Manifold
Coordinate teams of AI specialists on long-running, multi-step automation workflows.
- Source repo
- intelligencedev/manifold
- Stars
- ★ 504
- Last updated
- 1mo ago
- License
- MIT
- Primary language
- Go
- FA score
- 55/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Automation teams coordinating several specialists on objectives that may take hours and many steps.
- Not a fit if
- Teams requiring strong production stability guarantees
- Users unwilling to self-host and configure a model endpoint
- Website
- intelligence.dev ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-10-22latest release 2026.05.14 (2026-05-14)
- Source review
- 55/100 · Major gaps
What does this agent do, and when should you use it?
Manifold is an experimental platform for pursuing complex objectives with teams of configurable AI specialists over extended periods. Its interfaces include agent chat, a visual workflow editor, a specialist registry, isolated project workspaces, Pulse scheduling, and a playground for prompts, datasets, and experiments. It can call OpenAI, Google, and Anthropic models, as well as OpenAI-compatible endpoints backed by llama.cpp or vLLM. Built-in and MCP tools participate in workflows, while a saved workflow can itself become a specialist tool or a node inside another workflow. The recommended deployment is a self-hosted Docker Compose service accessed through a browser; a locally built binary can instead use SQLite, FTS5, Vec1, and process-local telemetry. The project is explicitly experimental and is not presented as production-ready for environments demanding strong stability guarantees.
Users define AI specialists in the Specialist registry and configure Projects as isolated workspaces. Agents load project skills from each workspace's skills/ directory and can discover universal read-only skills in $HOME/.manifold/skills and $HOME/.agents/skills through dedicated skill tools. Agent chat accepts objectives, supports collaboration across multiple turns, and can return text or configured visualizations. The Workflow editor exposes MCP tools as nodes, saves composed workflows, and makes those workflows callable by specialists or nestable inside other workflows. Pulse runs specialist tasks at intervals, daily, or once at a specified date and time, then sends results to Matrix or channels added through Skills and MCP. Manifold calls OpenAI, Google, Anthropic, or OpenAI-compatible model endpoints and also performs image generation through OpenAI, Google, or a custom ComfyUI MCP client. Its playground versions prompts, assigns them to agents, configures datasets, and runs experiments, while local telemetry can serve bounded metrics, logs, and traces.
- Automation teams coordinating several specialists on objectives that may take hours and many steps.
- Workflow designers who want to compose MCP tools visually and reuse a subworkflow as another tool.
- Operations or research teams scheduling specialist jobs and delivering their results to Matrix.
- AI application developers comparing prompt versions and datasets through repeatable agent experiments.
- Teams choosing between hosted model providers and self-hosted llama.cpp or vLLM endpoints.
- Organizations separating multiple client or research workspaces by project root and project-specific skills.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
docker compose up -d manifold
- Ports
32180- Environment variables
OPENAI_API_KEY- Config files
.envconfig.yaml- Runtime
- Node.js 22 · Go 1.26.3
- Install notes cover
- Docker
How do you install or deploy this agent?
A basic local deployment requires Docker with Docker Compose support, an LLM API key or reachable OpenAI-compatible endpoint, and a writable absolute host path for WORKDIR. Copy the examples, set at least OPENAI_API_KEY and WORKDIR in .env, and start the service:
cp example.env .env
cp config.yaml.example config.yaml
# Edit .env and set at minimum:
# OPENAI_API_KEY=...
# WORKDIR=/absolute/path/to/your/manifold-workdir
docker compose up -d manifoldOpen http://localhost:32180 after startup. Node 22, pnpm, and Go 1.26.3 are only required when developing outside Docker or building the binary locally; a host build also needs Chrome or another Chromium-compatible browser for browser-driven tools.
How do you use this agent?
After starting the service, open the web interface at:
http://localhost:32180Define specialists in the Specialist registry, configure isolated workspaces under Projects, and assign objectives through Agent chat. For structured automation, compose the automatically exposed MCP nodes in the Workflow editor and save the result; specialists can invoke that workflow, or it can be inserted into another workflow. For recurring work, use Pulse to select an interval, daily schedule, or one-time date and time, then configure Matrix or another channel exposed through Skills or MCP. When running a locally built binary with SQLite configuration, inspect storage before startup with:
./dist/manifold storage doctor --jsonThe standard host build is:
make build-manifoldTo expose beta UI links, use either supported build command:
make build-manifold-beta
make build-manifold FEATURE_GATE=betaWhat are this agent's strengths and limitations?
- Supports OpenAI, Google, Anthropic, and OpenAI-compatible endpoints, including open-weight models served through llama.cpp or vLLM.
- Automatically turns MCP tools into visual workflow nodes and allows saved workflows to be called or nested as tools.
- Combines specialist configuration, isolated projects, collaborative chat, scheduling, image generation, and prompt experiments in one self-hosted interface.
- Can operate with SQLite, FTS5, Vec1, and bounded process-local telemetry instead of requiring external database and telemetry services.
- The project is explicitly experimental and does not promise the stability needed for production-critical environments.
