Tale — Shared Workspace for Teams and AI Agents
Turn company problems into tasks your team and AI agents solve together, then review the reports and files they deliver.
- Source repo
- tale-project/tale
- Stars
- ★ 47
- Last updated
- 5d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 65/100 · Some gaps
At a glance
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
- You'll need
- Typical use
- A team reviewing a launch brief: create one project task with the brief and meeting notes, and have an agent produce a Markdown report of conflicting dates, missing owners and open decisions with citations.
- Main limitation
- Self-hosting is not turnkey: production needs DNS, TLS, backups and access controls on top of Docker and Compose, plus several GB for images and data.
- Website
- tale.dev ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-16latest release v0.5.107 (2026-11-16)+0.6 ★/day over 7 days
- Source review
- 65/100 · Some gaps
What does this agent do, and when should you use it?
Tale is an open-source project workspace where teammates and AI agents share a task board, project knowledge, agent configuration and review flow. You create a project task with the problem statement, source files and completion criteria, then assign it to a person or a project agent whose instructions, harness, model and tools you configure individually. Agents run in persistent sandbox workspaces, with concurrency limited by your configured capacity, and a manager agent can delegate ready work when granted the required tools. Around the board, the platform provides Chat with an Arena mode that compares two model responses side by side, Knowledge for retrieval-ready documents and web content, versioned Automations workflows, and Connectors backed by credentials your workspace controls. Administrators manage members, roles, providers, policies, usage and audit records. You can self-host it with Docker Compose on your own infrastructure or use the managed Cloud service; the code is MIT licensed.
Tale reads the task brief, source files and completion criteria you attach to a project task and hands the work to a configured project agent. Before running, you choose that agent's harness (Claude Code, Codex, Cursor, Gemini CLI, Hermes, OpenClaw, OpenCode, Pi, Qwen Code and others), its model and its tools; the agent then executes in a persistent sandbox workspace and returns reports and delivered files to the board for review. You follow progress on the board, steer the agent with @mentions in task comments, inspect the reports and files, and either ask for corrections or accept the result; a manager agent with the right tools can delegate ready tasks and coordinate follow-up work. The wider platform adds Chat and Arena for comparing two model responses, Knowledge for preparing documents, knowledge entries and website content for retrieval (upload and indexing are separate steps), versioned Automations that connect steps and run manually or through configured triggers, Connectors for external services using workspace-controlled credentials, and Administration for members, roles, providers, policies, usage and audit records.
- A team reviewing a launch brief: create one project task with the brief and meeting notes, and have an agent produce a Markdown report of conflicting dates, missing owners and open decisions with citations.
- Engineering teams that want agents to actually work in a repository or build an internal tool: assign the task to a project agent running Claude Code or Codex in a sandbox, then review the delivered files.
- A project manager who needs one shared board where human-assigned and agent-assigned cards sit in the same status columns with their deliverables.
- Knowledge-heavy teams turning documents, knowledge entries and website content into retrieval sources and checking the cited sources behind answers.
- An operations team repeating a defined process as a versioned Automation that starts on a schedule or through connector approvals, then reviewing each run's steps and inputs.
- An administrator consolidating governance: members, roles, providers, policies, usage and audit records, plus Guardrails pages for content-safety and data policy status.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://raw.githubusercontent.com/tale-project/tale/main/scripts/install-cli.sh | bashbun install --frozen-lockfile
- Install notes cover
- macOS · Linux · Docker
How do you install or deploy this agent?
The documented local path uses the published CLI on macOS or Linux; no repository clone and no Bun installation are needed:
curl -fsSL https://raw.githubusercontent.com/tale-project/tale/main/scripts/install-cli.sh | bash
tale init my-project
cd my-project
tale devYou need Docker with Compose, several GB of space for images and data, and a model provider credential for your first reply. On Apple Silicon, Docker Desktop's amd64 emulation is needed by the bundled object store; on ARM64 Linux, configure emulation before starting. The CLI can help install or start Docker.
