# AgentsKit — Composable MIT-licensed TypeScript packages for agent runtime, tools, memory, RAG and chat UI.

> Source: The Agents Index — https://theagentsindex.com/agentskit (structured, researched, re-verified)
> Facts last verified: 2026-10-04

AgentsKit is an open-source TypeScript library for building AI agents from six contracts: adapter, tool, skill, memory, retriever and runtime. Developers install only the packages they need and swap models, tools, memory or interfaces without rewriting the agent.

| Fact | Value |
| --- | --- |
| Website | https://agentskit.io |
| Price | The listed products are free and open source. |
| Free tier | Yes |

No paid tier at all. MIT-licensed packages installed from the @agentskit npm org with no account, seat or usage cap; your only cost is the model provider's inference, and a demo adapter runs with no API key.

## What it does

**Writes code** · Yes — A bundled coder skill writes real multi-file TypeScript against scoped filesystem and allowlisted shell tools, so the agent you assemble authors code rather than snippets.

> Implements features from specs — TypeScript-first, TDD-leaning, opinionated about code quality.
> — https://www.agentskit.io/docs/agents/skills/coder · Page subtitle under the heading 'coder'

**Debugs and fixes** · Partly — The docs describe a self-debug pattern where agents read their own traces and retry with corrections, but it is a primitive you wire up, not a product that lands a verified fix.

> Agents that read their own traces, diagnose failures, retry with corrections.
> — https://www.agentskit.io/llms.txt · Agents section, Self-debug entry

**Writes tests** · Partly — The coder skill's documented behaviour includes writing tests beside the implementation under Vitest conventions, but test generation is skill behaviour rather than a guaranteed product feature.

> Tests alongside code (Vitest conventions).
> — https://www.agentskit.io/docs/agents/skills/coder · Behavior bullet list

**Reviews code** · Yes — A codeReviewer persona ships with the skills package and reviews diffs for bugs, security and style in pull-request comment format; you host the review loop yourself.

> Reviews diffs for bugs, security, style. Comments in PR-review format.
> — https://www.agentskit.io/llms.txt · Agents section, codeReviewer entry

**Whole-codebase work** · Partly — The coder skill is documented for multi-file edits, package additions and refactors, but nothing on the site claims a migration carried across a whole repository or repo-wide indexing.

> Multi-file edits, package additions, refactors.
> — https://www.agentskit.io/docs/agents/skills/coder · 'When to reach for it' bullet list

**Ships it** · Partly — An allowlisted shell tool can run git and package-manager commands, so shipping is possible but you build the deploy path; the library itself does not deploy or open pull requests.

> `shell` (allowlisted) | `pnpm`, `node`, `git`.
> — https://www.agentskit.io/docs/agents/skills/coder · 'Tools it expects' table, shell row

**Done-for-you service** · No — There is no services engagement here: AgentsKit is distributed as npm packages you install and run yourself, described explicitly as kit-shaped rather than a managed build.

> AgentsKit is kit-shaped: install one package, grow into the full stack, no required project layout.
> — https://www.agentskit.io/docs/compare · 'vs Mastra' section

## How much it does unattended

**Runs autonomously** · Yes — The runtime package loops observe-think-act unattended until the model returns a final answer or hits maxSteps, with a default ceiling of ten iterations.

> It runs a ReAct loop — observe, think, act — until the model produces a final answer or a step limit is reached.
> — https://www.agentskit.io/docs/agents/runtime · Intro paragraph under 'Runtime'

**Multi-agent** · Yes — Configured delegates become auto-generated delegate_<name> tools the parent model calls; each child runs its own ReAct loop with its own skill, tools and optional different model.

> When you configure `delegates`, the runtime auto-generates tools named `delegate_<name>`. The parent LLM calls them like any other tool. Each delegate runs its own ReAct loop and returns a result.
> — https://www.agentskit.io/docs/agents/delegation · 'How It Works' section

**Agent permissions** · Yes — Any tool can be marked as requiring confirmation, which pauses the run and waits for an explicit approve or deny, and a sandbox policy layer adds allow/deny lists on top.

> Runtime pauses on invocation and emits `tool.awaiting-approval`. Resume with `chat.approve(toolCallId)` or `chat.deny(toolCallId, reason)`.
> — https://www.agentskit.io/docs/agents/hitl · 'Gate a tool' section, paragraph after the code sample

## Which models it runs on

**Claude** · Yes — A native Anthropic adapter covers the Claude family with streaming, tool calls, vision and long context, and the homepage demo configures Claude as the runtime adapter.

