AXF

Agent eXoskeleton Framework for workspace-native agent capabilities.

AXF (Agent eXoskeleton Framework) gives teams a framework for building workspace-native agent exoskeletons: small, self-describing capabilities that encode how each codebase is built, tested, searched, diagnosed, and operated safely.

The goal is not to automate judgment. The goal is to compress repeated local ceremony into reliable workspace capabilities, so agents can spend more attention on the actual problem.

When AXF helps

• agents keep guessing which build, test, or validation command applies
• agents repeatedly search for the same repo facts before editing
• local safety rules live in prose, convention, or human memory
• brittle shell chains keep getting rebuilt for repo-specific chores

AXF is not

AXF is not a universal command catalog, a replacement for judgment, or an MCP-only product. It is the framework and control plane for building workspace-native agent exoskeletons from workspace-owned capabilities. MCP is one adapter surface; workspaces own the repo-specific capabilities agents use as their local exoskeleton.

Status: alpha. The core loop is in place: scout, inspect, execute, scaffold, and promote capabilities through one contract. Manifest version axf/v0 is the current alpha contract.

Install and MCP

License

This repository is source-available, not open source.

You may view, fork, modify, and run this project for personal, non-commercial use under the SmarterGPT Source-Available Personal Use License.

Commercial use, organizational use, employer/client use, production use, hosted-service use, redistribution, sublicensing, or embedding in another product or platform requires a separate written license from Joseph Gustavson / Guffawaffle / SmarterGPT.

Public visibility on GitHub does not grant open-source rights or business-use rights.

The npm package name is @smartergpt/axf. It installs two bins:

• axf — the CLI entrypoint, including the stdio MCP launch subcommand axf mcp
• axf-mcp — the stdio MCP server entrypoint

MCP clients may launch axf-mcp directly. For registry-driven or package-driven launch, prefer axf mcp so the package can stay centered on the base axf command surface. Both entrypoints start the same stdio MCP server. The server exposes exactly one MCP tool named axf, and that tool routes into AXF's existing capability surface for the bound workspace.

Global install

npm install --global @smartergpt/axf

Then use either bin directly:

axf doctor
axf mcp
axf-mcp

Audit a Codex MCP package pin before assuming the configured server matches the installed CLI:

axf integrate codex --check
axf integrate codex --write
axf integrate codex --check --smoke

Writes are scoped to the AXF package spec in Codex config.toml; restart or reopen Codex afterward. For one-call AXF + Lex bootstrap, see the packaged templates/session-context workspace recipe and docs/13-repo-onboarding.md.

Repo-local dev launch

From a local clone, run the bins directly with Node:

node /path/to/axf/bin/axf.js doctor
node /path/to/axf/bin/axf.js mcp
node /path/to/axf/bin/axf-mcp.js

A repo-local MCP configuration can either call the direct MCP bin or the CLI subcommand. Prefer the CLI subcommand when the client supports command arguments cleanly:

{
  "mcpServers": {
    "axf": {
      "command": "node",
      "args": ["/path/to/axf/bin/axf.js", "mcp"],
      "cwd": "/path/to/project",
      "env": {
        "AXF_PROJECT_ROOT": "/path/to/project",
        "AXF_EXECUTION_ROOT": "/path/to/project"
      }
    }
  }
}

Direct-bin configs remain valid:

{
  "mcpServers": {
    "axf": {
      "command": "node",
      "args": ["/path/to/axf/bin/axf-mcp.js"],
      "cwd": "/path/to/project",
      "env": {
        "AXF_PROJECT_ROOT": "/path/to/project",
        "AXF_EXECUTION_ROOT": "/path/to/project"
      }
    }
  }
}

npm / npx launch

Published-package smoke tests must run from a clean directory such as /tmp, not from the AXF repo root. Running from the repo root can let npx or npm exec pick up a repo-local or globally installed axf binary instead of the package-installed one.

The registry-friendly launch shape is axf mcp, but the direct MCP bin remains valid for manual configurations. These equivalent package-driven forms use the published npm package:

npx -y --package @smartergpt/axf axf doctor
npx -y --package @smartergpt/axf axf mcp
npx -y --package @smartergpt/axf axf-mcp

npm exec --yes --package @smartergpt/axf -- axf doctor
npm exec --yes --package @smartergpt/axf -- axf mcp
npm exec --yes --package @smartergpt/axf -- axf-mcp

Windows-native launch

For registry-style Windows MCP configs, prefer the base axf.cmd launcher with mcp as an argument:

{
  "mcpServers": {
    "axf": {
      "command": "axf.cmd",
      "args": ["mcp"],
      "cwd": "C:\\src\\my-project",
      "env": {
        "AXF_PROJECT_ROOT": "C:\\src\\my-project",
        "AXF_EXECUTION_ROOT": "C:\\src\\my-project"
      }
    }
  }
}

Direct-bin .cmd configs are still valid:

