Kody Personal Assistant Platform
A portable MCP home for an assistant’s memories, secrets, code, packages, and automations.
- Source repo
- kody-hq/kody
- Stars
- ★ 858
- Last updated
- today
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 70/100 · Some gaps
At a glance
- How it runs
- Works with
- Platform-specificChatGPT (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- An individual wants the same assistant memories, secrets, code, and automations available from multiple compatible MCP hosts.
- Not a fit if
- Teams that need to run outside Cloudflare's stack
- Users wanting a turnkey setup without OAuth and multi-worker deploys
- Teams needing a permissive license for competing commercial use
- Source review
- 70/100 · Some gaps
What does this agent do, and when should you use it?
Kody is a multi-user personal-assistant platform on Cloudflare Workers, not a general-purpose agent harness. Every signed-in user receives an isolated set of packages, jobs, secrets, memories, and related state exposed through OAuth-protected MCP endpoints. Its Nx monorepo contains a Remix 3 UI and origin entrypoint in packages/worker, plus dedicated platform, runtime, and jobs workers. MCP state resides in Durable Objects on kody-platform, while D1 supplies the database and KV supports sessions and OAuth. Instead of publishing a broad static tool catalog, Kody centers its MCP interface on search and Code Mode execute flows. It is best suited to adopters comfortable with Cloudflare infrastructure who want assistant state and automation to remain usable across compatible MCP hosts.
Requests enter through packages/worker/src/index.ts, where OAuth traffic is handled first, followed by MCP requests and static assets; remaining non-asset requests fall through to the server handler and router. kody-production serves Remix, /mcp HTTP, OAuth, email, and queues. It delegates MCP, mailbox, meter, and repo-session state to Durable Objects in kody-platform; package apps, the invoke API, and StorageRunner to kody-runtime; and cron work to kody-jobs, whose JobManager calls back through JobsHost. Client assets are bundled into packages/worker/public/ and served with the ASSETS binding, while kody.run routes package-app traffic to the runtime worker. MCP hosts discover capabilities through search and run code through Code Mode execute, with packages, jobs, secrets, and memories isolated by user.
- An individual wants the same assistant memories, secrets, code, and automations available from multiple compatible MCP hosts.
- A team needs to host personal assistants for several users while isolating each user’s packages, jobs, secrets, and memories.
- A developer prefers extending an assistant through search and Code Mode execute instead of exposing a large catalog of fixed MCP tools.
- A Cloudflare-oriented team wants an assistant backend built from Workers, D1, KV, Durable Objects, queues, and cron jobs.
- A project expects ChatGPT to be its likely primary host but wants to preserve practical interoperability with other MCP hosts.
How do you install or deploy this agent?
Local development requires Node.js 26, npm, and the Wrangler-managed Cloudflare Workers environment. Clone the repository, then run these commands from its root:
npm install
npm run devThe development server defaults to localhost:3742. If that port is occupied, the CLI selects the next free port and prints the resolved URL. Wrangler automatically manages the local Workers runtime and D1 database. The supplied material does not include a production deployment command, Cloudflare account setup procedure, or exact environment-variable list, so a complete production installation cannot be reproduced from this source alone.
How do you use this agent?
Run npm run dev and open the local URL printed by the CLI to access the Remix UI. An MCP client connects to the OAuth-protected /mcp HTTP endpoint and then uses the search and Code Mode execute flows. In production, kody.codes routes to kody-production, while the kody.run package-app zone routes to kody-runtime. The source does not provide a copyable MCP client configuration, OAuth registration values, sample search or execute payloads, or first-login credentials, so those details must be supplied before a first remote invocation.
What are this agent's strengths and limitations?
- MCP-first design explicitly aims to make assistant memories, secrets, code, and automations portable across compatible hosts.
- Per-user isolation covers packages, jobs, secrets, memories, and related state, with no privileged runtime account.
- The compact search and Code Mode execute surface avoids requiring a large static MCP tool inventory.
- The worker fleet separates origin routing, platform state, package execution, and scheduled jobs into distinct components.
- Local startup is concise, and Wrangler automatically provisions the local Workers runtime and D1 database.
- The deployment architecture is tied to Cloudflare Workers and uses D1, KV, Durable Objects, queues, and zone routing, creating substantial adaptation work for another cloud runtime.
- The repository uses FSL-1.1-ALv2 rather than an immediately permissive license; Competing Use is restricted until each version converts to Apache-2.0 after two years.
- Remix 3 is identified as beta, introducing potential upgrade and compatibility risk.
- The supplied documentation lacks complete production commands, environment-variable values, OAuth client setup, and an end-to-end first-call example.
- ChatGPT is described only as a likely primary target, and compatibility with other named MCP hosts is not individually demonstrated.
How does this agent compare with similar options?
Compared with MCP servers that expose a large collection of fixed tools, Kody explicitly favors a smaller surface centered on search and Code Mode execute. It also distinguishes itself from a general-purpose agent harness: its stated scope is a Fair Source personal-assistant platform for persistent state, packages, and automation.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Kody Personal Assistant Platform This agent | 70 · Some gaps | CLIFree + model costs | ★ 858 | today | TypeScript | — |
| Talorys | 63 · Some gaps | — | ★ 1.4k | 1d ago | TypeScript | — |
| Agentic Inbox | 72 · Some gaps | Self-hosted serviceFree + model costs | ★ 8.5k | 6mo ago | TypeScript | Claude Code |
| Mirix Personal Memory Assistant | 36 · Major gaps | Self-hosted serviceFree + model costs | ★ 3.5k | 2mo ago | Python | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The security policy makes userId-scoped access a core invariant and identifies OAuth/PKCE, tokens, secret storage, credential stripping, and runtime isolation as security boundaries. Workflows use scoped GitHub permissions, commit-pinned github-script actions, disabled checkout credential persistence, and controlled environment-based secret synchronization. Deductions apply because the supplied material does not show user confirmation before agent actions, least-privilege scopes for each production token, a complete data-retention/data-flow policy, or the secret-handling implementation. Deployment performs migrations, resource creation, secret synchronization, and multi-worker updates with SHA guards, sequencing, and health checks, but no explicit application rollback procedure is shown; backup and restore are visible only as script names. Copyright and licensing attribution are clear, but the publisher is unverified by the stated registry, so attribution is not full-score.
