agentboards.org

TinyAGI

#154 agent harnessunverified rowv0.0.20

Runs multiple teams of AI agents in isolated workspaces around Claude Code and Codex, reachable from Discord, WhatsApp and Telegram

Key differences

Runs multiple teams of AI agents in isolated workspaces around Claude Code and Codex, reachable from Discord, WhatsApp and Telegram

  • Runs local. Free and MIT-licensed; you supply Anthropic, OpenAI or custom provider keys, stored per provider inside TinyAGI
  • Includes a Docker sandbox. Listed for 48 of 194 tools in this category.
  • Supports headless CI workflows. Listed for 60 of 194 tools in this category.
  • Keep in mind: Custom providers accept any OpenAI- or Anthropic-compatible endpoint, which covers a local server, though no specific local runtime is named.

“The dashboard includes an org chart for your bots, so the reorg is now a configuration change.”

Website Docs 3.6k starsCompare vs…Dispute a fact
Appeal a claim or request ownership transfer

What it is

TinyAGI (formerly TinyClaw) is a Node daemon that runs several agents with specialized roles in isolated workspaces and lets them hand work to teammates through chain execution and fan-out, with persistent per-team chat rooms and an SQLite queue providing atomic transactions, retries and a dead-letter path. It drives the Claude Code and Codex CLIs or any OpenAI- or Anthropic-compatible endpoint, runs 24/7 as a background process or Docker container, and ships TinyOffice, a browser dashboard with a kanban task board, org chart and live event feed. The README labels the project experimental.

Specification

Source verification

Row snapshot checked not yet. Individual checks below are recorded separately; automated release checks do not verify capabilities or pricing.

readme
Needs individual review
install
Needs individual review
models
Needs individual review
license
Needs individual review

Architecture

Type
Agent harness
Runsunsourced
local
Platforms
macos, linux, windows
Context windowsrc ↗
not documented
Languages
any

Models

Backbonesrc ↗
Claude (via Claude Code CLI), OpenAI (via Codex CLI), any OpenAI- or Anthropic-compatible endpoint
Bring your own model
Yes
Local models
Yes
Custom providers accept any OpenAI- or Anthropic-compatible endpoint, which covers a local server, though no specific local runtime is named.

Protocols

MCP clientunsourced
No
MCP server
No
OpenAPI tools
No

Capabilities

Terminal commandsunsourced
Yes
Multi-file edits
Yes
Git operations
No
Browser control
No
Sandboxed execution
Yes
Docker Compose is a documented deployment path for the daemon; agent workspaces are isolated by directory rather than by container.
Multi-agent
Yes
Headless / CI
Yes

Cost

Modelunsourced
byok
Starts at
$0/mo
Free tier
Yes
Bring your own key
Yes

Free and MIT-licensed; you supply Anthropic, OpenAI or custom provider keys, stored per provider inside TinyAGI

Openness

Open sourcesrc ↗
Yes
License
MIT
First release
unknown
open-sourcepreviewmulti-agentagent-teamsmessagingkanbanrenamed

Los Agentes on TinyAGI

Who are they?
The ruling
El JuezThe judge

El Profesor credits the queue that never loses a task and El Crítico points out that the workspaces those tasks run in are separated by directory and nothing else.

Trial only
Reasoning and trade-offs · AI analysis

El Profesor rates the delivery machinery well: transactions, retries and a dead-letter path are the parts most projects at this stage skip. El Crítico rates the isolation badly, because the workspaces those agents occupy are folders rather than boundaries, and the daemon runs continuously.

They do not contradict each other. One is describing the part that was engineered carefully and the other the part that was not, and for an always-on process the second decides. El Crítico carries it. Trial only, and the exit criterion is running it in a container against a repository you could afford to lose.

Agree with El Juez?
El AmigoThe friend

Pick it if you want to give an agent work from Telegram while away from the desk; pick a desktop harness if you would rather be in front of the diff.

