RELATED 01
Ruflo
Agent swarms with shared vector memory and task routing for Claude Code and other MCP clients
81kGitHub stars
87k/wknpm downloads
89popularity score
Agentic Flow
Agent runner, ReasoningBank pattern memory, local code edits and graph solvers for coding assistants over MCP.
By rUv (individual) TypeScript Local · stdio AI and MLDeveloper tools
83 popularity score ★ 838 ↓ 152k/wk
Connect your client
Local (npm)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Claude CodeTerminal
claude mcp add agentic-flow -- npx -y agentic-flow mcp startClaude Desktopclaude_desktop_config.json
{
"mcpServers": {
"agentic-flow": {
"command": "npx",
"args": [
"-y",
"agentic-flow",
"mcp",
"start"
]
}
}
}Cursor~/.cursor/mcp.json
{
"mcpServers": {
"agentic-flow": {
"command": "npx",
"args": [
"-y",
"agentic-flow",
"mcp",
"start"
]
}
}
}VS Code.vscode/mcp.json
{
"servers": {
"agentic-flow": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"agentic-flow",
"mcp",
"start"
]
}
}
}Codex~/.codex/config.toml
[mcp_servers.agentic-flow]
command = "npx"
args = ["-y", "agentic-flow", "mcp", "start"]Gemini CLITerminal
gemini mcp add --scope user agentic-flow npx -- -y agentic-flow mcp startSettings 6
AGENTIC_FLOW_MCP_PROFILE |
Optional: Which tools get registered: readonly, retrieval, learning or admin. |
|---|---|
AGENTIC_FLOW_MCP_ENABLE_DARWIN |
Optional: Set to true with the admin profile to register harness_repair. |
ANTHROPIC_API_KEY secret |
Optional: Key for the default Anthropic provider when running agents. |
OPENROUTER_API_KEY secret |
Optional: Key for the openrouter provider. |
GOOGLE_GEMINI_API_KEY secret |
Optional: Key for the gemini provider. |
AGENTIC_FLOW_PROMPT_INJECTION |
Optional: Prompt-injection policy (block, strict or warn) that prompt_injection_scan uses unless given one. |
What it does
Agentic Flow is a Node.js toolkit for running specialised coding agents with learned routing and a memory of past results (ReasoningBank on AgentDB). Running npx agentic-flow mcp start launches its stdio MCP server.
The README advertises 213 MCP tools, but the server that command starts (dist/mcp/standalone-stdio.js) registers 24. They cover listing, inspecting and creating agent definitions, running one agent on a task, picking a model for a task, storing and searching reasoning patterns, applying code edits with the local Agent Booster engine, building and checking signed manifests of harness files, a Darwin repair loop over a repository, and a few pure computations: splitting a task graph between two teams, linear solves, PageRank and a prompt-injection scan. The swarm, neural and GitHub tools the README names as examples are not among them.
What a client actually sees depends on AGENTIC_FLOW_MCP_PROFILE. Left unset it is readonly, and only the 15 tools that list, inspect, search or compute get registered. retrieval adds nothing beyond those at present, learning adds agentic_flow_agent and agentdb_pattern_store, and admin registers the rest. harness_repair also needs AGENTIC_FLOW_MCP_ENABLE_DARWIN=true.
Running an agent spawns npx --yes agentic-flow with your provider key. Anthropic is the default provider, with OpenRouter, Gemini and local ONNX as options.
What it needs
| Runtime | Node.js 20+ |
|---|---|
| Runs on | macOS, Windows, Linux |
| Anthropic | Paid planOnly for agentic_flow_agent with the default provider. OpenRouter or Gemini keys work instead, and the onnx provider needs no key. |
The default readonly profile needs no key. To run agents, set AGENTIC_FLOW_MCP_PROFILE to learning or admin and give ANTHROPIC_API_KEY (or OPENROUTER_API_KEY or GOOGLE_GEMINI_API_KEY with the matching provider). The package manifest asks for Node 20 or newer, although the README badge says 18.
Limits
- Not covered The stdio server registers 24 tools, not the 213 the README counts. swarm_init, github_pr_manage and the other README examples are not served.
- Limit AGENTIC_FLOW_MCP_PROFILE defaults to readonly, so agent runs, pattern storage and file edits are hidden until it is set to learning or admin.
- Limit harness_repair is registered only with the admin profile and AGENTIC_FLOW_MCP_ENABLE_DARWIN=true.