To develop from source you need the Bun version pinned in package.json, a compatible Node.js runtime, and Docker for backing services; full repository checks also need Python and uv:
bun install --frozen-lockfile
bun run setup:check
bun run devDocumentation-only work needs no platform database or provider:
bun run --filter @tale/docs dev
bun run --filter @tale/ui-docs devHow do you use this agent?
After tale dev starts, open the printed URL, create the first account and organization, add a credential under Settings > AI providers, and send your first message. Press Ctrl-C to stop; running tale dev again in the same directory resumes with your data.
A typical agent flow: create a project task with the problem, source files and completion criteria; configure a project agent (instructions, harness, model, tools); start the task and read the agent's report. An illustrative brief for a launch-review task looks like this:
Compare the attached launch brief and meeting notes. Produce a Markdown report listing conflicting dates, missing owners, and open decisions. Cite the source file and passage for each finding. Separate confirmed facts from questions, and leave the source files unchanged.Before accepting a result, open the delivered report, verify its citations against both files, and check that every requested category is covered; when evidence is missing, ask for corrections in the task or follow the task review guide. Before serving a team, complete production preparation — tale deploy uses separate data volumes from the local development instance.
What are this agent's strengths and limitations?
- People and agents share one workspace: the board, @mention steering in task comments, and report/file review all happen in place instead of across separate tools.
- Per-agent runtime choice is explicit — Claude Code, Codex, Cursor, Gemini CLI, Hermes, OpenClaw, OpenCode, Pi, Qwen Code — alongside model and tool selection.
- Agents execute in persistent sandbox workspaces with concurrency bounded by configured capacity, giving isolation and an explicit capacity limit.
- Bring your own provider API keys (routed through Tale's model gateway) or use supported vendor subscriptions with compatible runtimes.
- Two deployment paths with the same feature set: MIT-licensed Docker Compose self-hosting, or the managed Cloud service; Enterprise adds operation and support rather than features.
- Self-hosting is not turnkey: production needs DNS, TLS, backups and access controls on top of Docker and Compose, plus several GB for images and data.
- Vendor subscriptions only work with compatible runtimes, cannot power ordinary Chat, and bypass Tale's gateway metering and spending caps, so cost visibility suffers.
- Self-hosting does not automatically keep data local: model providers, connectors and external tools can process data outside Tale's stores, so residency has to be configured deliberately.
- Developer setup pulls in a mix of toolchains (Bun, Node.js, Docker, Python, uv), which raises the bar for source work.
- Runtime availability depends on your deployment, credentials and sandbox capacity rather than being guaranteed out of the box.
- Only four screenshots and the linked docs describe the UI; there is no public hosted demo in the repository itself.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Tale — Shared Workspace for Teams and AI Agents This agent | 65 · Some gaps | — | ★ 47 | 5d ago | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| Brigade | 53 · Major gaps | Self-hosted serviceFreemium | ★ 16 | today | TypeScript | ChatGPT · Codex · Claude Code · Claude.ai |
| Atom Platform | 67 · Some gaps | CLIFree + model costs | ★ 912 | today | Python | OpenAI API · Claude API |
| OpenCompany | 64 · Some gaps | Desktop appFree + model costs | ★ 1.1k | 1d ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence says agents can be configured with a runtime, model, skills, and tools, and work in persistent sandboxes; it describes capacity limits, human review, and approvals for connector actions. The README also warns that external tools and model providers may process data, and that subscription credentials can bypass gateway usage and spending limits. Deductions: the packet does not show granular permission controls, secret storage/redaction implementation, or a dependency security audit. Confirmation and rollback are described mainly as workflow guidance, not as enforced safeguards for every side effect.
The CLI routes uncaught exceptions and rejections through a shared error handler. The repository declares format, lint, type, build, and test gates, and the workflow sample corroborates some CI structure. Deductions: static evidence does not establish that these gates passed. The README's CLI install script uses a main-branch URL, and the Playwright MCP config uses latest, weakening dependency availability and reproducibility.