The README, package manifest, and deployment workflow align closely on Node 26, Cloudflare Workers, the multi-worker topology, MCP, and validation commands, supporting full self-consistency. Dependency versions, npm and engine requirements, overrides, and an audit command are handled well, but several dependencies use compatible ranges, Remix is identified as beta, and no lockfile content is supplied, so dependency availability is not full-score. Deployment uses strict shell behavior, timeouts, retries, exit-code handling, commit-SHA validation, health checks, and specific diagnostic messages, thoroughly supporting failure messaging.
The material identifies a multi-user personal-assistant audience, ChatGPT and other compatible MCP hosts, and distinct user, contributor, and local-development scenarios, but provides limited end-user task examples or deployment variants. Scope and non-goals are unusually clear: this is not a general-purpose agent harness, is MCP-first, and favors a compact search/execute surface, justifying full capability-boundary credit. Trigger precision is mostly an architectural assertion; tool schemas, invocation conditions, and ambiguity handling are absent from the supplied evidence. Cloudflare, D1, KV, Durable Objects, Node 26, and local-port expectations are explicit, but the environment is infrastructure-heavy and cross-host compatibility is qualified as practical rather than demonstrated.
The README supplies a clear repository map, topology, stack summary, and layered documentation index, while package, worker, script, and deployment names remain consistent. The quick start is adequate for ordinary local startup, but detailed environment and deployment requirements are deferred to files not included here, preventing full installation credit. Examples are mostly commands and topology, with no complete user workflow or FAQ. Scope, supported-version policy, and selected security boundaries are documented, but a comprehensive limitations treatment is not present in the supplied material. The complete FSL-1.1-ALv2 terms, copyright, use restrictions, and two-year Apache-2.0 transition justify full license credit. Versioning is weak because the root package remains 0.0.0 and no releases or changelog are provided. SECURITY.md clearly identifies a solo maintainer, latest-main support policy, and private reporting paths, so maintenance responsibility is thorough.
The repository describes a usable Remix UI, OAuth-protected MCP endpoints, memory/keys/code/automation storage, and compact search plus Code Mode execute flows, supporting practical output usability across MCP hosts. Centralized, portable agent state offers plausible marginal value over a large static tool catalog, but no concrete input/output examples, adoption evidence, or measured outcomes are supplied, preventing full scores. The multi-worker architecture, D1, KV, Durable Objects, queues, OAuth, and optional external services imply substantial deployment and operating costs; the evidence neither quantifies those costs nor compares lighter alternatives, so cost-benefit support is thin.
Major claims can be traced to concrete README, SECURITY.md, LICENSE, package.json, and deployment-workflow statements, and the stack, versions, topology, and operational controls corroborate one another across multiple files. Deductions apply because key isolation, sandboxing, encryption, and portability claims are primarily documentary assertions in this excerpt, without the corresponding implementation or tests; CI badges and script names do not establish successful execution. The prose generally distinguishes goals, conventions, and procedures, but it does not systematically label verified facts versus design intent or guarantees still requiring implementation evidence.
- This is a static review of only the supplied excerpts; no installation, tests, deployment, dependency audit, or security validation was executed.
- Production deployment can create cloud resources, migrate remote D1 databases, synchronize many secrets, and update multiple workers; verify token scopes, environment protections, recoverable backups, and rollback steps before use.
- Do not treat complete userId isolation, sandbox isolation, credential stripping, or encryption at rest as independently established; inspect the corresponding implementation and security tests.
- FSL-1.1-ALv2 restricts competing use and does not make every version immediately available under Apache-2.0; confirm the applicable date and intended use before adoption or redistribution.
- Only the latest main deployment is supported by a solo maintainer, with no maintained stable release lines, so assess upgrade risk and maintenance continuity.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v2026.10.25 Kody v2026.10.25
Release notes excerpt
Kody v2026.10.25 This release rethinks how package apps identify themselves: they now run as their publisher instead of whoever happens to visit, with signed visit tickets keeping visitor identity safe and separate. On the email side, you get reversible spam and archive triage for every inbox. Under the hood, a sweep of N+1 query fixes and a non-disruptive D1 backup lane make the platform noticeably faster and more reliable. New …View on GitHub ↗ -
v2026.10.22 Kody v2026.10.22
Release notes excerpt
Kody v2026.10.22 The signed-in workspace now feels like a real app with persistent navigation and a usage dashboard, first-time sign-up lets you choose your username, and the koala has a fresh new mark. Under the hood, a wave of reliability fixes tackles hung schedulers, stranded webhook deliveries, and MCP grants that were kicking each other out. New …View on GitHub ↗ -
v2026.10.21 Kody v2026.10.21
Release notes excerpt
Kody v2026.10.21 This release reshapes how hosted package apps are served—moving from /packages/{name} to a short /{name} path on your subdomain—while credentials gain a proper Disconnect that preserves your configuration, and fleet error diagnostics get per-error-name breakdowns for the first time. New …View on GitHub ↗