01
Controlled tool use
Only registered tools are available to the model. Each tool requires an explicit permission grant.
Open-source · Go · AI tooling
Useful AI agents.
Explicit control.
AgentForge is a lightweight Go runtime for model-driven tool execution. It gives AI models a controlled way to use practical business tools—with bounded execution, validated inputs, and permissions that default to deny.
Early-stage developer project · v0.3.0
agentforge / local run
# Analyze a fictional CRM export
$ go run ./cmd/agentforge -tool lead_pipeline_analyze -input-file examples/lead_pipeline_analysis.json
{
"total_leads": 8,
"won_leads": 1,
"lost_leads": 1,
"open_leads": 6,
"win_rate_percent": 50,
"open_pipeline_value": 44500
}
# Deterministic local tool · no API key required
Designed for practical work
Small runtime. Clear boundaries.
A focused foundation for building AI-assisted workflows without starting with a large platform or opaque automation layer.
02
Bounded execution
Limits model/tool turns and supports context cancellation and execution timeouts.
03
Validated data
Tool arguments and results are checked against the project's supported JSON Schema subset.
04
Provider boundary
A small provider interface separates orchestration from the current OpenAI Responses API adapter.
05
Local-first tools
Deterministic text and CSV analysis can run directly from the CLI without a model API key.
06
Testable foundation
Go tests cover provider behavior, tool execution, validation, and permission enforcement.
How it fits together
A deliberately simple architecture.
The model proposes a tool call; the runtime validates and authorizes it before execution.
Application / CLIPrompt · config · permissions
Agent runtimeBounded orchestration loop
Model providerProvider interface · OpenAI adapter
Registered toolsValidate · authorize · execute
Permission checks are application-level controls, not an operating-system sandbox. Tools execute with the privileges of the process running AgentForge.
Current toolset
Useful building blocks.
The current tools are intentionally small, deterministic, and easy to inspect.
echo
Echo
Confirms the basic tool-call path with a simple response.
text_analyze
Text analysis
Word, Unicode character, line, and estimated reading-time metrics.
csv_analyze
CSV profiling
Rows, columns, missing cells, and numeric-column summaries.
lead_pipeline_analyze
CRM pipeline analysis
Won/lost/open counts, closed-lead win rate, estimated values, and stage/source breakdowns.
Try it locally
A reproducible, no-key demo.
Run the synthetic CRM example from the repository. No hosted service or paid model call is needed for this direct tool execution.
1. Clone and run
git clone https://github.com/matteojose3112-bit/agentforge.git cd agentforge go test ./... go run ./cmd/agentforge -tool lead_pipeline_analyze -input-file examples/lead_pipeline_analysis.json
2. Inspect the result
{
"total_leads": 8,
"won_leads": 1,
"lost_leads": 1,
"open_leads": 6,
"win_rate_percent": 50,
"total_estimated_value": 63700,
"open_pipeline_value": 44500
}The example uses fictional records and illustrative values. The win rate is calculated from closed leads only (won + lost). AgentForge reports descriptive metrics; it does not forecast revenue or modify CRM records.
Project status
Built in the open.
AgentForge is an evolving developer project, currently focused on a dependable tool runtime and business-relevant examples.
- Go runtime and CLI
- OpenAI Responses API adapter
- Default-deny tool permissions
- Automated tests, vet, and CI build
- Local business-data analysis examples
What it is — and isn't
AgentForge demonstrates controlled model-to-tool orchestration and deterministic local analysis. It is an early-stage foundation, not a production-ready autonomous platform.
- Live model/API smoke test remains a separate setup step.
- No persistent task history or workflow scheduler yet.
- Application permissions do not replace sandboxing or security review.