The README clearly addresses team members, administrators, developers, and self-hosting operators, and gives research, document review, content creation, and development scenarios. It explains that runtime, credentials, roles, deployment, and capacity affect available capabilities. Deductions: only a sample of skills and source is provided, so trigger conditions, capability boundaries, and the full environment matrix are not established.
The README organizes material around a task workflow, FAQ, installation, development, and documentation entry points. It includes self-hosting steps, prerequisites, a sample task, and instructions for stopping and resuming a local instance; the license file states MIT. Deductions: the sample has no clear product release history or version support policy. Some limitations are disclosed but distributed across deployment, credentials, and data-residency guidance, and no specific maintenance owner is identified.
The product description and example show the intended value of assigning work, using persistent workspaces, coordinating agents, and reviewing reports and files. Deductions: the packet does not quantify time or cost savings or provide evidence that allows comparison with other workflows; costs also depend on model providers, credential paths, and execution capacity.
The README links important workflows and limitations to relevant guides, and its example asks for source-file and passage references while separating facts from open questions. The package manifest, CLI source, CI workflow, and LICENSE corroborate some static claims. Deductions: many product behaviors are only summarized in the README, with no corresponding implementation or full documentation in this sample; linked content and runtime behavior were not examined.
- Before use, check runtime and tool permissions, secret handling, and network access; self-hosting does not mean every request stays local.
- Confirm usage and spending controls for the chosen credential path: the README says supported subscription credentials may bypass the model gateway's usage tracking and spending limits.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.5.107 Tale v0.5.107
Release notes excerpt
Highlights - Sandbox sessions on Incus can spread over several hosts (experimental). The Incus sandbox backend reached exactly one host, so every session of a stack ran on that one node. The new SANDBOX_INCUS_ENDPOINTS takes a list of hosts, each with its own restricted client certificate, key and pinned server certificate, and optionally its own project, profile, network, storage pool, session cap and egress proxy. A new session goes to the host with the most room left by its session cap, its project's memory and disk limits and its pool's free space. A host marked drained takes n …View on GitHub ↗ -
v0.5.106 Tale v0.5.106
Release notes excerpt
Highlights - [Run Tale on Incus](https://docs.tale.dev/self-hosted/install/incus) (experimental). A workspace or a managed deployment can now run as Incus instances on one Linux host instead of Docker Compose containers, with no Docker daemon on that host. The services, images and settings stay the same: the tale CLI compiles the same Compose definitions into instances in one restricted Incus project, and the sandbox sessions run in a project of their own. Choose it in the reviewed declaration, never per command: runtime.engine: "incus" with a runtime.incus block in a workspace's …View on GitHub ↗ -
v0.5.105 Tale v0.5.105
Release notes excerpt
Highlights - Agents on Pi keep working through long conversations. Tale did not tell Pi how much context a model holds, so Pi took 128,000 tokens for every model it reached through Tale. Once a conversation came near that, Pi asked the model for a one-token answer and the run ended without a report ("the agent ended its turn without a final report"); a retry continued the same conversation and ended the same way. A long task stopped there, and so did a standing role whose runs continue one conversation. Pi now takes each model's context window from the model catalog, narrowed by the or …View on GitHub ↗
FAQ
What does self-hosting cost and which credentials are required?
What limits how many agents can run at once?
Does self-hosting mean no data leaves my infrastructure?
Is Community different from Enterprise?
Can I run the local development instance for my whole team?
tale deploy uses separate data volumes from the local development instance and points to production preparation (DNS, TLS, backups, access controls) before serving a team. Note too that bun run setup:check covers only part of the environment and does not prove databases, storage or a provider are configured.Is Tale — Shared Workspace for Teams and AI Agents open source?
Is Tale — Shared Workspace for Teams and AI Agents safe to use?
Is Tale — Shared Workspace for Teams and AI Agents still maintained?
Who makes Tale — Shared Workspace for Teams and AI Agents?
What is Tale — Shared Workspace for Teams and AI Agents's official website?
Thanks — we use this to improve the page.