{
  "mcpServers": {
    "axf": {
      "command": "axf-mcp.cmd",
      "args": [],
      "cwd": "C:\\src\\my-project",
      "env": {
        "AXF_PROJECT_ROOT": "C:\\src\\my-project",
        "AXF_EXECUTION_ROOT": "C:\\src\\my-project"
      }
    }
  }
}

WSL / Linux launch

For registry-style WSL/Linux MCP configs, prefer the base axf command with mcp as the first argument:

{
  "mcpServers": {
    "axf": {
      "command": "axf",
      "args": ["mcp"],
      "cwd": "/home/user/src/my-project",
      "env": {
        "AXF_PROJECT_ROOT": "/home/user/src/my-project",
        "AXF_EXECUTION_ROOT": "/home/user/src/my-project"
      }
    }
  }
}

Direct-bin configs remain valid:

{
  "mcpServers": {
    "axf": {
      "command": "axf-mcp",
      "args": [],
      "cwd": "/home/user/src/my-project",
      "env": {
        "AXF_PROJECT_ROOT": "/home/user/src/my-project",
        "AXF_EXECUTION_ROOT": "/home/user/src/my-project"
      }
    }
  }
}

WSL users should use a native WSL AXF install on PATH, not Windows npm shims or npm's _npx cache. Optional shared packs that launch native CLIs should follow the same rule: prefer Linux-native tools earlier on PATH and avoid crossing into Windows shims from WSL.

Root binding

axf, axf mcp, and axf-mcp resolve the active project root and execution root in this order:

1. explicit --project-root and --execution-root
2. AXF_PROJECT_ROOT and AXF_EXECUTION_ROOT
3. legacy CLI aliases: --workspace, --registry-workspace, --execution-workspace
4. legacy environment aliases: AXF_WORKSPACE, AXF_REGISTRY_WORKSPACE, AXF_EXECUTION_WORKSPACE
5. nearest axf.workspace.json from cwd
6. nearest axf.workspace.json from the installed script location
7. cwd fallback

For axf run, unprefixed root flags are global only before the subcommand; after run they are capability-owned. Non-run commands retain their legacy placement because they have no downstream argument owner. The --axf-project-root, --axf-execution-root, and --axf-workspace spellings are accepted later in the AXF-controlled portion of a command. After an axf run ... -- boundary, root-like names belong to the capability.

Run argument ownership

AXF reserves the --axf-* namespace for framework controls and leaves natural public argument names to capabilities. Use an explicit boundary when the ownership should be visible in the command itself:

axf --project-root /repo run global.logs.query \
  --axf-json --axf-any-lifecycle -- \
  --json --limit 20 --workspace downstream-name

Options before -- are AXF run controls. Options after -- are public capability arguments: AXF still parses them, validates and coerces them with argsSchema, maps them through argMap, and executes them through the declared adapter. The boundary never enables raw or schema-bypassing argv. The same name may appear on both sides with different ownership, but a duplicate within either explicit section is rejected.

Existing no-boundary invocations such as axf run echo say --message hello remain supported. In that compatibility form, legacy --json, --any-lifecycle, and --allow-draft remain AXF controls unless the capability explicitly declares the same argument. Use --axf-json and --axf-any-lifecycle when AXF ownership must be unambiguous.

Agent-first discovery

Start with the bounded workspace guide instead of requesting every full manifest:

axf guide
axf guide context --json
axf list --compact --search lex --limit 20 --json
axf explain global.lex
axf inspect global.lex.status --json

The MCP router exposes the same guide, compact/search list, explain, and inspect results. guide returns workspace- or family-declared context, validation, and handoff entrypoints without executing them. Compact list entries retain lifecycle, side effects, and source provenance while omitting full schemas and execution targets until inspect.

See Agent Discovery and Workflow Guide for recommendation declarations, missing-capability diagnostics, and CLI/MCP examples.

When discovery and execution should use the same repo, set both AXF_PROJECT_ROOT and AXF_EXECUTION_ROOT to the same path. When they should differ, set them independently. Legacy AXF_WORKSPACE remains a compatibility alias that binds both roots to one path.

For MCP clients, set cwd to the intended caller execution directory and set AXF_PROJECT_ROOT / AXF_EXECUTION_ROOT explicitly when possible. That keeps manifest discovery and caller-facing execution deterministic.

Set AXF_MACHINE_ROOT when you want a user- or machine-scoped AXF root to contribute optional shared packs across projects. Project-root families shadow machine-level families with the same family name.

From any directory once installed:

axf doctor
axf list
axf inspect echo say
axf run echo say --message hello
axf scout --check
axf run toy echo say --message hello
axf init capability global.acme.status

axf doctor reports the workspace source it selected and, under WSL, warns when axf, lex, node, or npm resolve through Windows PATH entries under /mnt/c/....

MCP routing and governance

The MCP server exposes one tool named axf. Supported operations are:

• help
• list
• inspect
• run
• doctor
• scout_check

Those operations route into AXF's manifest-backed discovery, resolver, lifecycle, policy, adapter, and executor paths. MCP does not add a second governance model or a raw shell bypass around AXF.