- Initial deployment requires Docker Compose, configuration files, a writable
WORKDIR, and either model credentials or a compatible endpoint. - Building outside Docker adds Node 22,
pnpm, and Go 1.26.3 toolchain requirements. - Observability is still described as a work in progress, while stricter workflow enforcement, tool-error recovery, and safe compaction modes require runtime configuration.
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 |
|---|---|---|---|---|---|---|
| Manifold This agent | 55 · Major gaps | Self-hosted serviceFree | ★ 504 | 1mo ago | Go | OpenAI API · Claude API |
| Keinsaas Navigator | 72 · Some gaps | Self-hosted serviceFree + model costs | ★ 1.2k | 14d ago | TypeScript | OpenAI API · Claude API |
| Agentspan | 60 · Some gaps | CLIFree + model costs | ★ 498 | 1mo ago | TypeScript | OpenAI API · Claude API |
| Agent Swarm | 62 · Some gaps | Agent plugin / skillFree + model costs | ★ 1.1k | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README describes project isolation and individually enabled MCP tools. The API skill requires confirmation before configuration changes or deletions and calls for redacting secrets; the auth guide describes httpOnly sessions and HTTPS settings for production. These support ordinary use, but the sample does not show a comprehensive permission model, data-retention policy, or rollback and recovery path. The TTS service also downloads models from HF by default. The dependency list has no audit evidence, and the release workflow creates an empty third-party notices file, so dependency security and attribution receive limited credit.
The README calls the project experimental and unsuitable for production requiring strong stability guarantees; observability is also marked as in progress. The CI file lists formatting, static checks, and tests, but this static review cannot establish that they ran successfully. Some paths or version cues are inconsistent: for example, the API skill's default URL is written as http:/localhost while the README uses http://localhost. The skill code gives clear handling for some errors, including a 400 response for an unknown voice, so failure feedback is comparatively well covered.
The README covers chat, image generation, scheduled tasks, workflows, and projects, and describes multiple model providers, compatible APIs, and local deployment options. The workflow skill defines its format and validation steps, while the API skill sets a clear scope and operating boundaries. The sample still gives limited guidance for different user audiences, deployment conditions, and feature maturity.
The README organizes features and provides Docker quick-start and developer build guidance. The auth guide includes provider configuration examples and the materials disclose experimental status, in-progress observability, and a GitHub email-mapping limitation. The MIT license text is complete. Deductions reflect distributed installation guidance with links to files not included in this sample, no visible changelog or clear maintainer/update channel, and duplicated Go contributor guidance in AGENTS.md and CLAUDE.md.
The product offers long-horizon multi-assistant work, a workflow editor, scheduled tasks, and MCP integration for a clear automation use case. The workflow skill provides an actionable document shape and validation loop. Deductions reflect the lack of quantified benefit or resource costs and limited guidance in the sample on model-call cost, operational limits, and human oversight for long-running tasks.
The README, skills, auth guide, and source provide traceable descriptions of several capabilities and configuration paths; the skill test files also show examples of interface behavior. The sample is limited, however, and many README feature claims are not corroborated here by implementation files or independent sources. CI configuration is evidence of a test plan, not evidence of executed results. Documentation distinguishes experimental and in-progress features from configuration facts, keeping unsupported inference relatively constrained.
- The project describes itself as experimental; CI configuration alone does not establish that this revision passed tests or has production-grade stability.
- The API skill writes its default address as http:/localhost:32180, which may be invalid; verify it before use.
- The release workflow creates an empty THIRD_PARTY_NOTICES.txt, and the sample shows no dependency security audit or complete third-party attribution.
- The TTS service permits automatic model downloads by default; confirm that network access and model sourcing suit the deployment environment.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * Adds Pulse tasks view by @Art9681 in https://github.com/intelligencedev/manifold/pull/166 Full Changelog: https://github.com/intelligencedev/manifold/compare/2026.05.8...2026.05.14View on GitHub ↗ -
Release notes excerpt
Release for master merge PR #165. What's Changed * Improves rolling summary, faster generation when local provider used by @Art9681 in https://github.com/intelligencedev/manifold/pull/165 Full Changelog: https://github.com/intelligencedev/manifold/compare/2026.05.6.2...2026.05.8View on GitHub ↗ -
Release notes excerpt
Release Summary This release ships the fix from PR #164: local provider tokenizer and summary flow corrections for self-hosted OpenAI-compatible backends. Highlights - Fixed the plain-text summary reducer so it uses the effective summary prompt budget instead of over-splitting work into many small serial summary calls. - Added a per-call summary timeout with a safe fallback that preserves prior summary state if the summary model stalls. - Stopped self-hosted chat completions from falling back to extra /tokenize requests when the provider already returns usage metrics. …View on GitHub ↗
FAQ
Must I pay for a commercial model API?
Does a deployment require an external database or telemetry stack?
Are project files separated?
skills/ directory, while universal skills in the two documented user directories are discovered as read-only resources.Where can scheduled task results be sent?
Is it suitable for a production-critical deployment?
Is Manifold open source?
Is Manifold safe to use?
Is Manifold still maintained?
Who makes Manifold?
What is Manifold's official website?
Thanks — we use this to improve the page.