Tokless
One CLI wires code analysis, context compression, and concise instructions into multiple coding agents.
- Source repo
- HoangP8/tokless
- Stars
- ★ 274
- Last updated
- 10d ago
- License
- MIT
- Primary language
- Go
- FA score
- 57/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformCodex · Claude Code
- You'll need
- Typical use
- A developer who alternates between Claude Code, Codex, and other supported clients and wants one place to maintain their shared tooling.
- Main limitation
- The documented installers pipe a GitHub-hosted script directly into Bash or PowerShell, so adopters must assess the remote script and its write scope before execution.
- Source review
- 57/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
Tokless is a cross-platform installation and configuration CLI for coding agents on macOS, Linux, and Windows. It assembles karpathy-skills, caveman, ponytail, rtk, codegraph, and context-mode into a coordinated toolchain, then connects them through each client’s supported instructions, hooks, plugins, extensions, or MCP configuration. Documented clients include Claude Code, OpenCode, Codex, Antigravity, GitHub Copilot, Factory Droid, Pi, Cursor, Grok Build, Kilo Code, Oh My Pi, and Cline, although their integration methods and feature coverage differ. The installer detects supported agents already present on the machine and lets the user configure one, several, or all of them; the same CLI handles updates, diagnostics, disabling, and removal. Its deployment boundary is the user’s local agent configuration and per-project codegraph indexes—it does not provide a model, hosted coding service, or common chat interface.
Running tokless detects installed supported coding agents and wires in selected combinations of rtk, caveman, ponytail, codegraph, and context-mode. The documented instruction packages, including karpathy-skills, establish coding and response rules; rtk filters command output before it reaches the agent; codegraph indexes source relationships, call paths, and impact; and context-mode supplies sandboxed context operations, focused analysis, and session memory. Tokless writes the appropriate Instruction, Hook, Allow, Plugin, Extension, or MCP setup for each selected client. Its exposed MCP operations include codegraph_explore, ctx_execute, ctx_batch_execute, ctx_execute_file, ctx_index, ctx_search, and ctx_fetch_and_index. The tokless index command builds per-project codegraph indexes, tokless doctor reports what is wired and warns about broken pieces, and tokless update updates the CLI, shows version differences, and upgrades the installed tools.
- A developer who alternates between Claude Code, Codex, and other supported clients and wants one place to maintain their shared tooling.
- A team working in a large repository that wants indexed source relationships, call paths, and change-impact exploration.
- A terminal-heavy developer whose command output consumes too much agent context and needs filtering before the output reaches the agent.
- A frequent coding-agent user seeking focused context analysis, session memory, and reduced context-window consumption.
- A cautious adopter who wants to inspect intended configuration changes with --dry-run before any local writes occur.
- A team maintaining macOS, Linux, and Windows workstations that wants a consistent coding-agent toolchain across operating systems.
How do you install or deploy this agent?
On macOS or Linux:
curl -fsSL https://raw.githubusercontent.com/HoangP8/tokless/main/scripts/install.sh | bashOn Windows PowerShell:
irm https://raw.githubusercontent.com/HoangP8/tokless/main/scripts/install.ps1 | iexAfter installation, run tokless for interactive agent selection, or run tokless --agents claude,opencode to wire only those clients. The source does not specify a language runtime, administrator privileges, or required credentials. Installation requires network access to retrieve the GitHub-hosted script and associated tools. Restart configured agents afterward so they load the new configuration.
How do you use this agent?
For a first working invocation, run:
tokless
This installs the tools and prompts you to select detected agents; it is documented as safe to rerun. Use tokless --agents claude,opencode to restrict the target clients, or tokless --tools rtk,caveman,ponytail,codegraph,context-mode to select components. Add --dry-run to preview without writing, --verbose to show every step, or --yes to skip confirmations. Run tokless index to build per-project codegraph indexes, tokless doctor to inspect wiring and broken pieces, tokless info to display installation and configuration paths, and tokless update to update the CLI and tools. tokless disable disables selected agents, tokless uninstall removes everything Tokless touched, and tokless self-update updates only the CLI.
What are this agent's strengths and limitations?
- A single cross-platform CLI covers installation, updates, diagnostics, disabling, and removal across numerous coding-agent clients.
- It uses each client’s documented integration mechanisms—hooks, plugins, extensions, instructions, or MCP—instead of assuming one mechanism works everywhere.
- The bundled components have separated responsibilities: rtk filters output, codegraph maps code relationships, and context-mode handles focused context operations.
- --dry-run, --verbose, and tokless doctor provide concrete ways to preview changes and inspect the resulting setup.
- Users can select only the agents and tools they want rather than accepting the complete bundle.
- The documented installers pipe a GitHub-hosted script directly into Bash or PowerShell, so adopters must assess the remote script and its write scope before execution.