AXF MCP is intentionally not full CLI parity yet. The CLI remains the authoritative mutation and control plane for commands such as init, promote, demote, scout --write, and other registry/materialization flows.

Capabilities such as global.echo.say are not separate MCP tools. Use operation=help to learn the router contract, operation=list to discover capabilities, and operation=inspect before operation=run. Treat capability lifecycleState, sideEffects, policies, and workspace binding as part of the execution contract.

scout_check is AXF's MCP-specific read-only structured scout diagnostics surface. It is not intended to promise literal CLI parity with axf scout --check.

Registry and manifest updates still happen through normal AXF CLI and filesystem/control-plane flows. The MCP server reloads registry state per request, so external AXF updates become visible on later MCP calls without restarting the server.

Full CLI parity may be considered later behind explicit policy and approval gates.

Shared packs appear through AXF only when the bound AXF project root or machine layer discovers or mounts those capabilities.

What's wired up

Built-in adapters

AXF ships two public built-in adapter types:

• internal — runs handlers in-process (adapters/internal/)
• cli — generic subprocess dispatcher with stdout JSON parsing (adapters/cli/)

Shared command families use the same adapter contract as project-owned capabilities. A shared pack can use the generic cli adapter, a provider adapter, or an internal handler, but AXF core does not require that pack as a runtime dependency.

Built-in capabilities

Capability	Provider	Lifecycle	Notes
global.echo.say	internal	active	smallest in-process capability example

Optional shared packs, including a Lex pack, can be added at machine or project scope. They route through the same resolver, lifecycle, policy, adapter, and executor path without defining the framework itself.

Toolspaces

• toy — smallest mount example; re-mounts echo.say with a local default

Repo onboarding

The recommended repo flow is:

1. Add axf.workspace.json at the repo root so workspace binding is explicit.
2. Add shared packs intentionally at machine or project scope when a repo wants them.
3. Add repo-specific capabilities separately under manifests/capabilities/ or manifests/families/.
4. Keep MCP optional; AXF works as a plain CLI capability router without it.
5. Mark mutating capabilities with sideEffects: "write". AXF does not yet have a first-class approvalRequired field, so approval gates stay a repo policy or review convention for now.

See docs/13-repo-onboarding.md for the concrete onboarding pattern and platform notes.

How to add a new provider

The contract is open. Every new provider goes through the same scaffolders and lifecycle gates:

# 1. Scaffold a draft provider adapter (only if the provider has an
#    envelope or quirks the generic cli adapter shouldn't carry):
axf init adapter --kind provider acme --composes cli

# 2. Scaffold each capability:
axf init capability global.acme.status

# 3. Edit the drafts, then:
axf doctor
axf run acme status --axf-any-lifecycle

The four canonical prompts under prompts/ walk an agent through discovery → planning → scaffolding → review against the actual file contract. JSON-first providers can usually use the generic cli adapter directly; provider adapters are for wrappers that need envelope, error, or argument normalization beyond the generic route.

Layout

axf.workspace.json               # workspace marker
bin/axf.js                      # CLI entrypoint
src/cli/                        # CLI parsing + main dispatch
src/core/                       # registry, resolver, executor, adapters, doctor, policy
adapters/<type>/                # type adapters (internal, cli, ...)
adapters/<provider>/            # optional provider adapters for wrapped CLIs
manifests/capabilities/         # capability manifests
manifests/toolspaces/           # toolspace mount manifests
prompts/                        # canonical prompts for agent-authored adapters
docs/                           # architecture, contract, lifecycle, prompts
test/                           # node:test suite

Reading order

1. docs/00-foundation.md — why axf exists
2. docs/01-vocabulary.md
3. docs/02-architecture.md
4. docs/03-capabilities-and-manifests.md
5. docs/04-adapter-contract.md — the two-kind adapter model (type + provider)
6. docs/05-lifecycle-and-promotion.md
7. docs/06-canonical-prompts.md
8. docs/07-v0-bootstrap-plan.md — alpha implementation milestones
9. docs/08-adapter-folder-shape.md — the concrete file contract
10. docs/09-launch-plans.md — interpreter-aware launch plans, env-bound roots, fallback paths
11. docs/10-command-families.md — family imports, public-to-provider arg mapping, materialization, drift
12. docs/11-normalization-guidance.md — JSON-first vs text-first providers, when to write a provider adapter
13. docs/12-layered-docs.md — caller / integrator / author paths through the docs
14. docs/13-repo-onboarding.md — workspace markers, shared packs, and platform guidance
15. docs/14-family-identity-and-layer-precedence-plan.md — family identity, layer precedence, and optional shared-pack direction

Tests

npm test

Uses Node's built-in node:test; no external test runner is required.

What is intentionally not here

• a broad command-alias layer
• privileged integration paths
• a plugin marketplace
• mandatory remote execution or MCP support
• agent-generated capabilities that auto-promote