> Anthropic chat adapter — Claude Opus / Sonnet / Haiku. Streaming, tool-calls, vision, long-context.
> — https://www.agentskit.io/llms.txt · Data section, anthropic provider entry

**GPT** · Yes — A native OpenAI adapter is documented for the GPT and o-series models with streaming, parallel tool calls and multimodal input, alongside an Azure OpenAI adapter for enterprise tenants.

> OpenAI chat adapter — GPT-4o, o-series, GPT-5. Streaming, tool-calls, parallel tools, multimodal.
> — https://www.agentskit.io/llms.txt · Data section, openai provider entry

**Open models** · Yes — Four local-runtime adapters point at Ollama, LM Studio, vLLM and llama.cpp on localhost, so open-weight models run fully offline with no vendor inference in the path.

> Run fully offline — Ollama, LM Studio, vLLM, llama.cpp.
> — https://www.agentskit.io/docs/data/providers/local · Page subtitle under 'Local runtimes'

**Model choice** · Yes — You choose the provider: hosted, local or a custom adapter all satisfy one contract, and the vendor's pitch is that changing model does not change the agent.

> OpenAI, Anthropic, Gemini, Kimi, local models, or your own adapter use one contract.
> — https://www.agentskit.io · 'Change the model. Keep the agent.' section

## Where you use it

**In your editor** · Partly — The docs give three DIY patterns for running an agent inside VS Code or Raycast, but no editor extension is shipped and the LSP route is deferred.

> Run AgentsKit agents inside your editor or launcher. Three integration patterns + a worked example for each.
> — https://www.agentskit.io/docs/production/embedded · VS Code · Raycast · embedded page, page description

**On the command line** · Yes — A first-party CLI covers scaffolding, interactive chat, one-shot runs, a dev server and environment diagnosis, and an Ink binding gives a terminal chat UI.

> Nine commands wrapping every part of AgentsKit — init, chat, run, dev, doctor, ai, tunnel, rag, config.
> — https://www.agentskit.io/llms.txt · Production section, CLI entry

**In your pipeline** · Yes — Eval results export as JUnit XML, Markdown and GitHub annotations so failures block pull requests, and recorded sessions replay in CI without network calls.

> Write eval results as JUnit XML, Markdown, or GitHub annotations so failures block PRs and surface in pull request checks.
> — https://www.agentskit.io/docs/production/evals/ci · Page subtitle under 'CI reporters'

**In a browser** · Partly — Headless chat bindings for seven frameworks let you build a web surface, but the hosted product chat experience is a separate sibling site rather than part of this library.

> Product chat UI is AgentsKit Chat (sibling), not this monorepo alone.
> — https://www.agentskit.io/llms.txt · Opening summary blockquote

## Whose machine it runs on

**Self-hosted** · Yes — Everything runs on infrastructure you own — npm packages in your Node, Bun, Deno or edge process, optionally against models served from localhost with no vendor cloud involved.

> Run fully offline — Ollama, LM Studio, vLLM, llama.cpp.
> — https://www.agentskit.io/docs/data/providers/local · Page subtitle under 'Local runtimes'

**Open source** · Yes — The vendor's own comparison table records the licence as MIT and the site states the listed products are free and open source, with packages published under the @agentskit npm org.

> License | MIT
> — https://www.agentskit.io/docs/compare · TL;DR comparison table, License row, AgentsKit column

## What it costs to run

**How it meters** · Partly — AgentsKit itself bills nothing: your only spend is the model provider's tokens, which the cost-guard observer caps with a hard dollar ceiling per run.

> Set a hard dollar ceiling per run and track token usage with either a zero-dep heuristic or provider-accurate counters.
> — https://www.agentskit.io/llms.txt · Production section, Cost + token accounting entry

**Free tier** · Yes — No call, account or card: the packages are free and open source on npm, and a demo adapter lets you run the runtime without any provider API key.

> The listed products are free and open source.
> — https://www.agentskit.io/llms.txt · 'Access and commercial model' section

**API access** · Yes — The product is itself a programmatic surface — typed TypeScript packages — and even the CLI's subcommands are exposed as named exports you can drive from your own code.

> Drive every agentskit subcommand from your own code — same internals as the CLI, exposed as named exports.
> — https://www.agentskit.io/llms.txt · Production section, Programmatic API entry

**MCP server** · Yes — The bridge works both ways: it consumes third-party MCP servers as tools and publishes your tools (and a live runtime inspector) as an MCP server over stdio.

