AI engineering terminal for Toolkit-LLM

Delegate repo work without leaving your terminal.

Toolkode is a CLI that turns plain-English engineering tasks into specs, plans, isolated agent work, reviewable diffs, and verification runs.

$npm i -g @toolkit-cli/toolkode
$curl -fsSL https://toolkode.com/install | bash
see the workflowopen browser app
~/repo

$ toolkode

tk> /specify add usage-based billing with audit receipts

created spec.md with acceptance criteria, audit boundaries, and rollback notes

tk> /plan

selected server-side meter path, idempotent receipts, guarded destructive retries

tk> /implement --agents=builder,reviewer,red-team

captainbuilderreviewerred-team
diff: 7 filestests: passingreview: 2 risks noted
200+commands across planning, build, review, and operations
14agent roles for research, implementation, audit, and recovery
gitisolated worktrees and visible file ownership
localOllama and OpenAI-compatible endpoints alongside Toolkit-LLM

From ask to evidence

Toolkode makes agents show their work.

The thing engineers care about is whether the agent can understand the task, isolate the work, leave a reviewable diff, and prove what changed.

A task becomes a trail.

Toolkode keeps the loop explicit: specify, plan, task, implement, test, review. The value is not another prompt box. It is the evidence left behind when agents touch a real repo.

tk> /specify the smallest safe fix for this issuescope
tk> /planapproach
tk> /taskswork units
tk> /implement --agents=builder,reviewerdiff
tk> /test && /reviewproof

Command surface

200+ commands, grouped around the engineering loop.

The number matters because it means coverage. Toolkode has commands for planning, building, verifying, coordinating agents, watching usage, and moving work toward release.

Plan

/specify/plan/tasks/clarify

Turn the rough ask into a spec, an approach, and reviewable chunks before files change.

Build

/implement/fix/refactor/edit

Land scoped changes with the repo context, conventions, and constraints already loaded.

Verify

/test/typecheck/review/audit

Make the agent produce evidence, argue with the diff, and surface residual risk.

Operate

/agents/worktrees/usage/deploy

Coordinate longer runs, inspect ownership, watch cost, and move toward release.

Capsule zero-SaaS stack

Remove SaaS control planes from the infra stack.

Capsule is the Toolkode runtime path for agents that need more than code edits. The default development stack is local and project-scoped: auth, users, orgs, data, Git, objects, jobs, billing ledger, analytics, previews, deployments, and receipts are inspectable without making a vendor dashboard the first blocker.

0 required vendor control planesuser-owned adapters onlylocal project objectsreceipt-backed gates
capsule project mapvendorControlPlanes: 0
authusers, orgs, sessions
databranches, databases, backups
gitrepos, commits, hooks
jobsqueues, runs, receipts
billingledger, invoices, rollback
deploylocal, staging, prod lanes

$ capsule parity --json

auth: ready · data: ready · git: private · adapters: explicit

GET /api/projects/capsule-app/parity

nextGate: prove receipts before production handoff

Inspectable sessions

CLI first. Browser when the session gets long.

The browser app is a control surface for permissions, sessions, diffs, model selection, and long-running work. It complements the CLI instead of replacing it.

toolkode session
permission request

builder wants to edit billing/meter.ts and run npm test

diff review

+312 -84 across 7 files

  • idempotency guard added
  • receipt schema migrated
  • rollback note required
model route

reviewer: local ollama

builder: Toolkit-LLM managed route

Pricing posture

Start locally. Add managed routing when you need it.

The CLI remains useful with local config and bring-your-own providers. Managed Toolkit-LLM routes are for teams that want shared controls and model access without stitching every provider together themselves.

Local / BYO

$0

CLI, toolkode.json, Ollama, OpenAI-compatible endpoints, private-by-default workflows.

Managed Toolkit

team

Toolkit-LLM access, routing, shared controls, and longer-running agent workflows.

Run it on a real repo

Install Toolkode, then make it prove itself.

$npm i -g @toolkit-cli/toolkode
$curl -fsSL https://toolkode.com/install | bash

tk> /specify the smallest safe fix for this issue