, what may be touched, how it should happen, what to check, how to recover, and when the job is actually done.
.agents/skills/**/SKILL.md A procedure with owner, authority, inputs, outputs, touched files, recovery path, references, and definition of done.
.forge/commands.json A short invocation such as /goal, /debug, /design, /ship, or /page-audit.
.resonance/ Project-owned context and typed ledgers for decisions, lessons, metrics, customers, and experiments.
.forge/ The compiler and validators that produce host-specific shims from one source.
AGENTS.md, CLAUDE.md, tool folders The files each tool loads so the operating standard and memory reach the agent.
.forge/exec/ Hooks for real tests, browser checks, validators, and outcome-based verification.
software, marketing, sales, finance, leadership A thin orchestration skill that coordinates specialists without replacing their doctrine or auto-shipping.
Command map
Commands are shortcuts into repeatable work. The important part is not the slash. It is the procedure, memory, approval boundary, and check that run around it.
/init/brief/plan/grill/council/venture-model/gtm-thinker/market-research /goal/build/debug/design/studio/refactor /test/audit/page-audit/review-pr/second-opinion/system-health /ship/incident/seo/friction/update-roadmap/update-resonance /improve/skill-author/capture/explain/handover/retro/voice-profile/call-intelligence/cold-call/sales-pipeline Beyond coding
The newer framework adds conductors for software delivery, campaign work, revenue motions, finance cycles, and leadership cycles. That matters because real teams rarely have isolated coding tasks. They have decisions, assets, numbers, customers, risks, owners, and follow-up.
Move from request to reviewed release candidate with scope, evidence, audit, and ship proposal.
deliver-change/grill/plan/build/audit Prepare campaigns with clear claims, channel assets, tracking, approvals, and outcome checks.
run-campaigncopywritercontent distributionanalytics Coordinate account research, lead ops, outbound, pipeline, forecast, and handoff as one controlled path.
run-revenue-motionaccount intelligencerevopscustomer success Run finance and leadership cycles from source data or team context to decisions, owners, and follow-up dates.
finance cycleleadership cycleledger/retro Grounded checks
Resonance points agents toward real checks: skill contracts, graph validation, ledger validation, project tests, browser checks, golden evals, planted-defect cases, definitions of done, and recovery paths. Product people can ask, "what evidence came back?" Technical people can inspect the check that produced it. Read the eval methodology.
python .forge/validate_skill.py --all --strict .agents/skills python .forge/validate_library.py --strict docs/SKILL_GRAPH.md npm run eval:score Setup
Current implementation details live in the README. New clones are direct. Existing projects should use the transactional updater so project memory, local decisions, and owned files stay protected.
git clone https://github.com/manusco/resonance
cd resonanceOpen the folder in your AI tool. The commands are already committed.
Use the updater from a clean Resonance checkout. Run the dry run first, review its JSON plan, then apply. It refuses user-owned conflicts, backs up replaced files, and keeps .resonance/ outside the managed write set.
After installation, run /init. That creates the `.resonance/` memory owned by your project.
FAQ
No. A prompt asks for an output. A Resonance skill defines how to do the work: who owns it, what it may touch, what it must return, how to recover, and how done is checked. Commands call those protocols from the tool you already use.
It does not train model weights or collect private project data. It gives your repository memory for project learning and an operator-run eval loop for skill improvement.
No. Engineering is one domain. Resonance also supports product work, campaigns, revenue motions, finance cycles, leadership cycles, design, research, people, and customer success.
No. Resonance is the operating layer inside those tools. It gives agents shared skills, commands, and memory while you keep using the host you prefer.
The framework can be upgraded while preserving .resonance/, which belongs to your project. Exact counts and implementation details live in the GitHub repo and current README.
Yes. Resonance is published under the MIT license at github.com/manusco/resonance.