TAKT Workflow Orchestrator
A YAML-controlled CLI for repeatable multi-agent coding workflows with review loops, checkpoints, and traceable runs.
Per-dimension scores and reasoning
Evidence shows: workflow steps can configure permissions (e.g., edit: false), tasks run in isolated worktrees, and user confirmation is required (e.g., /queue as task). Data flow transparency: logs and reports are recorded under .takt/runs/, routing decisions are recorded locally. Sensitive data handling: API keys are managed via environment variables, but no encryption at rest is mentioned. Dependency security: many dependencies, but no security audit or vulnerability scan evidence. External effects: can create PRs, post to Discord, etc., but requires user configuration. Rollback: task branches can be deleted, but no explicit rollback mechanism. Source attribution: author info in package.json, but publisher unverified. Deductions: insufficient evidence for sensitive data handling and dependency security.
Evidence shows: README and package.json are consistent, version number is clear. Dependency availability: dependency list is complete, but availability of all dependencies not verified. Failure messages: documentation mentions error handling, but no specific failure message examples. Deduction: insufficient evidence for failure messages.
Evidence shows: targets AI coding workflows, provides multiple workflows and customization options. Capability boundaries: defined via permissions and output contracts. Trigger precision: rule conditions are explicit. Environment fit: supports multiple providers and Node.js versions. Deduction: no major deductions.
Evidence shows: information architecture is clear, with documentation directory. Install notes are detailed. Naming stability: commands and terminology are consistent. Examples and FAQ: tutorials and examples provided. Known limitations: not explicitly listed. License: MIT license is clear. Versioning changelog: no CHANGELOG provided. Maintenance responsibility: GitHub Actions and release process present. Deductions: missing known limitations and versioning changelog.
Evidence shows: output usability: generates reports and logs. Marginal value: clear advantages over plain AI agents. Cost benefit: requires Node.js and external CLIs, but no performance benchmarks. Deduction: insufficient evidence for cost benefit.
Evidence shows: README claims partially align with code, but no independent verification. Cross-source corroboration: CI and test scripts exist, but no test results provided. Fact-inference separation: documentation distinguishes feature descriptions and recommendations, but not explicitly labeled. Deductions: lack of independent verification and test results.
- Publisher identity is unverified; assess supply chain risk carefully.
- Many dependencies without security audit; check for known vulnerabilities.
- Sensitive data handling details are insufficient; confirm API key storage.
- Known limitations and changelog are missing, which may affect long-term maintenance.
What does this agent do, and when should you use it?
TAKT is an open-source CLI for defining planning, implementation, review, fix loops, and human checkpoints as YAML workflows. A workflow consists of steps with personas, editing permissions, transition rules, and output contracts; rules can advance work, return it for fixes, or end it with COMPLETE or ABORT. Queued tasks run in isolated Git worktrees through `takt run`, while logs, reports, sessions, traces, and other run artifacts are kept under `.takt/runs/<run>/`. It supports Claude Code, Claude SDK, Codex SDK, OpenCode SDK, Cursor, GitHub Copilot CLI, and Kiro, and exposes both `takt-acp` for Agent Client Protocol and `takt-mcp` as a stdio MCP server. Although built primarily for coding workflows, its workflow model also applies to coordinated tasks that benefit from review and feedback loops.
A user starts with takt or takt add, refines a task in conversation, invokes /go, and can select Queue as task; TAKT stores queued task specifications in .takt/tasks/. takt run creates an isolated worktree for pending work and executes the YAML-defined steps according to initial_step, max_steps, persona, edit, required_permission_mode, and rules. Rules route a run to another step, COMPLETE, ABORT, or back to implementation for a fix loop; execution artifacts, including logs, reports, sessions, traces, and metadata, remain under .takt/runs/<run>/. takt list manages task branches with merge, retry, requeue, force-fail, instruct, and delete actions, while takt watch monitors and automatically executes pending tasks.
- A software team that wants a repeatable plan–implement–review–fix process can use the default workflow and its parallel peer-review and convergent fix loop.
- A maintainer of a CLI or backend service can assign separate personas, policies, knowledge, and edit permissions to planning, implementation, and review steps.
- A Git project that must keep queued changes out of the current working tree can execute them through
takt runin isolated worktrees. - A team using Codex SDK, Claude SDK, or OpenCode SDK can configure
provider,model, and API credentials for workflow execution. - A GitHub-based project can turn an issue into a task with
takt add #6ortakt #N, then manage the resulting task branch and PR flow. - An engineering team integrating a task queue with an MCP or Agent Client Protocol client can run
takt-mcportakt-acpover stdio.
What are this agent's strengths and limitations?
- YAML steps and transition rules make review and fix loops explicit instead of relying on an agent to remember a prompt-defined process.
- Each step can receive its own persona, policy, knowledge, instructions, output contract, edit setting, and required permission mode.
- Queued execution uses isolated Git worktrees by default and retains local run logs, reports, and traces for auditability.
- It supports both SDK-based and external-CLI providers, plus MCP, ACP, GitHub Actions, and pipeline-mode integration paths.
- It requires Node.js >=24.15.0, Git, and a configured provider; several supported providers also require a separately installed external CLI.
- Teams must design and maintain YAML workflows, facets, permissions, and routing configuration.
- The “Execute now” option writes directly to the current working tree rather than using the default worktree isolation.
- GitHub Issue workflows need the optional
ghCLI; GitLab integration depends onglaband remote-URL detection. - OAuth permissibility varies by provider and use case, so adopters must check the relevant provider terms.
How do you install or deploy this agent?
Use a Git repository with at least one commit, Node.js >=24.15.0, and Git. Install with npm install -g takt. For Codex SDK, create ~/.takt/config.yaml with provider: codex, model: gpt-5.6-sol, and language: en, then set export TAKT_OPENAI_API_KEY=sk-.... Nix users can run nix run github:nrslib/takt or install with nix profile install github:nrslib/takt.
How do you use this agent?
Run takt, describe the task, type /go, and select “Queue as task.” Run takt run to execute queued work in isolated worktrees. Use takt list to inspect task branches and merge, retry, or requeue them; use takt watch for a resident queue-monitoring process. To create and validate a custom workflow, run takt workflow init my-flow followed by takt workflow doctor my-flow.
How does this agent compare with similar options?
Compared with using plain AI coding agents directly, TAKT places process control in YAML: review and repair loops are explicit transitions, steps receive focused context, and runs leave logs and reports. Direct agent use instead relies chiefly on prompts and agent behavior to preserve the process.
FAQ
Does TAKT replace my coding agent?
Can I run it without an external provider CLI?
claude-sdk, codex, and opencode, which use Node.js and API keys. CLI-based providers such as Claude Code, Copilot, Cursor, and Kiro require their external CLIs.What is stored locally?
.takt/runs/<run>/. Routing decisions are written locally as NDJSON under .takt/events/ by default, and the README states they are not uploaded.What happens when a review finds an issue?
COMPLETE or ABORT, and workflows can request human judgment when needed.