> Consume or publish Model Context Protocol tools. Interop with Claude Desktop, Cursor, Codex, OpenClaw, and other MCP hosts.
> — https://www.agentskit.io/docs/agents/tools/mcp · Page subtitle under 'MCP bridge'

**Bring your own key** · Yes — Inference is always yours: adapters take your provider keys from your own environment, the CLI doctor command checks them, and local adapters need no key at all.

> Diagnose your environment. Providers, keys, tooling, versions.
> — https://www.agentskit.io/llms.txt · Production section, agentskit doctor entry

## Buying it for a team

**A company can buy it** · Partly — There is nothing to buy for the library itself; the managed product AKOS is the commercial path and is sales-led.

> Want a managed product rather than wiring this yourself? AKOS — the operating system for AI agents in production — packages governed flows, sandboxing, signed audit, and cost-metering on top of the library.
> — https://www.agentskit.io/docs/production · Production overview, callout above "Rollout"

**Pooled budget** · Partly — multiTenantCostGuard pools a dollar budget per tenant with a default for unlisted ones; there is no seat concept to share a balance across.

> Same accounting partitioned by tenant for SaaS deployments.
> — https://www.agentskit.io/docs/production/observability/cost-guard · Cost + token accounting page, multiTenantCostGuard section

**Admin controls** · Yes — A mandatory sandbox wraps every tool with allow lists, deny lists and per-tool validators, and the policy is snapshotted so later callers cannot widen it.

> calls that match the deny list, miss the allow list, or fail a validator are rejected before execution.
> — https://www.agentskit.io/docs/production/security/mandatory-sandbox · Intro paragraph under 'Mandatory sandbox'

**Audit log** · Yes — Runs can be recorded to a hash-chained, HMAC-signed log whose verify() throws on tamper; it is a library you attach as an observer, not a hosted admin console.

> writes each event with a sequence number, timestamp, HMAC, and a hash of the previous record — so tampering with any entry breaks the chain and `verify()` throws.
> — https://www.agentskit.io/docs/production/observability/audit-log · Intro paragraph under 'Signed audit log'

**Single sign-on** · Partly — OIDC ID tokens are verified in-library with WebCrypto, but SAML signature checking is explicitly delegated to an XML-DSig library you choose, so SAML is only half-shipped.

> SAML requires an XML / XML-DSig validator — bring your own.
> — https://www.agentskit.io/docs/production/security/sso · 'SAML assertions' section, first paragraph

## What happens to your code

**Data residency** · Partly — Jurisdiction is yours to choose because nothing is hosted by the vendor, and the router can pin selection to a region, but no residency guarantee is published.

> Require all selected adapters to match this data-residency region.
> — https://www.agentskit.io/docs/api/adapters/interfaces/RouterOptions · RouterOptions interface reference, region property

**Getting out** · Yes — There is nothing to exit from: MIT packages, no credits and no vendor store, with provider lock-in explicitly answered as none because every layer is a swappable contract.

> Provider lock-in | None (contract adapters)
> — https://www.agentskit.io/docs/compare · TL;DR comparison table, Provider lock-in row, AgentsKit column

**Certifications** · Partly — No vendor certification is published, which fits a library with no hosted service; what ships is compliance machinery — a tamper-evident log and erasure helpers.

> You want a tamper-evident audit log (SOC 2 / HIPAA friendly).
> — https://www.agentskit.io/docs/reference/packages/observability · @agentskit/observability package page, When to reach for it list

_Not established: Opt out of training_

## Provenance

Every fact above comes from the vendor. Nothing here is independently corroborated yet.

## Coverage elsewhere

Pages about this product that the vendor does not control. The label states how independent each one is.

| Source | Kind | Read |
| --- | --- | --- |
| mcpservers.org | Independent source | 2026-08-20 |

## What changed

- 2026-10-04 — Capabilities · Commercial terms · Free tier · Price
- 2026-09-09 — Capabilities · Commercial terms · Free tier · Price
- 2026-09-02 — Capabilities · Commercial terms · Free tier
- 2026-08-31 — Capabilities · Commercial terms · Free tier · Links
- 2026-08-30 — Commercial terms · Links
- 2026-08-20 — One-liner · Summary

## Correcting this record

Anything marked "Not established" is a question we have not answered — it is not a No. If you represent AgentsKit, or anything here is wrong or out of date, you can supply it with a source: the operations this site accepts from an agent are listed at https://theagentsindex.com/.well-known/agents.json — that catalogue needs no key to read, and each entry states its own auth.

Supplying evidence is free. Nothing that can be bought here changes what a field says, whether a gap is closed, or how this record is scored.