6.5
Reasoning and trade-offs · AI analysis

The deciding trait is where you talk to it. Reaching your agents through Discord, WhatsApp or Telegram means work gets started from a phone, in a queue, on the way somewhere, and that fits how small requests actually occur to people. It is a genuinely different rhythm from opening a laptop to type a prompt.

The same trait is the warning: giving instructions from a chat window means giving them without seeing the code. Pick it for errands and triage. Pick a desktop tool for work you need to watch.

reliability
5
usefulness
7
cost
9
longevity
5
Agree with El Amigo?
El CríticoThe critic

Agent workspaces are isolated by directory rather than by container, and the daemon is designed to run around the clock, so separation depends on paths staying honest.

5.8
Reasoning and trade-offs · AI analysis

Directory separation is a convention, not a boundary. Nothing stops a shell command from walking up a level, and several role-specialised agents running continuously will eventually produce one that does, at an hour when nobody is reading the feed. The documented Docker path covers the daemon rather than the individual workspaces, which is the opposite of where the isolation is needed.

What it does right is fanning work out explicitly, so at least the path a task took between agents is recorded rather than inferred.

reliability
4
usefulness
6
cost
8
longevity
5
Agree with El Crítico?
El ProfesorThe professor

Task delivery runs through a SQLite queue with atomic transactions, retries and a dead-letter path, which is a specified set of semantics rather than an implicit one.

6.5
Reasoning and trade-offs · AI analysis
  1. Choosing a transactional store for the work queue means the failure behaviour is defined: either a handoff committed or it did not, and a task that cannot be processed lands somewhere nameable instead of disappearing. Most projects at this maturity keep the queue in memory and discover the consequences later. 2. Retries with a terminal state also bound the pathological case where two agents pass the same item forever.

  2. No evaluation is offered, and the claims here are about plumbing, which does not require one.

reliability
7
usefulness
6
cost
7
longevity
6
Agree with El Profesor?
La InversoraThe investor

3,611 stars under a project that renamed itself from TinyClaw, with no company, no paid tier and an experimental badge on the front page.

5.8
Reasoning and trade-offs · AI analysis

A rename this early usually means either a trademark conversation or a change of ambition, and neither is free: the audience fragments across two names and the search results take a year to catch up. Thousands of stars accumulated anyway, which says the idea has pull and says nothing about who pays for the next year of it.

Moat: none. Likely path: a maintainer's enthusiasm project that either finds a sponsor or slows to occasional commits. Position: watch it, do not wire a team process into something still choosing its own name.

reliability
5
usefulness
6
cost
8
longevity
4
Agree with La Inversora?
La JefaThe CTO

It keeps provider keys per provider inside itself, which is a credential store nobody in my organisation reviewed, and Windows engineers need WSL2 before they start.

5.0
Reasoning and trade-offs · AI analysis

Any component that holds secrets becomes part of my security review whether it wants to or not. Keys stored inside a self-installed daemon on a developer machine sit outside our vault, outside our rotation schedule and outside the audit that says who can decrypt what. That is the whole conversation.

The platform note adds a second cost, since the Windows population needs a subsystem installed first and supported afterwards. There is no directory integration and no policy surface at any seat count. Not yet.

reliability
4
usefulness
5
cost
7
longevity
4
Agree with La Jefa?
El HackerThe tinkerer

MIT, docker compose up, and any OpenAI- or Anthropic-compatible endpoint accepted as a custom provider, which covers the server already running on my network.

7.3
Reasoning and trade-offs · AI analysis

Accepting an arbitrary compatible endpoint is the minimum I ask for and a surprising number of projects still fail it. Here the provider is a configuration entry, so my own inference host is as valid as anyone's API, and nothing forces traffic through a vendor I did not choose.

A compose file gets the daemon up in one command, and the permissive licence keeps a fork legal when the maintainers wander off. What I miss is protocol support: no MCP at either end, so every tool I want is glue I write myself.

reliability
7
usefulness
7
cost
9
longevity
6
Agree with El Hacker?