- Feature parity varies by client; Grok Build explicitly lacks RTK command rewriting.
- Agents must be restarted after installation to load the new configuration, interrupting active client sessions.
- The source does not document runtime versions, permission requirements, an offline installation path, or recovery behavior after network failures.
- The support table includes Kilo Code, but the documented --agents value list does not include kilo, leaving its CLI selection method unclear.
How does this agent compare with similar options?
Tokless does not replace karpathy-skills, caveman, ponytail, rtk, codegraph, or context-mode; it packages and connects those existing projects across coding-agent clients. Compared with installing and reconciling each component manually, it supplies one installation, update, diagnostic, and removal workflow. The tradeoff is that capabilities still depend on each client’s available hook, plugin, extension, and MCP mechanisms.
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| Tokless This agent | 57 · Major gaps | ★ 274 | 10d ago | Go | Codex · Claude Code |
| deepx-code | 63 · Some gaps | ★ 391 | 4d ago | Go | OpenAI API |
| LeanCTX Context Intelligence | 75 · Good | ★ 3.8k | 4d ago | Rust | Codex · Claude Code · OpenAI API · Claude API |
| Tura | 65 · Some gaps | ★ 639 | 10d ago | Rust | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The README provides interactive selection, agent/tool subsets, dry-run, verbose, disable, and uninstall controls, showing some attention to scoped changes, confirmation, and recovery. Its configuration matrix also outlines effects involving hooks, MCP, allow rules, and plugins. Deductions apply because installation pipes scripts directly from the moving main branch, dependencies use latest without pinning or integrity verification, and the exact files changed, network data flows, telemetry, credential handling, and other sensitive-data behavior are undocumented. Backups and recovery after partial failure are not shown. Third-party packages and the license holder are attributed, but publisher identity remains unknown and supply-chain provenance is incomplete.
The commands, tool inventory, and agent configuration are mostly coherent, while CI builds and tests on three operating systems. Deductions apply because go.mod specifies Go 1.22 while CI uses 1.23, and the advertised support list is not fully aligned with the --agents enumeration. context-mode is installed from npm at latest and its failure is ignored, with no pinned availability or fallback policy. doctor and verbose imply diagnostics, but the supplied files contain no concrete error messages, exit behavior, or remediation guidance.
The material clearly targets users seeking rapid, unified coding-agent setup, covers numerous agents and CLI/IDE/Desktop scenarios, supports agent/tool filtering, and backs OS compatibility with macOS, Linux, and Windows CI. Deductions apply because boundaries are largely limited to the configuration matrix and one Grok/RTK exception; there is no systematic account of unsuitable tasks, conflict resolution, or agent-version requirements. Command triggers are clear but selection semantics and noninteractive boundaries are not fully documented.
The README has strong organization across installation, tools, configuration, and usage, with useful command examples. The MIT license is complete, and the release badge, tag workflow, self-update command, and version-diff behavior establish an update path. Deductions apply for missing FAQ, troubleshooting, prerequisites, standalone changelog, and comprehensive known limitations. Minor naming or coverage inconsistencies appear around supported agents and flag values. Maintenance responsibility is only inferable from the repository and copyright name; no maintainer contact, support channel, contribution policy, or maintenance commitment is supplied.
A unified command surface for installation, updates, diagnostics, indexing, disabling, and removal is usable for ordinary administration, and the configuration matrix demonstrates plausible marginal value from integrating several existing tools across agents. Deductions apply because there are no run outputs, complete generated-config examples, performance measurements, or before/after comparisons. Claims such as “best tools,” “zero config,” and “under 30 seconds” are unsupported by measurements, while the resource, maintenance, and security costs of multiple hooks, MCP servers, and global npm tooling are not quantified.
Tool names, purposes, wiring methods, platform CI, and release mechanics are traceable to the README and workflows, and some facts—Go packaging, three-platform builds, tagged releases, and MIT licensing—are corroborated across files. Deductions apply because the central effectiveness and speed claims lack benchmarks, reported test results, or implementation evidence, and the supplied record omits application code, installation scripts, and tests. Marketing judgments are reasonably separated from configuration facts but are not explicitly labeled as measured results, inference, or subjective selection.
- Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
- The installation commands download and immediately execute shell or PowerShell code from the moving main branch; pin an audited revision and verify its contents or checksum first.
- Third-party tool versions are not pinned, and CI permits context-mode installation to fail, so installations at different times may produce different dependency sets or missing capabilities.
- Before enabling hooks, allow rules, plugins, and MCP configuration, separately inspect the actual installers, write paths, network access, permission scope, credential handling, and uninstall completeness.
- The “under 30 seconds,” “zero config,” and “best tools” claims are not supported by benchmarks or reported test results in the supplied files.