- Cost agentic_flow_agent calls Anthropic by default and bills your key. The onnx provider runs locally without one.
- Limit The agent tools start a second copy through npx --yes agentic-flow, which npx resolves itself and may download, rather than reusing the running server.
- Limit agentdb_clear_cache clears nothing in the code: with confirm=true it only returns a message about the cache.
- Limit The settings file written by init pins agentic-flow@3.2.0 and names the server claude-flow, an older release than the one published.
- Platform swarm_partition_tasks uses @ruvector/mincut-wasm on Node 22 and a built-in fallback on older Node.
Who it suits
Suits someone already using Agentic Flow agents in Claude Code who wants an assistant to list and inspect them, query the ReasoningBank memory or run graph and solver helpers. The default readonly profile changes nothing on disk. Opening up agent runs and file edits is a deliberate step through AGENTIC_FLOW_MCP_PROFILE. Anyone expecting the swarm and GitHub tools the README advertises will not find them in this server, and it is a community project, not one published by a model provider.
Tools 24
17 only read, 7 change things. A ? marks an optional parameter.
agentic_flow_agentchanges- Runs one named agent on a task through a chosen model provider and returns its output.
agent: stringtask: stringmodel?: stringprovider?: enumanthropicApiKey?: stringopenrouterApiKey?: stringstream?: booleantemperature?: numbermaxTokens?: numberagentsDir?: stringoutputFormat?: enumverbose?: booleantimeout?: numberretryOnError?: boolean agentic_flow_list_agents- Lists the agents the package ships.
agentic_flow_create_agentchanges- Writes a new custom agent definition under .claude/agents.
name: stringdescription: stringsystemPrompt: stringcategory?: stringtools?: array agentic_flow_list_all_agents- Lists package and local agents together, with where each comes from.
format?: enumfilterSource?: enum agentic_flow_agent_info- Shows one agent's source, category and a preview of its system prompt.
name: string agentic_flow_check_conflicts- Finds local agents that share a path with a packaged one.
agentic_flow_optimize_model- Recommends a model for an agent and task by quality, cost, speed or privacy.
agent: stringtask: stringpriority?: enummax_cost?: number agent_booster_edit_filechanges- Applies a marked-up code edit to a file inside the working directory with the local engine.
target_filepath: stringinstructions: stringcode_edit: stringlanguage?: string agent_booster_batch_editchanges- Applies several such code edits in one call.
edits: array agent_booster_parse_markdownchanges- Reads code blocks tagged with filepath= and instruction= from markdown and applies them as edits.
markdown: string agentdb_stats- Reports AgentDB table counts and memory use.
agentdb_pattern_storechanges- Saves a solved task with its reward and critique to ReasoningBank.
sessionId: stringtask: stringreward: numbersuccess: booleancritique?: stringinput?: stringoutput?: stringlatencyMs?: numbertokensUsed?: number agentdb_pattern_search- Finds stored patterns similar to a task description.
task: stringk?: numberminReward?: numberonlySuccesses?: booleanonlyFailures?: boolean agentdb_pattern_stats- Summarises rewards and critiques from recent patterns for a task.
task: stringk?: number agentdb_clear_cache- Meant to clear the query cache; the code only returns a notice.
confirm?: boolean harness_repairchanges- Runs the Darwin repair loop on a repository, keeping variants that score better on its tests.
repoRoot: stringgenerations?: integerchildren?: integerseed?: integersandboxMode?: enum harness_manifest- Builds a sha256 manifest over the given harness or agent config files.
files: arraycreatedAt?: string harness_verify- Checks an Ed25519-signed manifest and reports files that drifted on disk.
signed: object harness_capabilities- Reports which MetaHarness runtimes are available.
ecosystem_status- Shows which optional ruvnet packages are installed and what each powers.
swarm_partition_tasks- Splits a task dependency graph into two groups with the fewest links between them.
tasks: arraydependencies: array sublinear_solve- Solves a linear system Mx = b and checks the residual.
matrix: anyvector: arraytolerance?: numbermaxIterations?: integermethod?: enum sublinear_pagerank- Ranks nodes of a directed graph, optionally around seed nodes.
edges: arraynodes?: arraydamping?: numbertop?: integerseeds?: anyapproximate?: booleanepsilon?: number prompt_injection_scan- Scans text for prompt-injection patterns and returns the threats and a policy decision.
text: stringpolicy?: enum