{ }
Codelit now turns an agent idea into a production workflow: Skills, MCP servers, triggers, tools, specialist agents, model routing, approval gates, eval harnesses, and the architecture needed to run it.
Specialists with roles, inputs, outputs, model preferences, escalation rules, and tool access.
Progressive instruction packs with activation rules, resources, deterministic scripts, and risk labels.
Tool, resource, prompt, roots, and sampling boundaries for real context and action systems.
Slack, GitHub, Notion, Linear, Jira, Stripe, browser workers, databases, and custom APIs.
Pick cheap models for classification, stronger models for planning, and fallbacks for reliability.
Human approval, read-only defaults, audit logs, data boundaries, and blocked irreversible actions.
Eval, sandbox, replay, approval, observability, and prompt-regression gates before rollout.
Trace history, runbooks, source ledgers, and durable state planned before the agent learns bad habits.
What You Get
Most agent builders start at prompts. Codelit starts at the operating model: what Skills activate, which MCP servers expose context or actions, who can trigger the agent, what it may touch, what model handles each task, what gets logged, and where a human must approve.
Agent map with responsibilities, inputs, outputs, and tool permissions.
Skill registry with activation rules, resources, and deterministic scripts.
MCP plan for tools, resources, prompts, roots, and model sampling boundaries.
Workflow steps with success paths, failure paths, and escalation points.
Model routing plan that respects BYOK provider choices.
Guardrails, evaluations, and harnesses before production deployment.
One-click path into Product Board or Architecture views.
Starter Templates
These are not toy prompt shells. Each template includes Skills, MCP boundaries, agents, tools, model routes, approval gates, harnesses, evals, and a deployment shape you can turn into an architecture.
A Slack-native engineering agent that receives operational requests, gathers context from tickets and repos, routes work to specialist agents, and drafts auditable responses before anything risky happens.
A code review workflow that watches pull requests, assigns specialist reviewers, checks implementation risk, and posts concise review notes with links to evidence.
A controlled workflow for small teams that want an internal agent to answer operational questions, create tickets, inspect connected apps, and prepare approved actions across the company stack.
A guarded browser operator for internal tools: it can open approved apps, complete repetitive workflows, verify the result, and stop at human approval checkpoints before sensitive writes.
A Devin-style engineering workflow that turns a scoped ticket into a branch, implementation plan, code changes, test run, and pull request draft with review evidence.
A support workflow that reads product docs, checks account and billing state, drafts empathetic replies, and escalates sensitive requests with source-linked evidence.
A meta-agent workflow that generates red-team cases, runs regression suites, scores agent behavior, and blocks production rollout when a workflow violates its safety contract.
A research workflow that gathers competitor signals, checks source quality, extracts positioning patterns, and writes briefs that separate facts from interpretation.
A security workflow that watches alerts, gathers evidence from code and runtime systems, ranks blast radius, and prepares a human-approved remediation plan before any production action.
A finance-aware workflow that inspects subscriptions, invoices, usage limits, and customer requests, then drafts revenue-safe actions for approval.
A launch workflow that coordinates release notes, docs, changelog updates, social copy, customer comms, and post-launch monitoring from one evidence-backed plan.
A data-quality workflow that watches analytics schema changes, validates event health, explains metric drift, and opens owner-ready fixes when instrumentation breaks.
Use your provider keys, choose models per task, and make the permission model visible before the first connector is allowed to act.
Open Agent Workflow